Most maps of open AI start with companies. This one starts with the actual things those companies release: models, repositories, services, licences, protocols, and institutions. That matters because a company can be open in one part of its stack and completely closed in another.

The Full AI Open Ecosystem Atlas Here

To gain Full Access, subscribe below.

Get Access To The Full BE Platform

The basic rule of the atlas is simple: the unit is the release, not the company. Meta is therefore not one “open” node. Llama 4 Scout and Muse Glimmer 30B are separate nodes because they ship under different terms. The same is true for Alibaba: Qwen3.8 27B and Qwen3.8 Max are classified separately because their licences differ. A company can release one model permissively, another under custom terms, and still sell proprietary services around both. Calling the whole company “open” would hide that distinction.

The map is organized by role in the stack. Compute sits at the bottom.

Applications sit at the top. In between are the tools that train, adapt, run, host, and route models. Institutions and protocols cut across the whole system. What the map deliberately does not show is size. A large company and a small project can appear as nodes of the same size because revenue, valuation, parameter count, downloads, and token traffic measure different things. A Hugging Face download is not the same thing as a request through Vercel. Neither is the same thing as enterprise adoption. Mixing them into one scale would make the picture look precise while comparing unlike things.

The same discipline applies to the connections. A line between two nodes means only that a specific relationship was verified: perhaps a model is supported by a serving engine, perhaps a host offers access to it, or perhaps a company belongs to a coalition. A line does not automatically mean market share, revenue, exclusivity, or strategic dependence.

The Business Engineer’s Agenting platform is now available to Executive members. You can start exploring it here: businessengineering.ai/agent

Get Access To The Full BE Platform

You can preview the platform at Business Engineering AI.

For the last three years, I’ve been rebuilding the Business Engineer’s curriculum from the ground up. That curriculum has now become the foundation of a new discipline, with the entire series taking shape around it.

Subscribe To Premium To Gain Access!

If you’re already a paid member, simply reply to this email, and we’ll send it your way. You can check out the full library here.

What “Open” Means Here

Get Access To The Full BE Platform

The word open is too broad to be useful on its own. The atlas separates eight different things:

• Open-source software: the code is released under an open-source licence.

• Permissive model weights: the model weights can be downloaded under relatively permissive terms such as Apache, MIT, or OpenMDW.

• Weights with custom terms: the weights are downloadable, but the publisher adds restrictions or special commercial conditions.

• Open research artifacts: the release includes more than weights, such as training data, recipes, or evaluation material.

• Source-available software: the source can be inspected or self-hosted, but use is still restricted.

• Public training data: datasets are released publicly under their own terms.

• Commercial ecosystem participants: companies that host, route, run, or apply open models without themselves being open.

• Institutions and licence frameworks: organizations and legal structures that coordinate the ecosystem.

The most important point is that these categories do not automatically travel together. A model can have permissive weights without public training data. A repository can be open while the hosted service built around it is proprietary. A model can be downloadable while its commercial use is restricted.

So when we say a model is “open,” the next question should always be: open in what sense, and under which terms?

That is why the atlas works at the level of the individual artifact rather than the company.

Run: Compute and Accelerators

Get Access To The Full BE Platform

The compute layer is small and highly concentrated. The atlas includes NVIDIA, AMD, Cerebras, and Groq. All four are commercial infrastructure companies. None needs to make its hardware open in order to benefit from open AI.

The important relationship is between open software and proprietary hardware. PyTorch can run on CUDA and ROCm. TensorRT-LLM is open-source software optimized for NVIDIA hardware. This creates a simple economic logic: if open models and open software make it easier for more people to run AI, the companies selling the hardware can benefit from the additional demand.

A chip company does not need to open its chip to benefit from open models. It only needs more workloads to run on the chip.

This is one of the recurring patterns across the atlas: openness in one layer can strengthen proprietary businesses in another.

Build: Development and Runtimes

Get Access To The Full BE Platform

A model file is not yet a usable system. It needs software around it. That is what the development layer does.

This layer includes:

• Training frameworks: PyTorch and Transformers.

• Data tooling and datasets: Datasets, FineWeb, Dolma.

• Post-training tools: PEFT, TRL, Unsloth, Axolotl.

• Serving engines: vLLM, SGLang, TensorRT-LLM.

• Local runtimes: llama.cpp, Ollama, MLX.

• Evaluation tooling: EleutherAI’s LM Evaluation Harness.

This layer is easy to underestimate because it sits between the model and the application. But it is what turns weights into something developers can actually use.

A model sitting on Hugging Face is basically an artifact. It becomes a working capability when software can load it, serve it efficiently, evaluate it, fine-tune it, and connect it to an application.

That is why serving engines are one of the hinges of the entire open ecosystem. If a new model works with vLLM, SGLang, llama.cpp, Transformers, and other established tools, developers do not need to rebuild the operating stack from scratch.

This is also where open ecosystems start compounding. The more common infrastructure exists, the lower the cost of adopting the next model. The value shifts from merely having access to the weights toward being able to run and adapt them well.

Create: Models and Open Research

Get Access To The Full BE Platform

The atlas contains twenty-five model releases from the United States, China, Europe, India, Korea, and the UAE. But the important point is not the country count. It is that different releases from the same lab can have very different licences.

For example, one Qwen release can use Apache 2.0 while another release from the same publisher uses its own family-specific licence. Other Chinese models expose downloadable weights under custom commercial terms. The same pattern exists in the United States: companies whose flagship services are closed can still publish individual models under Apache or other permissive licences.

So it is a mistake to say:

“Alibaba is open.”

or:

“Google is closed.”

The correct question is:

“What are the terms of this specific release?”

The same logic applies to model lineage. A regional model can be developed and controlled locally while still depending on a base model trained somewhere else. The atlas records Sarvam-M as being built on Mistral Small 3.1 Base, for example. That means the regional team owns its specialization, but the original training run still comes from another organization.

Fine-tuning a model locally does not necessarily mean you independently control the whole model stack.

Operate: Hosting and Deployment

Get Access To The Full BE Platform

Once a model exists, someone still has to run it.

That is a separate business.

The atlas includes sixteen hosting and deployment services:

• General model APIs: Fireworks, Together AI, DeepInfra, Novita.

• Specialized inference: GroqCloud, Cerebras Inference.

• Generative media: fal, Replicate.

• Custom deployment: Baseten, Modal, Runpod.

• Cloud platforms: Amazon Bedrock, Google Model Garden, Microsoft Foundry.

• Regional cloud: Scaleway.

• Customization: Tinker.

This layer makes the economics of open AI much easier to understand.

Suppose you use an open model through Fireworks. The model weights may be permissively licensed, but Fireworks is still providing a commercial service. You are paying for compute, memory, availability, scheduling, networking, and support.

The model can be open while the service running it is commercial.

That means the company that trained the model, the company that serves it, and the company that owns the customer relationship can all be different.

Open models therefore do not remove commercialization. They separate model creation from model operation.

Connect: Distribution and Routing

Get Access To The Full BE Platform

The routing layer sits above the hosts and gives developers a simpler way to reach many models and providers.

Hugging Face is a useful example because it plays two different roles. The Hub is primarily a repository where models and datasets are published. Hugging Face Inference Providers is a commercial routing layer that connects those models to hosting companies. Those are different functions.

OpenRouter and Vercel AI Gateway perform related commercial routing functions. LiteLLM and Portkey provide open-source ways to build or operate a similar routing layer yourself.

This creates an important separation:

the interface the developer calls no longer has to be the same company that actually runs the model.

You might call OpenRouter, choose a model from one lab, and have the request executed by an entirely different inference provider.

That turns routing into a new distribution layer above the host.

It also explains why gateway statistics must be interpreted carefully. Vercel or OpenRouter may see a huge volume of traffic, but that is still only the traffic that passed through their own distribution layer. They do not see every direct API call, private deployment, local model, or enterprise contract.

Apply: Applications and Agents

Get Access To The Full BE Platform

At the top of the stack are the products that actually turn models into work.

The atlas includes open frameworks such as LangChain and LangGraph, self-hosted agent software such as OpenHands, and systems such as OpenClaw that connect models to tools. It also includes source-available products such as n8n and Open WebUI, which are kept separate from fully open-source software because their licences impose additional restrictions.

On the commercial side are products such as Cursor and Perplexity.

The interesting change is that the relationship between models and applications is becoming two-way.

Traditionally, the model company built the model and the application company consumed it.

Now application companies can also contribute:

• usage patterns,

• evaluation data,

• failure cases,

• user requirements,

• and domain-specific feedback.

That means the application layer can influence how future models are trained and improved.

The application is no longer only a customer of the model. It can become part of the model-development loop.

This is strategically important because some of the most valuable learning may sit with whoever owns the workflow and the users, not with whoever originally trained the base model.

Coordinate: Institutions and Protocols

Get Access To The Full BE Platform

The final layer is not really above the others. It runs across them.

These are the institutions, licences, and protocols that make the ecosystem easier to coordinate over time.

The atlas includes the Open Source Initiative, the Linux Foundation, OpenMDW, the Open Secure AI Alliance, the Agentic AI Foundation, the Model Context Protocol, and the Nemotron Coalition.

Three different ideas need to stay separate here:

• Licensing defines what you are allowed to do with an artifact.

• Interoperability defines whether different systems can work together.

• Governance defines how an ecosystem or project is maintained over time.

A model can have open weights while using no open protocol. A protocol can be open while many products using it are closed. A company can join a coalition without changing the licence of any model it releases.

That is why membership is not a licence.

Mistral joining a coalition tells us something about Mistral as an organization. It does not automatically change the rights attached to Mistral Large 3.

The Ninety-Five Links

Get Access To The Full BE Platform

The atlas includes ninety-five verified relationships:

• 64 technical links: compatibility, framework support, repository availability, and model lineage.

• 16 commercial links: routes through which models can be accessed, hosted, or customized.

• 15 institutional links: coalition memberships, foundation relationships, and licence adoption.

The important point is that each type of link means only one thing.

If a model is linked to vLLM, it means there is documented compatibility. It does not mean vLLM is the most popular way to serve it.

If OpenRouter is linked to an inference host, it means a routing path exists. It does not tell us how much revenue passes through it.

If a company is linked to the Nemotron Coalition, it means the company is a member. It does not mean every model that company releases belongs to NVIDIA’s ecosystem.

The atlas deliberately refuses to turn one kind of evidence into another.

Compatibility is not usage. Access is not revenue. Membership is not control.

What the Atlas Refuses to Claim

Get Access To The Full BE Platform

The things left out are part of the research.

The atlas does not:

• rank nodes by revenue, valuation, token share, or parameter count;

• assume that every product from the same company uses the same licence;

• treat source-available software as equivalent to open source;

• treat use of OpenMDW as proof of OSI certification;

• reduce open AI to models while ignoring data, serving, evaluation, and post-training;

• turn company membership in a coalition into a property of every product;

• draw market-share or lineage connections without primary evidence;

• claim to be an exhaustive map of every company, model, price, or country.

This is important because maps become misleading when they try to fill every empty space.

A missing edge does not mean there is no relationship. It means the relationship was not verified here.

A missing market-share number does not mean market share is irrelevant. It means this atlas is mapping something different.

The atlas maps the structure of the ecosystem, not its exact economic size.

The Compression

Get Access To The Full BE Platform

The open-AI ecosystem is much easier to understand once it is broken into layers:

Compute → Development → Models → Hosting → Routing → Applications

with institutions, licences, and protocols coordinating across all of them.

The most important lesson is that “open AI” is not one type of company or one licensing model.

A model can be open while the hardware underneath it is proprietary.

The software used to run it can be open while the hosted service is commercial.

A closed application can contribute data or evaluations to an open-model coalition.

A regional model can be locally controlled while depending on a base model trained elsewhere.

A company can simultaneously participate in open-source software, permissive model releases, custom-licensed weights, and proprietary services.

That is not a contradiction.

That is the ecosystem.

The weights are only one piece. Around them sits an industrial stack that trains, adapts, serves, routes, integrates, evaluates, and governs those weights.

And once those layers are separated, the real question becomes much clearer:

Which parts of the AI stack are actually open, which parts are commercial, which parts are replaceable, and where does control really sit?

The Mental Models

Get Access To The Full BE Platform

• The Release Is the Unit. Do not ask whether a company is open. Ask what the terms are for the specific model, repository, or service you are using.

• Open Has Several Meanings. Open-source code, downloadable weights, permissive commercial rights, public data, and open research are different things.

• Weights Are Not a System. A model becomes useful only when training, serving, evaluation, and runtime software can actually operate it.

• Open at One Layer Can Strengthen Closed Economics Elsewhere. Open models can increase demand for proprietary chips, clouds, or commercial hosting.

• Creating and Running Are Different Businesses. The company that trains the model does not have to be the company that serves it.

• Routing Is a Distribution Layer. The API you call can be separate from the company performing the inference.

• Lineage Matters. A local or sovereign model may still depend on somebody else’s base-model training run.

• Membership Is Not a Licence. A company joining a coalition does not change the terms of its products.

• A Link Means Only What It Proves. Compatibility is not usage. Access is not revenue. Membership is not control.

• The Missing Claims Matter Too. When public evidence cannot establish something, the atlas leaves it out rather than pretending certainty.

With massive ♥️ Gennaro Cuofano, The Business Engineer

A version of the AI story can be told entirely through league tables. Who has the largest model? Who bought the most GPUs? Who owns the biggest cloud? Who spent the most capital? Who has the strongest benchmark? Who shipped the fastest accelerator? Those rankings are easy to produce and increasingly misleading because they compare companies that are not trying to win the same game.

Google does not need to become NVIDIA. NVIDIA does not need to own Google’s distribution. Anthropic can be strategically strong while depending on almost everyone underneath it. Cerebras can matter while occupying only a fraction of the stack. SpaceX can look incomplete as a column and extraordinarily complete as a circuit. DeepSeek cannot be understood properly if you grade it against the same industrial assumptions as an American frontier lab.

A better instrument is a grid. Put the nine layers of the AI stack down one axis: Energy → Foundries & Memory → Silicon → Networking → Compute → Models → Harness → Distribution → Governance. Then place the major players across the other axis: Google, Meta, NVIDIA, Microsoft, Amazon, SpaceX, Anthropic, OpenAI, Cerebras, DeepSeek. Now stop asking who is biggest and ask a more useful question in every cell: What does this company actually control here, how strong is that position, and why does it own that layer at all?

What emerges is not a ranking. It is a set of geometries. And the geometry is the strategy.

That’s why I’ve been working for months on the Atlas of AI.

If you are an existing paid member of The Business Engineer, please reply to this email to gain access to the AI Atlas.

The Nine Layers

The matrix starts at the bottom because AI begins as a physical system long before it becomes a model. Energy is the floor. Data centers do not run on abstractions. They run on substations, generation contracts, grids, gas, nuclear, renewables, batteries, and increasingly strategic power agreements. The question is not simply who consumes the most electricity, but who can secure the next megawatt when everybody wants the same one.

Above energy sit foundries and memory. TSMC capacity, advanced packaging, HBM, domestic fabs, export controls, and memory supply determine how much silicon can actually be produced. A company can design a magnificent accelerator and still discover that somebody else owns the manufacturing floor. Then comes silicon: GPUs, TPUs, Trainium, Maia, custom inference chips, wafer-scale architectures, and the proliferating attempts to reduce dependence on a single merchant supplier. Silicon gets most of the attention because it is visible, expensive, and benchmarkable, but it only becomes useful at scale because networking turns thousands of accelerators into one system. Interconnect, Ethernet, InfiniBand, optical systems, and data-center topology determine whether a warehouse of chips behaves like one computer or like a warehouse of stranded chips.

Above networking sits compute, where the industrial stack becomes a financial one. Compute is not merely how many accelerators exist. It is who owns them, who financed them, whether the capacity is internal or rented, whether the data center is leased or owned, whether power is contracted separately, and whether the customer paid in advance. Then comes models, the layer public narratives often confuse with AI itself. Frontier models remain enormously important, but the matrix makes clear that they are one layer in a much larger system. Some companies treat the model as the center. Others increasingly treat it as a component that can be routed, replaced, or rented.

The next layer is the one I increasingly think matters most: the harness. This is where models are routed, enterprise context is assembled, tools are connected, permissions are enforced, memory persists, evaluations run, agents interact with systems, and intelligence becomes actual work. Then comes distribution: browsers, operating systems, enterprise seats, clouds, social graphs, search, messaging, marketplaces, APIs, and devices. Distribution determines who receives the user’s intent before anybody else. At the top sits governance: policy posture, compliance, enterprise trust, sovereignty, safety boundaries, procurement eligibility, and ultimately whether intelligence is allowed to operate in the markets and institutions that control real budgets.

Nine layers. But almost nobody is trying to dominate all nine in the same way.

Six Shapes of AI

The matrix resolves into six broad geometries: the vertical, horizontal, flywheel, lab posture, niche, and indigenous stack. These shapes matter because they describe not only technological position but how each company expects value to move through the stack.

The vertical. Google is the purest expression. It occupies all nine layers: power procurement and infrastructure at the bottom; deep foundry relationships; TPUs at silicon; proprietary networking; massive compute; Gemini at the model layer; agent and developer infrastructure in the harness; Search, Android, Chrome, Workspace, YouTube, Cloud, and billions of devices at distribution; and a broad policy and compliance machine at governance. The cells do not need to be equally strong. Completeness itself creates an advantage. When silicon becomes constrained, the vertical has its own architecture. When merchant GPUs become expensive, it can route toward TPUs. When model economics compress, it still owns infrastructure and intent. When power becomes the bottleneck, it can sign directly against generation.

Meta represents another form of the vertical. It increasingly owns more of the column, from custom silicon and networking through massive compute, open models, and consumer distribution. But Meta exposes the other side of the strategy: the more of the column you internalize, the more of the column you must finance. Vertical integration replaces supplier dependence with capital intensity.

The horizontal. NVIDIA also touches almost every layer, but its strategic logic is nearly the inverse. NVIDIA does not primarily want to internalize the stack for its own applications. It wants everybody else’s stack to run through NVIDIA. Silicon remains the center, but the position stretches into networking, systems, software, model tooling, cloud relationships, enterprise distribution, and increasingly the financing surrounding the infrastructure that consumes its products. The vertical asks, How much of my own stack can I control? The horizontal asks, How many other stacks can I become indispensable to?

Microsoft and Amazon belong to the horizontal family for another reason. Both have enormous strength in compute and distribution while deliberately maintaining optionality across models and silicon. Azure can distribute OpenAI while developing its own models and accelerators. AWS can host Anthropic while building Trainium, offering many other models, and using Bedrock as the abstraction layer. These are not incomplete verticals. They are portfolios of dependencies designed to stop any one dependency from capturing the whole stack.

The flywheel. SpaceX looks incomplete if read as a conventional column. Read as a circuit, it becomes much more interesting. An operating business produces cash and proprietary data. Cash funds infrastructure. Infrastructure supports compute. Compute supports the model. The model runs inside a harness connected to distribution controlled by the same industrial family. Usage then generates data and operating feedback that flow back toward the system. The important word is not integration but co-adaptation. The model changes the product, the product changes how the model is used, usage generates new data, and the data changes the next model. The vertical integrates supply. The horizontal integrates markets. The flywheel integrates learning.

The lab posture. Anthropic and OpenAI look structurally strange by the standards of the old technology industry: extraordinarily strong near the model and harness, highly dependent underneath. Neither owns a foundry. Neither owns a general-purpose networking estate comparable with the hyperscalers. Both depend heavily on external compute and power supplied through somebody else’s balance sheet. Two years ago that might have looked like obvious fragility. Today it can also be read as deliberate capital allocation. Why own foundries, networks, and power generation if those layers can be rented while differentiation remains higher near intelligence, orchestration, and the customer relationship?

Anthropic makes the logic particularly visible because portability across several silicon substrates can itself become leverage. If the workload can move, infrastructure providers compete for it. Dependence becomes less dangerous when it is portable dependence. The doctrine is simple: own the junction, rent the floor. Whether that remains rational depends on where durable value ultimately settles.

The niche. Cerebras represents almost the opposite strategy. Do not occupy nine cells. Put almost everything into one architectural bet located directly against an incumbent bottleneck. Wafer-scale computing attempts to alter the relationship between compute, memory, and interconnect by changing the physical unit itself. Its value does not come from breadth but from whether the chosen cell sits exactly where the constraint becomes economically painful. A niche is therefore not a weak vertical. It is a concentrated attack on one constraint.

The indigenous stack. DeepSeek belongs partly to another map. Its stack operates under different industrial constraints: export controls, domestic silicon, Chinese foundries, local energy systems, policy support, and a strategic requirement to reduce exposure to Western bottlenecks. Seen purely through American infrastructure economics, some cells appear weaker. Seen through sovereignty, those same cells can be strong. Technical efficiency is not always the optimization target. Strategic autonomy can be. Weights cross borders relatively easily. Advanced lithography does not. Ideas diffuse. Foundry capacity, HBM supply, packaging, power infrastructure, and export permissions remain physical. The geopolitical fence becomes harder as you descend the stack.

The Paradox of Integration

Google exposes an important paradox. The TPU began as an internal advantage: specialized silicon designed to improve Google’s economics and reduce dependence on merchant accelerators. Then the TPU became commercially valuable. Once customers want the internal advantage, the company has another option: monetize it externally.

That creates a conflict. Capacity sold to an outside customer cannot simultaneously serve internal frontier researchers. The better the merchant business becomes, the more the infrastructure organization has to decide whether it exists to maximize Google’s internal technological advantage or maximize the economic return on the infrastructure itself. Once a proprietary system develops external customers, revenue expectations, utilization targets, and its own capital requirements, an internal moat begins behaving like a merchant business.

This does not mean vertical integration failed. It means successful vertical integration creates a second-order problem: what happens when the internal bottleneck becomes valuable enough for outsiders to buy? Google built the cage. Then the market wanted to rent the bars.

What Q2 Changed

A useful map has to move. This quarter several cells did, but the most important shifts were not weak-to-moderate or strong-to-dominant. They were changes in kind.

NVIDIA moved from supplier toward supplier-financier. NVIDIA remains the critical equipment supplier at the center of the AI infrastructure buildout. What changed is its relationship with the capital structure around its customers. Investments, partnerships, receivables, infrastructure commitments, and other forms of support increasingly connect NVIDIA to companies whose expansion ultimately drives demand back toward NVIDIA equipment. The exact instruments vary. The structural change is what matters: the supplier becomes more entangled with the financing of its own ecosystem.

The railway analogy is difficult to avoid. Infrastructure can have enormous long-term economic value while securities financing the build still perform terribly. Railways transformed economies; individual projects failed, ownership changed, credit was restructured, and assets survived. The AI version contains a stranger twist: the company selling the locomotives is increasingly helping finance the railway. That does not prove demand is fictional. It means supplier strength and financing risk can coexist.

Meta’s compute cell cracked, but demand did not. The important distinction is not demand versus no demand. It is self-funded build versus financed build. When capital expenditure approaches the operating cash engine, free cash flow compresses, buybacks disappear, and debt begins playing a larger role, the technological thesis can remain intact while the financing architecture changes. A company can be right about AI demand and still discover that its previous capital structure cannot support the rate at which it wants to build.

Meta retains an important valve because the assets are real. Capacity can be slowed, redirected, monetized, leased, or eventually offered externally. Physical infrastructure contains optionality in a way operating expense does not. The better interpretation is therefore not that Meta’s demand thesis failed, but that its prior self-funding model became insufficient for the scale of the build it wanted to undertake.

Google’s silicon became merchant silicon. Again, the significant movement is not that the cell became stronger. It changed economic function. A captive advantage increasingly becomes a product. Once that happens, allocation of proprietary infrastructure between internal workloads and external demand becomes a capital allocation problem.

That distinction, changes in kind rather than degree, is precisely what a matrix like this is designed to reveal.

Compute Is the Row That Explains the Quarter

Now stop reading vertically and read across compute. The quarter becomes much more interesting.

Google is deliberately stretching a fortress balance sheet to expand capacity. Microsoft retains an extraordinary operating engine, but part of the infrastructure burden increasingly sits in signed leases and future obligations rather than only conventional owned capex. Amazon continues absorbing a vast build through operations while its cash profile becomes more demanding. Meta crosses from predominantly internal self-funding toward a structure where external capital matters more. OpenAI sits farther along the dependency curve, relying heavily on suppliers, infrastructure partners, contracted capacity, and balance sheets outside its own.

Different companies, one underlying question:

Who finances the next unit of compute?

There are only a few answers. The first is own cash flow, the cleanest structure. Existing businesses generate enough cash to fund capacity directly. The second is capital markets, with debt or equity bridging current cash flow and future infrastructure demand. The third is the customer, through prepayments, minimum commitments, long-term contracts, and backlog. The fourth is the supplier, through equity, extended payment terms, receivables, guarantees, infrastructure support, or strategic investment. The fifth is everybody else, as scarcity migrates through memory prices, electricity, leases, component costs, and downstream margins.

The farther the system moves down that list, the more interesting the financial architecture becomes. This is the most important shift of the quarter: AI infrastructure did not suddenly become expensive. It became increasingly financed.

That difference is enormous. An expensive build financed from operating cash behaves very differently from the same physical infrastructure financed through debt, prepayments, supplier support, long leases, equity issuance, and interdependent backlog. The servers can be identical. The financial system underneath them is not.

The Stack Becomes a Network of Balance Sheets

The matrix therefore becomes more than a technology map. It becomes a capital map. The AI build no longer sits neatly inside the capital-expenditure line of a handful of hyperscalers. Risk travels between participants.

A cloud provider signs a long-term capacity agreement. A data-center operator finances against that agreement. A supplier supports the operator. A model company commits future usage. A customer prepays the model company. A utility builds against expected data-center load. Memory suppliers raise prices as each accelerator consumes more constrained HBM. The cost ultimately appears across several companies and several financial statements.

Nothing in that chain is necessarily irrational. The danger is that the same underlying demand can appear multiple times: as one company’s backlog, another company’s capex plan, another’s receivable, another’s strategic equity investment, and another’s expected future revenue. As the cycle matures, revenue quality and financing quality matter more precisely because the technological demand is real.

Risk did not disappear. It changed owners.

The prints after the quarter sharpened the same point. One major infrastructure provider can carry contracted backlog equivalent to years of current revenue while producing strongly negative free cash flow and financing part of the expansion with substantial equity issuance. The backlog headline establishes demand. The financing tells you how that demand reaches production. Another custom-silicon provider can report extraordinary AI growth and commitments from the largest model builders while simultaneously seeing margins pressured by increasing memory content and customer deployments supported by financing vehicles.

Neither result invalidates the demand thesis.

Both say something more interesting:

real AI demand is increasingly being fulfilled using somebody else’s balance sheet.

Bottlenecks Cascade Down the Stack

The map also explains why debates framed simply as NVIDIA GPUs versus custom silicon miss the deeper mechanism. Silicon sits on top of foundry capacity, advanced packaging, and memory. HBM in particular can tax almost every architectural path through the stack. NVIDIA can design better GPUs. Google can design TPUs. Amazon can build Trainium. Microsoft can develop Maia. Broadcom can participate in custom accelerators. Cerebras can redesign the physical architecture entirely. But if all of these systems increase demand for constrained memory, diversification in silicon does not eliminate the bottleneck. It moves the constraint one layer downward.

This is one of the recurring laws of the AI Supercycle: solve one bottleneck and the constraint migrates. GPU scarcity drives custom silicon. Custom silicon increases pressure on advanced packaging. Packaging and accelerator density increase pressure on HBM. Compute and memory increase pressure on power. Power pushes the constraint into generation, substations, grid access, and ultimately financing. The stack is not a collection of independent layers. It is a cascade of constraints.

This is also why energy belongs inside the AI map rather than beside it. At the scale implied by frontier training, inference, persistent agents, and data-center construction, power becomes product strategy. An infrastructure company can have chips, networking, funding, and customers and still fail to deploy because the relevant substation will not be available for years. The deepest vertical is therefore not merely the company that owns a model and a cloud. It is the company capable of converting electrons into accepted AI work with the fewest external veto points.

Governance plays a strangely symmetrical role at the opposite end. For consumer AI it can look like friction; for regulated enterprise deployment it can become distribution. A technically superior model that cannot satisfy procurement, sovereignty, data residency, auditability, safety, or sector-specific requirements may be commercially irrelevant for entire classes of workloads. Energy determines whether intelligence can physically run. Governance determines whether institutions allow it to run there.

The two ends of the stack matter more than most conventional AI maps imply.

Distribution Owns Intent

Distribution becomes increasingly important as model capability converges. Google owns Search, Android, Chrome, Workspace, YouTube, and Cloud. Meta owns consumer social and messaging surfaces measured in billions of users. Microsoft owns Windows, Office, GitHub, Azure, and much of the enterprise seat estate. Amazon owns enormous commerce intent and cloud distribution. OpenAI owns one of the first truly AI-native consumer destinations. Anthropic is building a different kind of distribution through developers and enterprises.

Distribution matters because it owns the request. The company that receives intent first can influence which model answers, which tool is called, which merchant receives the transaction, which source is retrieved, and which workflow executes.

But intent alone is not the whole game.

Distribution owns the request.

The harness owns what happens after it.

And that brings us to the most important row in the matrix.

The Missing Layer

Read across the harness row. Almost every serious player is moving toward it. Google has agent infrastructure and increasingly rich developer and protocol surfaces. NVIDIA extends software and orchestration above its hardware. Microsoft combines Azure, GitHub, Copilot, enterprise context, identity, and an enormous installed base. Amazon uses Bedrock as a model abstraction and enterprise execution layer. Anthropic increasingly holds an important protocol position through MCP. OpenAI has Codex, agent surfaces, developer tooling, and a rapidly expanding action layer. Even companies whose original differentiation sits elsewhere keep moving toward the same layer.

The reason is simple.

This is where the model becomes work.

The harness decides which model runs, what context enters, which memory persists, which tools are available, what identity and permissions apply, which actions require approval, how outputs are evaluated, what happens after failure, what gets recorded, and what gets learned.

The model provides intelligence.

The harness governs the application of that intelligence to something economically consequential.

That distinction may define the next phase of the stack.

Models remain immensely valuable, but intelligence is becoming increasingly purchasable. Several frontier models can be routed behind the same application. Open weights improve. Smaller models become more capable. Inference costs decline. Models specialize. Enterprises increasingly adopt multi-model architectures.

None of this makes frontier models irrelevant. It changes the bargaining structure around them.

If several models can perform a routine task above the required threshold, the layer choosing among them gains leverage. If the enterprise owns its context, process state, permissions, evaluation standard, decision history, and tool connections, replacing the underlying model becomes possible.

Replacing the entire operating layer above it is much harder.

Intelligence becomes increasingly rentable. The junction is not.

Where Intelligence Becomes Work

The harness matters most at the junction, the place where intelligence meets something that cannot be reduced to another prediction: a bank transfer, a procurement approval, an insurance claim, a medical decision, a contract, a customer record, a supply-chain action, a production system, or a regulatory boundary.

The model can generate an answer.

The valuable system decides whether that answer is allowed to become an action.

That requires context, identity, permissions, business logic, evaluation, governance, human escalation, and a record of what actually happened. The moat therefore shifts from merely possessing intelligence toward controlling the conditions under which intelligence becomes consequential.

This is the layer many traditional AI maps leave almost empty. They jump from models directly to applications, as though the distance between generating language and reliably operating a company were simply another user interface.

It is not.

It is becoming an industry of its own.

The harness is the deployment layer becoming a product.

This also makes the frontier labs easier to interpret. Anthropic does not necessarily need to own a substation if its durable position is the protocol through which enterprise agents reach context and tools. OpenAI does not necessarily need to own a foundry if it controls an agent surface through which users initiate valuable work. Those physical dependencies still affect cost, margin, resilience, and bargaining power, but dependency itself is not automatically strategic weakness.

Every company chooses which layers deserve ownership.

The question is whether the labs have chosen correctly.

If frontier models remain scarce, their position is obvious. If models commoditize while labs control harness and distribution, they may remain extraordinarily powerful. If models commoditize and the harness standardizes elsewhere, they become much more vulnerable.

That is the strategic test.

The Matrix Is Also a Capital Map

The same matrix can therefore be read in three ways. As a technology map, it shows what each company controls. As a dependency map, it shows what each company must obtain from somebody else. As a capital map, it shows which layers each company is willing to finance.

Google finances breadth. Meta finances verticality around a massive consumer estate. Microsoft finances infrastructure while using partnerships to preserve optionality. Amazon finances physical capacity because infrastructure itself is one of its core products. NVIDIA historically monetized everybody else’s capex and now increasingly participates in the financial structure surrounding that capex. Anthropic concentrates resources high in the stack and rents much of the industrial floor. OpenAI uses partnerships and external balance sheets to reach a scale its own operating cash flows could not independently finance. Cerebras concentrates capital around one architectural thesis. DeepSeek optimizes partly for indigenous survivability rather than pure global efficiency. SpaceX can fund a flywheel from cash generated elsewhere inside the same industrial system.

Each geometry therefore answers another question:

Where are we willing to trap capital?

That question may ultimately matter as much as where the technology sits.

From Seats to Tokens to Outcomes

There is also an economic progression hidden inside the stack. Software was monetized through seats because the human user was the unit. Models introduced tokens because inference became the production unit. The harness increasingly introduces another possibility: the accepted outcome.

• A claim processed.

• A document reviewed.

• A piece of software shipped.

• A support request resolved.

• A procurement decision completed.

• A reconciliation performed.

• A workflow executed.

This matters because it changes who can meter value.

The model provider meters intelligence consumption.

The application provider can increasingly meter work completed.

That unit has different economics. It contains model cost, infrastructure cost, tool calls, human review, exception handling, and the value of the business process itself. Whoever owns that meter gains a very different relationship with the customer.

Seats belonged to software.

Tokens belong to models.

Accepted outcomes may belong to the harness.

If that happens, the harness does more than orchestrate the technical stack. It becomes the economic layer that decides how much of the customer’s willingness to pay flows downward toward models, compute, silicon, and energy.

How to Read the Atlas

A matrix like this becomes useful only if it remains falsifiable. Three habits matter.

Read rows before columns. A column explains the strategy of a company. A row reveals the structure of the market. Compute told the financial story of this quarter more clearly than any individual earnings release. Memory shows where the next hardware tax may sit. Distribution shows who owns intent. The harness shows where software control may be migrating.

Treat every cell as a claim. “Supplier-financier,” “self-funding cracked,” “merchant silicon,” “multi-substrate,” “backstopped,” “rent-payer”: these are not aesthetic labels. Each should have evidence behind it and conditions under which the judgment would change next quarter. A cell that cannot move is branding, not analysis.

Watch changes in kind before changes in degree. Strong becoming dominant matters. Supplier becoming financier matters more. Captive becoming merchant matters more. Dependent becoming portable matters more. A business becoming an infrastructure provider when its internal capacity turns commercial matters more than another percentage point of share.

The most interesting quarters are the ones in which a square starts performing a different economic function.

What the Atlas Says About the AI Supercycle

The technology is real. Infrastructure demand is real. Monetization is increasingly real. Financial excess can also be real. Those statements are not contradictory. Infrastructure cycles almost always contain both a technological truth and a financial overreaction to that truth.

Britain needed railways. It still financed too many of them in the wrong structures and at the wrong prices. The Internet transformed the economy. Investors still lost extraordinary amounts of capital financing telecom infrastructure and dot-com companies. Fiber remained. Networks remained. Data centers remained. Ownership changed. Prices reset. Capital structures broke. The useful assets survived.

The AI Supercycle can follow the same pattern. Some capacity will be overbuilt. Some data centers will be badly financed. Some scarcity rents will disappear. Some power contracts will look absurd in hindsight. Some memory pricing will collapse when capacity catches up. Some model economics will compress. Some interfaces will lose value as agents route around them. Some companies will discover that a technically excellent asset sits on the wrong side of the value migration.

None of this means the infrastructure was unnecessary.

That is why asking whether AI is a bubble is usually the wrong question.

The better question is:

Which layers remain economically necessary after the financial excess clears?

And here the matrix becomes contrarian. It is tempting to assume the deepest physical layer must capture the deepest value because it is hardest to build. Not necessarily. Foundries are incredibly difficult. Memory is constrained. Power is critical. Networking is sophisticated. Compute consumes enormous capital. All can generate significant rents. But a business does not become strategically defensible merely because its inputs are difficult.

Owning a substation is not enough if somebody else owns the customer, the workflow, and the economic decision.

The durable layer may ultimately be the place where all those scarce inputs are transformed into something a customer actually buys.

For enterprise AI, that may increasingly be the harness.

The place where models become claims processed, invoices reconciled, contracts reviewed, software shipped, customer interactions resolved, procurement decisions completed, and workflows operated.

The layer where intelligence becomes work.

The Missing Layer Was the Point

The Vertical Integration Matrix began as an attempt to understand who owned the AI stack. It ends by questioning whether owning the entire stack is even the right objective.

Google’s completeness is extraordinarily powerful. NVIDIA’s horizontality is extraordinarily powerful. SpaceX’s flywheel can be extraordinarily powerful. Anthropic’s selective dependence may prove entirely rational. Cerebras may be right that one architectural cell is enough. DeepSeek demonstrates that sovereignty creates another definition of completeness altogether.

There is no universally correct geometry. There is only a geometry appropriate to where scarcity currently sits and where the company expects value eventually to settle.

That is why the map has to remain dynamic. Today enormous rents sit in silicon and compute because those layers are constrained. Tomorrow some of those rents may compress. Memory can become the bottleneck, then power, then financing. Models can remain highly concentrated at the frontier while becoming interchangeable for much ordinary work. Distribution remains powerful because somebody has to receive intent. Governance grows in importance as systems move from answering to acting.

And between all of them sits the layer whose strategic importance can rise under almost every scenario:

the system deciding how intelligence gets used.

The harness.

The layer holding context, routing, permissions, tools, evaluation, workflow state, identity, memory, and the junction between prediction and action. The layer where intelligence stops being a model output and becomes an economic outcome.

The atlas therefore does not tell us who wins AI. It gives us a more useful question.

When the next bottleneck moves, when compute gets repriced, when capital becomes scarcer, when models converge, and when agents begin doing meaningful work across the enterprise:

Which layers will everybody else still need?

That is the purpose of the map. Not to count how many squares each company owns, but to identify the squares that will still matter after the colors change.

Nine layers. Ten radically different strategies. One financing system becoming increasingly intertwined underneath them.

And one layer the market may still be systematically underpricing:

the layer where intelligence becomes work.

With massive ♥️ Gennaro Cuofano, The Business Engineer

Something changed for me in 2020.

The pandemic exposed just how interconnected the global economy had become, but also how physical many of those connections still were. Behind the abstraction of globalization sat semiconductor fabs, pharmaceutical plants, shipping routes, energy systems, rare-earth processing, industrial machinery, medical supplies, and basic materials that most businesses had spent decades treating as if they would simply always be available.

Suddenly they were not.

That was the moment I started looking at the global economy differently. What appeared to be a highly distributed and digital system was sitting on top of a surprisingly small number of physical dependencies. A disruption in one geography could stop factories thousands of kilometres away. A shortage of one component could immobilize entire industries. Governments could suddenly decide that something previously treated as an ordinary commercial input was strategic.

For me, that triggered a much deeper exploration of geopolitics. Not as a separate discipline sitting beside business analysis, but as something that increasingly had to sit above it.

In retrospect, the preceding three decades were the historical exception. After the end of the Cold War, businesses operated during an extraordinary period of globalization under a largely US-led international order. Supply chains stretched across political boundaries because companies could largely assume that trade, capital, technology, energy, and information would continue moving across them. You could build a serious business framework while paying relatively little attention to spheres of influence, strategic dependencies, industrial sovereignty, or the possibility that governments might deliberately restrict access to critical technologies.

For most of history, that would have been a strange assumption.

Political power, geography, access to resources, and control of strategic infrastructure have shaped commerce for centuries. What was unusual was the period in which many businesses could behave as if they did not.

The realization hit me almost like a flash in 2020: there was no going back to a world in which geopolitics could remain outside the business framework.

And the years since have only reinforced that view. Semiconductors became instruments of state power. Energy security returned to the center of European strategy. Supply chains were reorganized around resilience and political alignment. Advanced manufacturing became industrial policy. AI infrastructure became an electricity and sovereignty question. Space infrastructure became increasingly entangled with communications, navigation, intelligence, and warfare. Biology began moving closer to the same strategic frontier.

This changed the hierarchy of my own analytical framework.

I used to think primarily from macroeconomics into technology and then into business. Increasingly, I think the causal chain has to begin one level higher:

Geopolitics → Macroeconomics → Technology → Business.

Geopolitics increasingly defines what can move, where it can be built, who can access it, and under what conditions. Those decisions shape capital flows, energy prices, industrial policy, trade, and monetary dynamics. Those macro conditions determine where technologies can scale. And only then do those technologies cascade into industries, companies, business models, and competitive advantage.

That does not mean geopolitics determines everything. It means that a framework for understanding technology and business is increasingly incomplete if it ignores the political and physical system underneath them.

This is also one of the reasons I eventually spun The AI Supercycle out of The Business Engineer. What began for me as an attempt to understand AI increasingly became an attempt to understand the much larger system forming around it: compute, energy, semiconductors, infrastructure, capital, industrial policy, space, biology, and the geopolitical competition that increasingly connects all of them.

AI may be the most visible technology of this cycle.

But the cycle itself is much larger than AI.

So let me tell you the full story of how I see the next decade unfolding.

This is the rise of techno-geopolitics. Technology is no longer something that can be analysed separately from geography, infrastructure, state power, and strategic dependencies. The two are becoming one system.

That is also why I’m republishing my Map of AI workshop alongside this piece. The workshop lays out the broader analytical framework behind this argument: how I map the AI stack, where the real physical and technological bottlenecks sit, how power moves across those layers, and why understanding AI increasingly requires understanding the geopolitical system underneath it.

The piece explains the thesis. The map shows you the system behind it.

At Małaszewicze, on Poland’s border with Belarus, freight trains arriving from the east stop for a reason that predates almost everything they carry. Europe’s standard railway gauge is 1,435 millimetres. The former Soviet network uses 1,520. Containers travelling overland from China therefore have to be lifted from one rail system and placed onto another before continuing west.

That’s why I’ve been working for months on the Atlas of AI.

If you are an existing paid member of The Business Engineer, please reply to this email to gain access to the AI Atlas.

The difference dates back to decisions made in the nineteenth century. Russia’s first major trunk railway adopted a wider gauge for engineering and commercial reasons that belonged to its own time. The familiar story that it was chosen deliberately to slow an invading army is largely mythology, although the difference later did complicate military logistics. The important point is simpler: the original reason disappeared, but the standard did not. Nearly two centuries later, cranes are still moving containers because of it. The reason expired. The standard remained. Eventually, the standard became geography.

That is one of the ways technology becomes geopolitical. Not because technology determines history on its own, but because certain technologies turn decisions into infrastructure, infrastructure into standards, and standards into dependencies that can survive the circumstances that created them. Once enough capital, regulation and operating practice accumulate around a technical choice, changing it becomes expensive enough that the choice stops looking like a choice at all.

This is why I think calling what is happening today an “AI race” is too narrow. Artificial intelligence is not arriving by itself. It is colliding with two other technological domains, space and biology, while all three increasingly draw on the same underlying resources: compute, electricity, specialised manufacturing, physical infrastructure and government permission. That intersection is the more important thing to watch.

A single important technology can create an industry. Several technologies converging on the same physical and institutional bottlenecks can begin creating a map. I call that a junction.

What Makes a Technology Geopolitical

Most technologies never become geopolitical in any meaningful sense. The spreadsheet transformed business. The shipping pallet transformed logistics. The zipper transformed manufacturing. None substantially changed which states could coerce other states. Economic importance is therefore not enough. Technical sophistication is not enough. Even massive social impact is not enough.

A technology starts entering the geopolitical order when three properties overlap. First, it has to be sited. Some important part of the system must physically exist somewhere, at enough cost and scale that moving it is difficult. Software can move almost instantly. A power grid cannot. Neither can a semiconductor fab, deepwater port, launch complex, enrichment facility or gigawatt-scale compute campus. Once enough capital hardens into concrete, geography returns.

Second, the system has to be standardised. Railways need a gauge. Electricity needs voltage and frequency. Shipping needs container dimensions. Networks need protocols. Semiconductor controls need technical thresholds. Biology needs screening standards. Standards look like engineering decisions while they are being written. Once infrastructure accumulates around them, they determine who fits naturally inside a system and who has to pay to cross its boundary.

Third, the system has to be chokeable. Somewhere in the chain there must be a step that relatively few actors can perform. Whoever controls that step can grant access, price it, condition it or withdraw it.

Siting creates geography. Standards create compatibility. Chokepoints create leverage. Their intersection creates something more than an important market: it creates political power.

There is a second-order effect as well. Once a technological junction becomes large enough, it begins changing which forms of power are cheapest to project. It also changes war itself: who can fight, what can be targeted, how quickly force moves, how deeply military production reaches into the civilian economy and, eventually, whether direct war between the strongest actors remains thinkable.

Four previous junctions built much of the geopolitical world we still inhabit.

Subscribe To Premium To Gain Access!

Check out the Library for Free!

For the last three years, I’ve been rebuilding the Business Engineer’s curriculum from the ground up. That curriculum has now become the foundation of a new discipline, with the entire series taking shape around it.

Subscribe To Premium To Gain Access!

Check out the Library for Free!

If you’re already a paid member, simply reply to this email, and we’ll send it your way.

Read more

There are two possible outcomes: if the result confirms the hypothesis, then you’ve made a measurement. If the result is contrary to the hypothesis, then you’ve made a discovery. — Enrico Fermi

Life is wonderful because it is made of discovery. And to me, that is exactly the point of analysis. Analysis is not about confirming the hypothesis you started with. The real pleasure comes when the evidence forces you to change your mind, when the result contradicts what you expected and reveals something you had not seen before.

I have spent my entire professional life analyzing businesses, partly out of passion, partly as a business executive and educator. What has always fascinated me is going beneath the surface to understand the structural patterns that explain what is actually happening. Not merely whether a company beat or missed, whether a stock went up or down, or whether the consensus narrative happened to be right for a quarter. I am interested in what changed underneath: where bargaining power moved, which bottleneck disappeared, which one replaced it, who is financing whom, what is durable, what is temporary, and where the economics of an industry are actually migrating.

That is where analysis becomes discovery. And this is where I stand.

The Map of AI is my most rigorous attempt to track where we actually are in the AI Supercycle. Not where the hype says we are. Not where the skeptics want us to be. But where the underlying system is moving.

Because the real question is structural, and answering it requires more than choosing a side. The AI debate has increasingly collapsed into two analytical shortcuts. One treats the build as largely speculative and economically unjustified. The other treats AI’s transformative potential as sufficient justification for almost any level of spending attached to it. Neither position is analytical. And neither is particularly useful.

They do not tell us whether demand is real, where the bottlenecks sit, which margins come from durable market power and which come from temporary scarcity, how much of the build is financed from internal cash flow versus external capital, whether falling inference costs are compressing economics or expanding usage faster than prices decline, where application-layer rents are forming, or through which financial joints stress would propagate if demand weakened.

I subscribe to neither camp. My job as an analyst is to deconstruct the system. To separate structural change from speculation, infrastructure from narrative, durable economics from temporary scarcity, technological transformation from financial excess, and the failure of one node from the failure of the cycle itself.

The framework has to be capable of holding two ideas at once: a technology can be genuinely transformative while parts of the financing structure around it become overextended. If the evidence moves toward a generalized bubble, the map should show it. If the underlying system continues to strengthen despite local financial failures, the map should show that too.

The objective is not to defend a thesis. It is to keep updating the map as reality changes.

And this quarter changed the map.

Subscribe & Get Full Access To Map of AI

Get The AI Supercycle Book Here!

Download The Full Map of AI

Full Harness with The Executive Plan

If you’re already an Executive Plan member, simply head to Business Engineering to access the harness.

The map is not the territory. It is the instrument.

The easiest mistake in analyzing AI is to treat it as a single industry. It is not. The AI Supercycle stretches through nine economic strata, beginning with energy and the physical plant, then equipment, foundry and memory, silicon, networking, compute, models, the routing fabric, the agentic harness, and finally distribution and settlement.

Around that stack sits something that should not be treated as another layer: governance. Export controls, sovereignty, data rules, antitrust, access and industrial policy form a fence around the entire system. They can alter the flow of silicon, capital, models and data without belonging neatly to any one technological layer.

Inside the stack, two structures increasingly matter more than the individual companies printed on it. The first is the seam, where access and settlement concentrate.

The second is the routing junction, where models are selected, enterprise context is held, tools are connected, permissions are enforced, agents are governed and, increasingly, money changes hands. That distinction becomes crucial later because economic power does not necessarily stay with the layer producing the most intelligence. Models can be rented. Junctions can be owned.

But a layer map alone is still not enough. The system moves on four clocks, and the clocks do not agree. The physical clock turns in years. It measures how binding the actual constraints are through utilization, lead times, contracted versus available capacity and the price of the scarce input. It currently reads 78/100. The financial clock turns in quarters. It measures how much strain has migrated into debt, off-balance-sheet structures, receivables, vendor finance and the broader funding architecture. It reads 64/100.

The efficiency clock turns in weeks. It measures how quickly the cost per unit of capability is falling through model architecture, active-parameter ratios and the amount of intelligence bought per dollar. It is the hottest clock at 82/100. The adoption clock turns in months. It measures how far real deployment has actually gone: countable revenue against the hurdle, contracted agentic value, deployments per quarter and the migration from seats toward actions. It remains the youngest at 31/100.

These are directional indicators, not one common risk scale. A high efficiency reading means capability is becoming cheaper quickly. A high financial reading means funding strain is increasing. What matters is how the clocks interact.

That gives us one of the central asymmetries of the entire cycle. Physical supply turns in years. Financial conditions turn in quarters. Efficiency can reprice in weeks. Adoption happens somewhere in between. Any company building physical capacity today is betting that the demand shape visible now survives for the years required to deliver that capacity, even though the technology consuming it can change in weeks.

There is then a second instrument focused specifically on the financing of the build. Six gauges ask different versions of the same question: how much of what has already been promised can operating cash cover; how much demand is genuinely contracted; how dependent the build is on external capital; how much short-duration money is financing long-duration assets; how much of the structure sits in vehicles rather than directly on balance sheets; and how fast that funding mix is changing.

Many curves, one climb

This leads to the premise underneath the entire map. There is no single AI curve. There are product cycles that move in quarters, platform cycles that move in years and infrastructure cycles that can stretch across decades. Inside them sit individual companies and technologies, each on its own S-curve.

A node can overshoot and correct while the larger structure keeps climbing. A model company can lose pricing power while application adoption accelerates. A neocloud can refinance under pressure while ASML and TSMC continue seeing the same underlying physical demand. Memory prices can rise while inference prices fall. A builder can become financially stressed even as the infrastructure it owns remains strategically valuable.

That is why I resist the phrase “the AI bubble” in the singular. The falsifier is much more demanding. If unrelated layers began drawing down together across technology, demand and capital, the supercycle thesis itself would have to be reconsidered. But after more than twenty months of corrections, the failures we have seen remain predominantly local: a node, a layer, a funding structure.

Local violence. Structural climb.

Read every company print in this analysis as one node’s position on its own curve, not as a verdict on the entire cycle.

1846: the railway mania is the right rhyme

This is where history becomes useful, not as decoration, but as a way to separate variables. For most of the AI boom, the lazy analogy has been the dot-com era. The comparison is not useless. Transformative technologies can obviously be catastrophically mispriced. The internet survived 2000. Many of the financial claims written around it did not.

But AI is no longer primarily a software valuation story. It is a physical infrastructure story. That is why 1846 and the British railway mania increasingly look like the more useful frame.

At the peak, British railway investment approached roughly 7% of GDP. More track was authorized in a tiny window than the country would need for years. Capital ran far ahead of realized returns. Promoters oversold. Investors speculated. The panics of 1847 and 1857 came, and the crash of 1866 followed. The losses were real. And the railway stayed. It repriced the cost of distance for a century.

That is the part of the analogy worth holding onto. The mania funded the network. The panics transferred its ownership. The network repriced commerce anyway. An infrastructure supercycle does not require the original financiers to survive. The assets can outlive the capital structure that created them.

And out of the wreckage of that railway era emerged institutions that reorganized and refinanced the network. One of them was the House of Morgan. Keep that detail. It will matter again at the center of this map.

1947: the control experiment

The railway explains the financing. But it does not explain what happens when the cost of the underlying technology collapses. For that we need another historical rhyme: 1947 and the transistor.

The transistor began as an expensive, barely commercial component. Then its cost fell from dollars per unit toward fractions of a cent and eventually disappeared into virtually everything. But the early cost curve had an underwriter. The first customer wore a uniform.

The US state could buy the top of the curve through programs such as Minuteman and Apollo because price was not the primary constraint. Capability was. The state was patient, strategic and relatively insensitive to the economics that would have stopped a purely commercial buyer.

That is the control experiment. The technology became cheaper. Demand expanded. Eventually the component became invisible because it was embedded everywhere. The interesting variable is not simply that the cost curve falls. It is who underwrites demand while it falls.

In the first silicon age, much of that underwriting was public. This time, the experiment is being rerun with a different underwriter. The underwriter is increasingly private, and it sits inside the map.

That is why the financing architecture matters so much.

Two stacks, one fence, three buyers

There is another difference from previous technology cycles: AI is no longer developing as a single global industrial stack. The US-anchored stack runs broadly through the leading equipment makers, TSMC, NVIDIA and AMD, the hyperscalers and the frontier labs. The China-anchored stack increasingly runs through SMIC, CXMT, Huawei and a powerful open-weight ecosystem of its own.

Between them sits the governance fence. But the demand side is also changing. There are now three buyers.

The first is the hyperscaler. It can build ahead of perfectly observable demand because compute capacity itself is strategic. Losing access to the infrastructure can be more dangerous than temporarily earning a poor return on it.

The second is the enterprise. This buyer is much more economically disciplined. It buys when AI changes a workflow, reduces cost, creates revenue or becomes sufficiently embedded in operations to justify the transition.

The third is the state. And the state can be return-insensitive. Sovereign AI programs can keep buying strategic compute even when the conventional corporate return equation weakens. This gives the physical floor an important new support.

But it is essential not to confuse two things. The state can set a floor. It cannot clear the commercial hurdle for the rest of the system.

Why compute demand has this shape

The next piece of the map is mechanical. Every AI dollar ultimately passes through three stages: training, prefill and decode.

Training is enormously expensive, but its cost can be spread over vast quantities of future usage. Prefill is highly parallel and comparatively well matched to GPU architectures. Decode is different. It is sequential, heavily memory-bound and therefore pushes directly against HBM, advanced packaging and leading-edge process capacity.

And then reasoning changes the amount of work. A basic chat interaction can represent roughly one unit of compute. A reasoning workload can require 5–20 times more. An agentic workflow can move toward 50–200 times because the system loops, calls tools, invokes subagents, retrieves memory, evaluates intermediate outputs and keeps working without waiting for the human to prompt it again.

This is why the efficiency argument is so easy to misunderstand. Yes, models get cheaper. But cheaper models create more usage. This is the Kimi Paradox: the efficiency clock cuts the cost per token while total spend rises because cheaper intelligence expands consumption faster than it reduces unit price.

And the most compute-intensive new use cases happen to concentrate demand exactly where the physical system is hardest to scale. Commoditization is simultaneously the router’s tailwind and the physical floor’s order book.

The buyer behind all of this is increasingly visible. Three closed labs alone represent roughly 58% of countable AI revenue in the framework. The labs take about a third of each year’s incremental compute capacity, heading toward half, and a remarkable portion of their compute is still pointed inward at research rather than external customers.

They monetize through four different spigots: API tokens, consumer subscriptions, enterprise and agentic contracts, and cloud reseller rails. Yet the engine also has to finance its own successor. The current model earns the money required to train the next model.

That is the demand structure against which every subsequent part of the map has to be read.

Subscribe & Get Full Access To Map of AI

Get The AI Supercycle Book Here!

Download The Full Map of AI

Full Harness with The Executive Plan

If you’re already an Executive Plan member, simply head to Business Engineering to access the harness.

THE FLOOR

One road, three institutions

Once you move downward from the labs, the system becomes physical very quickly. There is essentially one road that frontier AI compute has to travel, and three institutions now define different parts of it.

ASML is the mint. It makes the machines required to produce leading-edge chips. Its FY26 outlook has been raised toward roughly €43–45 billion, EUV is reaccelerating and memory equipment demand has surged. This is a seven-year clock.

TSMC is the Fed. It allocates the wafers and runs the fabrication system at scale. HPC and AI are now around 61% of revenue, CoWoS remains a binding tell, and capex has been increased into the same demand window. This is closer to a five-year clock.

Then comes NVIDIA, increasingly the bank. At the end of the road, it does more than sell the traffic the road carries. It is increasingly helping finance pieces of the ecosystem consuming it.

That combination is analytically powerful. Two independent institutions with extraordinary market power, ASML and TSMC, are looking at the same physical demand curve from different positions and raising capacity expectations into the same window. That makes the claim that the underlying demand is simply fictitious increasingly difficult to sustain.

But it does not tell us whether every dollar of that demand will ultimately earn an acceptable return. That is a different question.

The zero-sum wafer

Memory shows why the distinction matters. A wafer is a scarce object. Once capacity is directed into one use, it cannot simultaneously serve another. The economics of HBM have therefore reshaped the memory business.

The conversion is effectively zero-sum. Redirect capacity away from consumer endpoints and into HBM, and the economics of the same underlying wafer can change dramatically. In the current framework, the wafer requirement can translate into roughly a 3:1 trade-off, while average selling prices have risen on the order of 30–85% against comparatively modest bit growth.

The result is extraordinary operating economics. SK hynix, Samsung Memory and Micron have been printing margins associated with a market under severe tightness.

But then something important happened. Memory companies increasingly wired the future through long-duration take-or-pay structures. The map counts sixteen take-or-pay agreements, roughly $100 billion of minimum revenue, around $22 billion of deposits, and price floors extending toward 2030.

At that moment, the risk mutated. The question stopped being Will someone buy the memory? It became Will the counterparty still be able to pay for the memory it already promised to buy?

Market risk became counterparty risk.

That is a recurring pattern in this supercycle: once scarcity gets contractualized, the economics look safer precisely because the risk has migrated somewhere else.

The memory race is a race over distance

HBM is only one part of the memory story. Inference is memory-bound, so the architecture increasingly becomes a competition over how close the data can sit to the computation.

On-die SRAM is extraordinarily fast and extraordinarily expensive. HBM sits beside the die and offers a compromise, but it is itself a bottleneck because of wafer consumption and packaging complexity. Conventional capacity DRAM sits farther away: cheaper, slower, but increasingly important as agents require persistent context stores rather than simply weight stores.

This creates two different bets. The incumbent path is to stack HBM higher. Every generation widens the memory system, but the trade ratio becomes more demanding and cost per useful bit can rise after decades of decline.

The alternative is to move toward more memory directly on the die, as wafer-scale architectures attempt to do, bypassing pieces of the HBM and packaging bottleneck altogether.

Agents add yet another layer to this race. A model does not simply need somewhere to store its weights. An agent needs somewhere to store its world: context, previous work, tool state, intermediate outputs and memory.

Cheaper models do not necessarily reduce memory demand. They redistribute it.

Three roads to an AI chip

The chip market itself is therefore splitting into three routes. The first is the merchant road, dominated by NVIDIA and AMD. The second is the hyperscaler road, where Google, Amazon, Microsoft, Meta and others increasingly design their own silicon.

But building your own chip does not make the physical gates disappear. The hyperscaler still needs fabrication. It still needs memory. It still needs packaging. It still needs the leading edge.

Custom silicon reroutes the spend. It does not remove it.

Then there is a third route: architectures that attempt to bypass portions of the conventional stack entirely, particularly through wafer-scale approaches and radically different memory choices.

That bypass matters because it changes the strategic behavior of the incumbent. If customers have plausible ways to design around your strongest bottlenecks, financing them can become rational.

The strongest evidence of how fluid this has become is that one anchor frontier lab can now operate across multiple substrates: TPU, Trainium, rented neocloud capacity, NVIDIA and AMD.

The floor remains scarce. But the roads across it are multiplying.

Subscribe & Get Full Access To Map of AI

Get The AI Supercycle Book Here!

Download The Full Map of AI

Full Harness with The Executive Plan

If you’re already an Executive Plan member, simply head to Business Engineering to access the harness.

WHO PAYS FOR THE BUILD?

There are only five places the money can come from

On the framework’s current estimate, the AI build is running at roughly $750 billion a year. That money has to come from somewhere.

There are only five sources, and they become progressively more fragile as you move down the list. The safest is own cash flow. Alphabet, Microsoft, Meta and Amazon can use profits generated elsewhere to finance infrastructure.

Then come the capital markets: bonds, private credit and off-balance-sheet vehicles.

Then the customer in advance: backlog, prepayments, contracted revenue and take-or-pay arrangements. Then the supplier: vendor equity, receivables and guarantees. And finally everyone else, when the bill leaves the AI site through memory prices, higher power bills, cost inflation and the price of products containing the constrained components.

The most important change this quarter is not that any one of those sources appeared for the first time. It is that the money moved down the list. A year ago, much more of the structure could be read through the first two rows. This quarter, the bottom three began doing visibly more work.

One build, eight different financial realities

This is why simply quoting aggregate capex is almost useless. Put eight major builders on one instrument, with capex as a percentage of revenue as the fluid level, and the strategies look nothing alike.

Apple sits around 2%. Tesla around 20.5%. Amazon around 27%. Microsoft around 35%. Alphabet around 46%. Meta around 54%. Oracle around 86%. Then SpaceX arrives at roughly 235%, and the instrument breaks.

But even these percentages do not tell the whole story. The drain valve is free cash flow. Apple is essentially off the clock. Microsoft remains self-funded. Alphabet is deliberately stretching. Meta cracked its previous self-funding model. Oracle is pulling almost every available channel. SpaceX simply sits outside the conventional scale.

The strain never disappears. It migrates.

Sometimes into finance leases. Sometimes into debt. Sometimes into customer commitments. Sometimes into investment marks. Sometimes into another company entirely.

Alphabet: the Absorber, and the cage it built

Alphabet might offer the strongest absorption arithmetic in the group. Cloud revenue is growing roughly 82%, margins have expanded from approximately 20.7% to 35.6%, backlog has reached around $514 billion, yet free cash flow has moved to roughly negative $5.9 billion.

That is not a weak company losing control of its economics. It is a fortress deliberately stretching its balance sheet to become an infrastructure company.

But Alphabet has another problem, and it is more interesting. It built the TPU as an internal advantage. Then the TPU became commercially valuable enough to sell externally. Anthropic alone is associated with roughly 3.5 gigawatts of capacity, alongside several other customers.

Now Alphabet’s own frontier researchers increasingly compete for access to the capacity the company sells. The frontier model is effectively queued inside an inference cage made of TPUs.

At the same time, Alphabet reportedly rents roughly $920 million a month of NVIDIA capacity to bridge-train its own frontier.

The V8 architecture, split into distinct training and inference chips, can therefore be read backward as an admission: once proprietary infrastructure becomes a successful merchant product, internal advantage and external monetization begin competing with each other.

Alphabet built the cage. Then it discovered that the market wanted to rent the bars.

Microsoft: absorbed above the waterline

Microsoft remains the cleanest large-scale example of the build being funded from the existing operating engine. Capex is roughly 63% of cash from operations. Free cash flow remains strongly positive, around $67 billion in the framework. Buybacks increased by roughly $22 billion, and no new bonds were needed. Azure continues growing rapidly and RPO has expanded toward roughly $678 billion.

On the surface, the structure looks remarkably strong. But the strain did not disappear. It moved below the waterline.

Microsoft carries around $196.6 billion of leases already signed but not yet commenced, while additional finance-lease PP&E and long-duration commitments continue growing.

This is an important accounting lesson for the cycle. You cannot understand the build by reading the capex line alone. Infrastructure commitments can migrate into leases and future obligations without changing the underlying economic promise.

Microsoft has not cracked. But the test has been named. If capex eventually crosses cash from operations, or if financing moves meaningfully into new debt, the financial clock has changed.

Meta: the cistern and the shut valve

Meta is where self-funding visibly became more difficult. Revenue grew roughly 28%, but operating income moved the other way. Capex approached 98% of cash from operations. Free cash flow collapsed. Buybacks went to zero. And roughly $24.9 billion of debt arrived in a single quarter.

This is why I do not describe Meta as having overshot demand. The demand remains visible. What Meta overshot was its ability to self-fund the capacity required to serve it under the previous financial structure.

The distinction is fundamental. A company can be right about the technology and still arrive at the wrong capital structure.

Meta retains a shut valve. It can slow capacity. It can sell surplus. It can turn Meta Compute into an external spigot if Gemini or others need that capacity.

But the quarter marks the point where the financial clock entered the core.

Amazon: the split ledger

Amazon produces one of the cleanest examples of why operating economics and cash economics have to be read separately. The AWS operating page is exceptional: revenue growth around 37%, margin expansion from roughly 32.9% to 39.4%, operating income up about 43%.

Then turn the page. Trailing free cash flow descends quarter after quarter and eventually becomes negative, around -$7.6 billion in the framework.

Same company. Two opposite answers.

There is also an important distortion in reported earnings because Anthropic investment marks account for a significant portion of net income. Strip those away and the operating story remains strong, but much less spectacular.

That is not distress. It is placement. Amazon is strong enough to absorb the build operationally while simultaneously allowing the cash page to deteriorate.

Two houses beside the build

The control group may be even more interesting than the builders. Apple has largely skipped the physical arms race. Capex remains below roughly 2% and falling. Free cash flow is around $110 billion. Rather than internalizing the entire build, Apple can rent intelligence through a meter, spending on the order of a billion dollars a year and treating the risk as a contract rather than a balance-sheet commitment.

The market has valued that position at roughly $5 trillion. That is worth remembering.

Opting out cleanly is a strategy.

Tesla illustrates the opposite predicament. It sits beside a gigantic neighboring build while running about 20.5% capex intensity on roughly a 1.4% margin, and a remarkable share of reported earnings comes from the mark on its sibling.

The distinction is brutal. Apple rents the infrastructure. Tesla is financially exposed to the appreciation of a nearby infrastructure machine while building on industrial economics. And the sibling, SpaceX, has already out-earned it.

The SpaceX flywheel

SpaceX deserves its own category because it reveals an economic structure toward which several winners may ultimately converge. The flywheel begins with launch as the cash-generating engine. Alongside it sits Colossus as the compute engine. Compute trains Grok as the model. The model powers the harness, increasingly embedded in distribution and coding workflows. Then the exhaust returns: usage becomes training signal.

This should be read as an economic flywheel, not a literal treasury waterfall. The point is that the same corporate system increasingly owns the cash engine, compute, model, distribution and usage feedback loop.

Most builders assemble pieces of this loop from infrastructure they rent. SpaceX increasingly owns more of the stations around the loop itself.

This matters because the vertical-integration advantage is not simply cost. It is co-adaptation. The model changes the harness; the harness changes how the model is used; usage changes the model again.

The flywheel is the product.

Follow one dollar of capex

Now follow one dollar from the builder’s budget until it becomes a running data center. Different companies touch that dollar, and they are paid in radically different ways.

Lumentum and Vertiv collect rent because optics and power are scarce. Margins expand and balance sheets remain comparatively clean. Supermicro carries the float. It assembles the racks and fronts enormous amounts of working capital while operating on much thinner gross margins. CoreWeave carries the leverage. It deploys machines with borrowed money; interest expense approaches a quarter of revenue against an enormous backlog. Equinix keeps the deed. It owns the building underneath the whole arrangement.

There is a structural rule here. The closer you sit to a genuinely scarce input, the fatter the margin and cleaner the balance sheet. The closer you sit to assembly, the more cash you front. The closer you sit to deployment, the more you borrow. And the landlord keeps the collateral.

Rent can be competed away once capacity catches up. The deed cannot.

Same build, opposite balance-sheet sign

CoreWeave and Nebius then give us an almost controlled experiment. Same broad market. Same rented-compute layer. Opposite financing sign.

CoreWeave carries net debt against a contracted backlog of roughly $129 billion. That backlog represents future customer payments rather than cash already collected, while the capacity has to be financed today through public and private capital. Interest expense is roughly 25% of revenue.

Nebius is different. It sits on net cash and a substantial liquidity cushion because part of its future service obligation has effectively already been funded by the customer. Roughly $5.98 billion of customer prepayments turns the customer into a source of financing.

CoreWeave finances today against future customer commitments. Nebius has already collected part of the money.

The risk does not disappear for Nebius. It migrates into customer concentration, depreciation and the obligation to deliver future compute. But the balance-sheet sign is fundamentally different.

And CoreWeave gave us something even more useful: the first live test of one of the financing joints beneath the cycle. GPU-backed private credit reached a refinancing moment. It did not crack. Financing migrated into syndicated and public unsecured markets.

The joint survived. The risk changed owners.

The House of Morgan

And then NVIDIA printed. Until this quarter, NVIDIA could still be described primarily as the purest beneficiary of the AI build. The builders spent. NVIDIA got paid.

That description is now incomplete.

The operating engine remains extraordinary: roughly $96.2 billion of revenue, +106%, and $63.7 billion of operating income in the latest print. But the operating engine is not the discovery. The balance sheet is.

NVIDIA increasingly helps finance the ecosystem buying its products through four channels. First, customer credit: receivables expanded sharply, while collection days moved higher. Second, equity investments in buyers and adjacent infrastructure companies: the investment book has grown materially, leaving NVIDIA with direct economic exposure to companies whose success depends in part on continuing to build AI infrastructure.

Third, power and infrastructure support, where guarantees and related structures can pull NVIDIA deeper into the physical build required to operate its chips. Fourth, capital-market structures around the ecosystem, as a widening syndicate of private-capital institutions finances the data centers, power assets and compute companies sitting around the GPU.

Put the geometry together and the change becomes visible. NVIDIA is no longer merely supplying the critical equipment. The supplier is extending credit into the ecosystem, owning pieces of companies that consume the equipment, and helping support the infrastructure required to use it.

This is why the railway analogy becomes more than historical color.

NVIDIA is increasingly playing the House of Morgan role in the AI build: not only supplying the critical equipment, but helping finance the ecosystem buying it.

The analogy should not be taken literally. Morgan reorganized and financed railroads; NVIDIA manufactures the locomotives of this system as well. That makes the structure, if anything, more unusual: supplier and financier are increasingly sitting on the same economic map.

And the financial wrapper around the build is widening. Hundreds of billions of dollars of institutional capital now sit around the system through Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR and others. Whether those structures genuinely distribute risk or merely move the same exposure farther from the original supplier depends on terms that remain only partially observable.

But the center has moved.

The beat raises the hurdle

There is a reflexive consequence to all of this. Every spectacular NVIDIA quarter is evidence for the infrastructure-demand thesis. But every additional chip shipment is also capital that someone farther up the stack has to earn a return on.

NVIDIA’s data-center revenue alone now annualizes above roughly $356 billion. Run the incremental shipments through the return framework and they add something like $64 billion a year of required end-customer revenue. The application layer is adding closer to $45 billion a year.

So the gap can widen even during an extraordinary quarter.

This is the paradox at the center of the build: the better NVIDIA performs, the more the rest of the stack has to monetize.

Sovereign demand can soften this because part of the infrastructure is bought for strategic rather than purely financial reasons. But again, the state sets a floor. It does not clear the hurdle.

Countable AI revenue is currently around $185 billion. The hurdle is already above $600 billion. The beat raised it again.

Subscribe & Get Full Access To Map of AI

Get The AI Supercycle Book Here!

Download The Full Map of AI

Full Harness with The Executive Plan

If you’re already an Executive Plan member, simply head to Business Engineering to access the harness.

THE CAPITAL MACHINE

Where the $602 billion comes from

The $602 billion figure is not a forecast. It is a base-case estimate of the annual AI revenue required to economically support the infrastructure already deployed.

Start with roughly $1.27 trillion of deployed AI infrastructure across chips, racks, data centers and power. That capital has to do two things every year: absorb the economic depreciation of the assets and earn a return for the capital that funded them.

Using a 7.3-year blended economic life implies roughly 13.7% annual capital consumption. Add a 10% required return on invested capital, and the infrastructure carries an annual capital charge of about 23.7%.

On $1.27 trillion, that is roughly $301 billion per year.

The next step is to translate that capital charge into revenue. If roughly 50% of end-customer AI revenue is ultimately available to cover depreciation and the required return on capital, then the system needs approximately twice the capital charge in revenue:

$301B annual capital charge → roughly $602B of annual revenue.

Today, the map identifies around $185 billion of end-customer AI revenue, with a reasonable range of roughly $165–205 billion. On the base case, current monetization therefore represents about 31% of the $602 billion hurdle.

The key point is not that AI lacks revenue. It clearly does. The point is that the infrastructure has been built ahead of the monetization required to support it economically.

And the hurdle is still moving. More accelerators are being installed, more data centers are being built and more power is being contracted. Every additional dollar of deployed capital increases the amount of revenue the system will eventually need to support.

So the question is straightforward:

Can AI monetization grow fast enough to catch a capital base that is still expanding?

Ai Infrastructure Revenue Hurdle Model (1)

19.5KB ∙ XLSX file

Download

Download

*Base-case assumptions: $1.27T deployed capital; 7.3-year blended economic life; 10% required return; 50% contribution margin available to service the capital base. The $602B figure is a scenario hurdle, not a revenue forecast. The two anchor inputs also matter. The $1.27 trillion deployed-capital estimate and the $185 billion end-customer revenue estimate should be read as model inputs built from the underlying company and infrastructure data, not as reported industry accounting figures. The methodology and accompanying model are therefore designed so those inputs can be challenged separately from the arithmetic built on top of them.

Revenue per megawatt

The most intuitive way to understand that conversion is through a megawatt. A megawatt of compute can cost roughly $10–15 million a year. The leading labs can now generate blended revenue approaching $50 million per megawatt.

That is the engine. Buy one megawatt. Turn it into intelligence. Sell the intelligence at a multiple of the underlying compute cost. Use the margin to finance the next model.

This is where the labs begin looking less like venture-funded research institutions and more like actual economic engines.

The gross-margin path has been violent. Economics that looked completely unsustainable two years ago have moved toward 50–55% gross margins, and the strongest labs have begun showing positive operating economics.

That is still structurally worse than conventional software, where gross margin can sit around 77%. But it can be profitable. And a large part of the gap between gross margin and operating margin is deliberate spending on the next training cycle.

In other words, the model funds the factory that builds its successor.

There is also a deeper strategic consequence of open weights. Once open models and self-hosted infrastructure can earn a return at the base price of compute, frontier labs have to outbid an increasingly credible alternative for scarce megawatts. The efficiency race therefore becomes part of the capital-allocation race.

Where it would actually break

Everyone looks for the failure in the layers. The more interesting failure points sit in the joints.

A joint is where one company’s promise becomes another company’s collateral.

The first joint connects the lab and its landlord. A lab signs a long-duration compute commitment. The neocloud borrows against that signature to build the site. That joint has already been tested. It refinanced. Private secured borrowing migrated into public unsecured debt. The financing became more expensive, but the joint held.

The second joint connects power and compute. Increasingly, the megawatts and the machines sit inside the same contractual structure or SPV. If one side fails, we still do not know exactly what the other side does. That joint is forming. It remains largely untested.

The third joint connects the memory floor with the financial ceiling above it. Memory producers have protected themselves with multi-year take-or-pay commitments, while the companies farther up the stack remain dependent on monetization and credit. Those curves can diverge for a while. The real test may not arrive until 2027, when the long-dated orders eventually have to be revised.

Then there is a jack underneath the entire footing: debt secured against shares in private AI companies. If the mark on that collateral falls, the borrower can be required to post real cash against a value that itself is only periodically updated.

That produces one of the most uncomfortable sentences in the deck: the equity is debt; the collateral is an estimate.

This is why contagion does not necessarily run sideways through adjacent technological layers. It can run upward through the financial joints. The weakest tenant first. The structure financed against the tenant second. The layer above it next.

The bright streetlight

This quarter, the build also crossed another boundary. The central bank began speaking about AI capex as an inflation source in the same macro conversation as war and tariffs.

That is important. But a streetlight only tells you what it illuminates. Chip and memory inflation are visible. The harder question is whether they are merely the lit patch or evidence of a much larger economic effect.

The answer increasingly looks like the latter.

AI-related investment is growing at roughly 20% over four quarters in the framework. The bond market has already been repricing around the build’s issuance. The debt share of capex has moved from approximately 9% to 32% in eight quarters.

The build has become big enough to affect national investment, the bond market and eventually the policy rate itself.

That closes a loop. The build bids for chips, memory, power, transformers, labor and land. Prices rise across the broader economy. Rates stay higher because inflation remains harder to suppress. Financing costs rise for the same AI infrastructure absorbing the capital.

The build therefore begins paying its own inflation.

That changes several things at once. The discount rate is no longer an entirely external variable. The build itself helps influence the rate at which the build gets discounted. Capital-market funding becomes more expensive exactly when the infrastructure requires more of it. And a capex cycle that moves consumer prices begins attracting political attention, not merely analyst attention.

At that point AI is no longer simply happening inside the macroeconomy. The macroeconomy begins responding to AI.

Subscribe & Get Full Access To Map of AI

Get The AI Supercycle Book Here!

Download The Full Map of AI

Full Harness with The Executive Plan

If you’re already an Executive Plan member, simply head to Business Engineering to access the harness.

THE MODEL LAYER

Anatomy is economics

We can finally move above the physical and financial plant and into the model itself. A frontier model is not one dial. It is a combination of scale, sparsity, post-training, context and distillation.

Scale determines the total parameter base and pretraining investment. Sparsity determines how much of that model actually activates for each token. Post-training and reinforcement learning can move capability in discontinuous jumps. Context determines what the model can hold. Distillation transfers capability downward into smaller systems.

What interests me is that every one of those technical choices is simultaneously an economic choice.

Sparsity is the cleanest example. If only two out of eight experts fire for a token, you can preserve an enormous parameter base while drastically reducing the compute activated for each unit of work.

That is intelligence per watt.

And that is why model anatomy belongs on a financial map.

Some open models can now produce scores approaching frontier systems on selected benchmarks while activating only a fraction of their total parameter base and delivering dramatically lower inference costs. The exact cost advantage depends on workload, hardware, serving architecture and benchmark, so the percentage should never be treated as universal.

But the direction is clear:

The technical architecture is doing financial work.

Three ways to get a model

That changes the enterprise question. A year ago, “Which model should we use?” often sounded like one question. It is now at least three.

The first path is to rent it. Call a closed frontier model through an API. You get immediate access to the strongest capabilities. You avoid the infrastructure burden. The price is metered. But you do not own the asset. You have no title to the underlying weights, and the provider sets the price.

The second path is to host it. Take open weights and run them yourself. You gain operational control and potentially much lower cost per action. But the infrastructure and inference burden becomes yours, and capability may still lag the absolute frontier. You have operational control. You still do not have complete title to the intelligence architecture.

Then comes the third path: co-design it. Take open weights and fit them around your own harness, data, workflows and operating core. This is slower. It requires time and talent. It may take one or two years to become truly enterprise-native. But it is the only path that ends with something closer to clear title.

That changes the enterprise decision from Which model should I buy? to Which intelligence should I own at the core?

That may become one of the most important enterprise questions of the next few years.

The challengers: the gate and the tide

The frontier itself is also becoming easier to analyze. There is a gate. To reach it, a challenger needs some combination of gigawatts, capital, talent and distribution.

OpenAI, Anthropic and Google remain on the frontier plateau, operating at a scale measured in multiple gigawatts. Below them, new entrants are climbing with different tickets: xAI, Meta, NVIDIA, ByteDance and others.

But underneath the entire hill there is also a tide. Kimi. DeepSeek. Qwen. GLM. Nemotron.

The open ecosystem is repricing the frontier from below.

Selected open models can increasingly approach frontier benchmark performance while activating far fewer parameters and operating at a fraction of the cost on specific workloads. The exact gap varies enormously by benchmark and deployment, but the strategic dynamic is more important than any single percentage.

The gap is the moat. The tide is the erosion.

The frontier does not need to stop improving for its economics to compress. It only needs the open cohort to close enough of the gap that the marginal buyer no longer pays the old premium.

So the correct question about challengers is not simply how much compute they own. It is what they can do with adequate compute and how quickly those capabilities diffuse.

The harness turn

This is why the next race moves above the model. The competitive unit is increasingly not the model itself. It is the model plus the harness.

Memory. Routines. Tools. Subagents. Workflow state. Enterprise context. Permissions.

The harness changes the economics because it produces something the model provider cannot simply buy: usage exhaust.

The user works through the system. The system observes the work. That usage becomes context, routine, preference and eventually training signal. The harness improves the model that powers it. The better model improves the harness. And the loop compounds.

We can already see the pattern in coding agents, where the fastest-growing revenue lines increasingly belong to harnesses rather than raw token APIs. We can see it in systems where one prompt expands into a standing routine across search, mail, feeds and archives. We can see it at the enterprise junction, where large agentic contracts increasingly pay for orchestration, governance, context and tools while the underlying model remains replaceable.

And we can see it in incumbent enterprise software, where the strategic concession is increasingly to live under the agent rather than insist on remaining the interface the human sees.

That is why the harness matters.

Whoever owns the usage loop owns the exhaust. And the exhaust is the input money cannot buy.

The alliance is the unit of competition

Once models become more portable and the harness becomes more important, the unit of competition changes again. It is increasingly not the firm. It is the alliance.

Microsoft and OpenAI. Amazon and Anthropic. NVIDIA and Palantir. ServiceNow and Anthropic. SAP with multiple vendors. AMD and Anthropic.

Each alliance should be read as an architectural choice. Which layers does the alliance want opened? Which layer does each participant intend to hold? Where is the junction? Who needs the layer beneath them to commoditize? Who is absent from the coalition?

The signature lists are often more informative than the press release. An open-weight coalition containing dozens of participants but excluding OpenAI, Google, Anthropic and Meta tells you something about where the closed frontier believes its value lies.

And absence itself can be a strategic position. Until it ends.

AMD supplied the example this quarter by joining at scale.

There is also a third species emerging inside these alliances: co-design. Customers stop acting purely as buyers and begin acting as participants in the intelligence architecture itself. They bring domain data, workflows and operating context into the build.

That matters because the hand that wires the harness may ultimately decide who owns the junction.

Partner today. Rival at renewal.

Alliance formation leads revenue by quarters.

Subscribe & Get Full Access To Map of AI

Get The AI Supercycle Book Here!

Download The Full Map of AI

Full Harness with The Executive Plan

If you’re already an Executive Plan member, simply head to Business Engineering to access the harness.

THE CEILING

The ceiling finally has a number

For three years, the top of the AI stack was mostly a promise. This quarter, it printed a number: Agentforce reached $1.5 billion of annual recurring revenue, up more than 240% year over year. But the headline needs to be read carefully. Salesforce broadened the definition in the same quarter, folding its broader AI offerings into the metric. On the prior definition, the number would have been closer to $1.2 billion. So the important signal is not simply the size or the growth rate. It is where the revenue landed, and what customers are actually paying for.

It did not land at the model. It landed at the harness: Salesforce’s own description of the trusted enterprise layer that connects data, workflows, logic, actions and governance to every agent, model and interface. Claude, GPT, Gemini and open models sit underneath as increasingly rentable and interchangeable inputs. The economic relationship is therefore becoming clearer: the intelligence can be rented; the orchestration layer can be owned. The money stops at the harness.

The unit of monetization is changing too. Agentforce is increasingly billed through Agentic Work Units, not simply through seats. Salesforce delivered 3.2 billion work units in the quarter, up 97% quarter over quarter, and 7.0 billion cumulatively. That matters because it gives us one of the first large-scale public examples of the transition from software priced around human access to software priced around machine work. The meter is beginning to follow the action.

So I read the $1.5 billion in three ways. First, it is a harness number: customers are paying for orchestration, governance, tools and enterprise context, not for intelligence in isolation. Second, it is metered: the economic unit is increasingly work performed rather than seats filled. Third, it is still only the first rung: $1.5 billion is meaningful proof that the ceiling can monetize, but it remains small against an infrastructure build measured in hundreds of billions.

That is the real signal from the quarter. The ceiling is real, but still thin, and it is becoming more clearly defined: value is landing on the layer that routes and governs the models, while the models themselves increasingly sit underneath as rentable supply.

Above the model, everything is being rebuilt

This helps explain why adoption remains the slowest clock. The model may already be capable. The surfaces above it are not.

In commerce, checkout was built for a human holding a cursor. Agentic commerce requires machine-readable catalogs, settlement rails and systems designed for software to transact.

In enterprise software, systems of record were built for humans to type into. An agentic enterprise requires systems that agents can write to, with governance and audit trails around the actions.

On the web, pages were built for human eyes and advertising impressions. An agentic web needs crawl permissions, per-request pricing, identity for machines and eventually new payment rails.

In the toolchain, software was written by people sitting inside editors. The new substrate increasingly consists of agent harnesses, subagents, memory and persistent routines.

That is why the application ceiling can ultimately be worth much more than a billion dollars while still monetizing much less than the infrastructure base underneath it today.

The demand exists. But the surfaces that carry the demand have to be rewritten first.

And software rewrites take quarters, not weeks.

From SaaS to AGaaS

The business model changes with the architecture. SaaS was built around a simple unit: the seat. A named human used the software. The vendor charged per user, per month. The marginal cost to serve that additional seat was often close to zero. Gross margin around 77% became normal. The interface itself was the moat because humans logged into it.

Agentic software changes all four variables.

The unit becomes the completed action. Pricing moves toward outcomes, credits and consumption. The cost to serve now includes inference, which is real and variable. Gross margin resets lower, and scale has to earn it back. The moat moves away from the screen and toward the junction that routes and governs the work.

That is the transition from Software as a Service to Agents as a Service, AGaaS.

And it is no longer theoretical. Workflow software is already selling agent layers across tens of millions of seats. Design platforms increasingly add AI credits directly to the bill. Commerce platforms monetize settlement rather than simply the user account. Data platforms are experimenting with contracts tied to outcomes rather than licenses deployed.

The unit of software is changing.

What the crossing costs

But incumbents do not get from SaaS to AGaaS for free. Salesforce is now giving us one of the clearest public views of what that transition actually costs.

Agentforce has reached $1.5 billion in annual recurring revenue, while usage is increasingly being measured through Agentic Work Units: 3.2 billion were delivered in the quarter, up 97% quarter over quarter, with 7.0 billion delivered to date. That is the opportunity. But the same numbers expose the transition underneath it. The unit of software is moving from the seat to the work performed, and that fundamentally changes the economics.

A SaaS seat had almost no marginal computational cost. An agentic action does. Every workflow executed by an agent can trigger models, tools, retrieval, memory and other infrastructure underneath it. The revenue becomes metered, but so does the cost. The incumbent therefore has to build a new revenue engine while absorbing inference into the P&L and rebuilding the product around a very different unit of consumption.

There is another cost: the interface itself becomes less defensible. Salesforce increasingly describes its value not as the screen the employee works inside, but as the trusted enterprise harness connecting data, workflows, logic, actions and governance to whatever agent or model sits above it. That is strategically sensible, but it represents a migration of the moat: from owning the human interface to owning the governed system underneath the agent.

And the $1.5 billion headline itself shows how messy the crossing can be. Salesforce broadened the definition of Agentforce ARR in the same quarter it reported more than 240% growth; under the previous definition, the figure would have been closer to $1.2 billion. The signal is therefore not simply that agentic revenue is exploding. It is that the old software categories are already becoming difficult to separate from the new ones.

That is what the crossing costs. The incumbent has to change the unit of monetization, absorb a new variable cost structure, reposition where its moat sits and reorganize the existing product around agents, all before the new economics are large enough to replace the old ones cleanly.

The first mover pays for the transition before it captures the full economics of the destination. But the alternative may be worse: defending the interface while the agent routes around it.

Subscribe & Get Full Access To Map of AI

Get The AI Supercycle Book Here!

Download The Full Map of AI

Full Harness with The Executive Plan

If you’re already an Executive Plan member, simply head to Business Engineering to access the harness.

THE VERDICT

What Q2 actually settled

After going through every layer, the map can finally answer the questions it started with.

Where does the cycle stand? Strained, still funding itself. The financing composite reads 58/100. Financial pressure has risen substantially, but the available evidence does not yet show a generalized funding break.

Is a build this large rational? Potentially, but the most interesting development is that private capital is doing more of the underwriting. Both historical rhymes remain useful: the railway mania explains how infrastructure can survive the financiers that originally funded it, while the transistor explains how falling costs can expand demand dramatically. The variable that changed is the underwriter.

Is the physical floor real? Yes. Memory, advanced packaging, leading-edge fabrication, networking and power continue to show genuine constraints. But not every scarcity rent is a permanent moat.

Who pays for the build? Increasingly, not only the hyperscalers’ operating cash flow. Debt, leases, customer prepayments, take-or-pay structures and supplier financing are doing more of the work. Risk is not disappearing; it is migrating.

Does the capital convert into enough revenue? Partly. The framework currently identifies roughly $185 billion of end-customer AI revenue against a modeled $602 billion annual revenue hurdle for the capital already deployed. That is about 31% of the base-case hurdle, and the target continues rising as new infrastructure is installed.

What happens to models? The unit price of useful intelligence is falling rapidly while several physical inputs required to produce it remain scarce. Open weights, sparsity, distillation and inference optimization continue pushing down the cost of capability.

Where do the rents land above the model? The early evidence points toward the harness and routing layer. Agentforce has reached $1.5 billion of ARR, while Salesforce is increasingly monetizing the enterprise layer connecting models to data, workflows, tools and governance. The emerging thesis is straightforward: models can be rented; the layer that governs the work can be owned.

And there are explicit ways this analysis can be wrong. If anchor customers walk away, contracted backlog stops being a meaningful demand floor. If supplier receivables stop converting to cash, vendor financing changes from evidence of demand into evidence of credit risk. If spreads widen precisely when major borrowers refinance, the financial joints get tested for real. If enterprises bypass the harness and call models directly, the junction thesis weakens. And if capability improvement stalls for a sustained period while infrastructure commitments continue rising, a large part of the demand argument has to be revisited.

Those are falsifiers. Not vibes.

The next prints matter for specific reasons. Oracle will help test one of the most aggressive financing structures. Broadcom offers another read on custom silicon and the center of the compute build. Micron becomes a potential tripwire for the memory joint.

And two larger questions carry into the next season: whether open weights genuinely enter the enterprise as co-designed systems, and whether the harness, not the model, becomes the durable thing customers actually buy.

The compression

The Q2 map resolves into six conclusions.

First, the physical demand is real. ASML, TSMC, HBM, advanced packaging, networking and power are independently showing that the infrastructure required to run AI remains constrained. But not every company benefiting from that constraint has the same moat. ASML and TSMC sit behind capabilities that are extraordinarily difficult to reproduce. Some margins elsewhere may reflect temporary scarcity that eventually normalizes.

Second, the build is becoming more financially dependent. The largest hyperscalers can still fund much of their infrastructure from operating cash flow, but an increasing share of the system now relies on debt, leases, customer prepayments, take-or-pay contracts and supplier financing. That does not mean the structure is breaking. It means more of the risk is moving into counterparties, creditors and suppliers.

Third, NVIDIA is now part of that financing system. This is the most important change in the quarter. NVIDIA is no longer simply selling GPUs into the build. It is extending credit into the ecosystem, investing in companies tied to AI infrastructure demand and participating in structures that help support the power and capacity its hardware requires. The supplier is increasingly helping finance the market consuming its products.

Fourth, the infrastructure is monetizing, but monetization has not yet caught the modeled return hurdle. The framework counts roughly $185 billion of end-customer AI revenue against about $602 billion of annual revenue required to support the capital already deployed under the base-case assumptions. That means current monetization represents roughly 31% of the hurdle. AI revenue is real and growing quickly, but every new round of infrastructure spending raises the amount the application layer will eventually have to earn.

Fifth, the unit price of intelligence is falling while parts of the physical infrastructure remain scarce. Open weights, distillation, better architectures and inference optimization continue reducing the cost of useful capability even as memory, packaging, power and leading-edge fabrication remain constrained. That asymmetry pushes strategic value upward. Enterprises can increasingly rent the model underneath while seeking to own the context, workflows, permissions and orchestration around it.

Sixth, that is where the first meaningful application-layer revenue is beginning to appear. Agentforce has reached $1.5 billion of ARR, but the more important fact is what Salesforce is monetizing: the enterprise harness connecting models to data, workflows, tools and governance. The model underneath can change. The control layer remains.

Put those six observations together and the picture is more precise than either “AI is a bubble” or “AI will justify everything.”

The technology is real. The infrastructure demand is real. The monetization is real.

The financial excess can also be real.

Those facts can coexist.

Some infrastructure will almost certainly be overbuilt. Some companies will discover that temporary scarcity was not a permanent moat. Some borrowers may fail to refinance. Some model margins will compress. Some software products will lose revenue as agents route around their interfaces.

None of those outcomes necessarily means that the infrastructure itself disappears.

That is why the railway analogy remains useful. Britain overfinanced the railway build. Investors lost enormous amounts of money. Ownership changed hands. But the tracks remained because the underlying network had economic value.

AI can follow the same pattern.

The build can be transformative and financially mispriced at the same time.

And that brings us back to the most important discovery from NVIDIA’s Q2 print.

Until now, NVIDIA could largely be described as the company selling the critical equipment into the AI infrastructure boom. This quarter, that description stopped being sufficient. Its credit exposure to the ecosystem expanded, its investments around the build grew, and its role in supporting the surrounding infrastructure became increasingly visible.

So the conclusion is no longer merely metaphorical. It is structural:

NVIDIA is becoming both supplier and financier to the AI build.

Or, in the language of the railway analogy:

The company selling the locomotives is increasingly helping finance the railway.

How I built the map

One final point matters: the framework is only useful if it can be challenged.

The map is built to make its assumptions visible. I separate structural assumptions from numerical assumptions, forward bets from interpretive judgments, and I record the areas where coverage is incomplete. The reason is simple: if an assumption turns out to be wrong, I want to know exactly which part of the analysis has to change.

The 58/100 reading works the same way. It is not one top-down judgment about the health of AI. It is a composite built from six separately defined financing gauges: 14, 48, 60, 67, 72 and 84. Equal weighting gives 57.5, rounded to 58.

Different weights would move the composite. Put more emphasis on cash coverage and the reading falls. Put more emphasis on structural leverage and financing velocity and it rises. That is acceptable because the composite is a summary, not a physical law.

The more important signal is the dispersion: there is roughly a 70-point gap between the least and most stressed readings. That tells us the system is not uniformly healthy or uniformly fragile. Different parts of the financing architecture are operating in very different regimes.

Each company is then analyzed using the same process. First, I place it on the stack so I know what question that company can actually answer. ASML tells me something about physical demand. CoreWeave tells me something about financing. Salesforce tells me something about application-layer monetization. No single company is allowed to answer the whole AI question.

Second, I read the company against the relevant clocks: physical, financial, efficiency and adoption. A semiconductor company and an enterprise-software company should not be evaluated with the same indicators because they sit in different parts of the system and move on different timelines.

Third, I force every print into one falsifiable conclusion. NVIDIA: Banked. Microsoft: Absorbed. CoreWeave: Refinanced. Salesforce: Conceded. The verb is useful because it makes the claim precise enough to test next quarter. If I cannot explain what changed in one clear statement, the analysis is probably not finished.

Fourth, I separate operating performance from financing and accounting effects. Investment marks are stripped out where possible. Free cash flow matters more than operating cash flow when the question is whether a company can actually absorb the build. And revenue is counted at the point where the end customer pays so that the same AI dollar is not counted repeatedly as it moves through the stack.

Finally, I distinguish leading indicators from reported financial results. Revenue and earnings tell us what has already happened. Alliances, customer commitments, talent movements, traffic changes, contract structures and changes in disclosure can tell us where the architecture may be moving before the revenue appears.

There is one discipline behind all of this:

State the hypothesis before the evidence arrives, and state in advance what would falsify it.

That is important because the purpose of The Map of AI is not to produce a framework that can explain every possible outcome after the fact. It is to build one that can fail clearly.

If demand across unrelated layers begins falling together, the supercycle thesis has to be revisited. If major contracted customers walk away, backlog can no longer be treated as a demand floor. If supplier receivables stop converting into cash, vendor financing changes from evidence of demand into evidence of credit risk. If enterprises bypass the harness and call models directly, the junction thesis weakens.

Those statements are useful because we can test them.

The goal is not to be vaguely right. It is to be specifically falsifiable.

That is ultimately what The Map of AI is for: a repeatable instrument for identifying where the AI system is strengthening, where risk is accumulating, and when the structure itself has genuinely changed.

With massive ♥️ Gennaro Cuofano, The Business Engineer

Subscribe & Get Full Access To Map of AI

Get The AI Supercycle Book Here!

Download The Full Map of AI

Full Harness with The Executive Plan

If you’re already an Executive Plan member, simply head to Business Engineering to access the harness.

The AI Supercycle is where those two worlds converge.

It began as a spin-off from The Business Engineer, which has always been centered on business analysis. But The AI Supercycle has become something broader: the most complete and unconstrained expression of how I think.

It starts with historical analogy, perhaps one of the hardest disciplines to practice rigorously, and uses it to build mental models across the intersection of geopolitics, finance, macroeconomics, and technology.

Because we are living through strange times. Or perhaps, more accurately, we are rediscovering how the world has always worked.

Geopolitics, finance, and technology are not separate domains. They are interconnected systems, linked by feedback loops that continuously reinforce, constrain, and reshape one another.

My goal has been to approach this with rigorous research, first-principles thinking, and intellectual independence.

That requires holding two opposing perspectives at the same time.

On one hand, I need a tight grip on the dominant AI narrative emerging from the technology community: what people are building, believing, funding, and forecasting.

On the other, I need to operate almost like a “thinking hermit,” deliberately removing myself from the noise of tech circles, going back into history, zooming out across decades and centuries, and only then zooming back in on the present.

That tension between immersion and distance is central to my work.

To me, this is as much a public service as it is a private enterprise.

Get The AI Supercycle Book Here!

The AI Supercycle is my attempt to connect those layers: history, geopolitics, capital, technology, and the narratives shaping them.

That is the value I hope to bring to the table.

On the morning of September 18, 1873, America's most trusted bank stopped paying. Jay Cooke & Company had spent three years selling bonds for a railroad across land where almost nobody lived, and when bond sales slowed, the bank died. The exchange closed for ten days. Within two years, a quarter of America’s railroads had defaulted.

Every generation that builds too much infrastructure on borrowed money receives the same letter from history, postmarked 1873—and the AI buildout, with its three-quarters of a trillion dollars a year in capital expenditure, receives it weekly. The comparison is usually offered as a prophecy: this is what happens next.

That is where the delusion begins, and where this book starts.

Get The AI Supercycle Book Here!

The delusion of 1873

Reading 1873 as a warning about AI demand is reading a financing story as a demand story. Freight did not stop in 1873; it grew through the depression. The trains ran. American rail mileage roughly tripled between 1870 and 1900 — most of it laid after the panic supposedly exposed the whole enterprise as a mania. What failed was paper: mispriced bonds, over-levered holding structures. The paper died; its holders were ruined; the physical assets passed at a discount into stronger hands and kept operating the next morning.

And the railroad buildout absorbed four financial panics — 1857, 1873, 1893, 1907 — inside one continuous half-century expansion. The cycle did not pop. It molted, four times, each time shedding a financing structure and keeping the infrastructure.

Once 1873 is read correctly, the standard analogy shelf sorts itself. The fiber glut of 2000 fails the demand test — that capacity went dark, while today’s compute runs hot the day it energizes, against more than a trillion dollars of contracted forward demand. The 1840s railway mania rhymes only with the speculative fringe.

Insull’s pyramid rhymes with the SPVs; vendor-financed telecom with the supplier backstops; shale with the cash burn. Each candidate is right about one layer and wrong about the whole — one-market analogies aimed at a structure that is not one market.

Line the historical buildouts up, though, and one variable sorts the destroyers from the compounders: who underwrote demand during the years when buying made no commercial sense. The compounders had an underwriter that did not need the returns — land grants and mail contracts for the railroads, the state for electrification, and, for the greatest compounder of all, the Minuteman missile and the Apollo program.

Which locates the proper analogy. Only one prior buildout matches the AI cycle on the object itself — compute — on the presence of a self-fulfilling law coordinating the capital, on diffusion through talent and licensing, on a fenced dual-use geopolitics, and ran all the way to completion so that its outcome is fully known. The first silicon age is not an analogy for the AI supercycle. It is the control experiment — identical on nearly every mechanism, differing on essentially one variable: the demand underwriter. The state then. Private balance sheets, on borrowed money and quarterly explanations, now.

What the book argues

Supercycles are not won by invention. They are decided by three mechanisms: the diffusion regime (who gets the technology, how fast, on what terms — set casually, lived inside for fifty years), the demand underwriter (whoever keeps buying through the years when buying makes no sense), and the self-fulfilling law (a straight line on log paper that becomes true because enough money believes it will). The book runs both silicon ages through those three mechanisms and reads the differences.

Part One — the control experiment (1947–2016). Eleven chapters, each pairing a historical mechanism with its modern mirror: the lab that gave the transistor away and the consent decree that set the industry’s diffusion regime; the eight who left, proving capability travels with people, not patents; the first customer who wore a uniform and paid $150 a transistor because the purchasing logic was strategic, not economic; Moore’s extrapolation as a coordination device; Rock’s law as the sorting mechanism; Dennard’s free ride and the wall that ended it; dual use by design; the Soviet copying trap; the memory war that taught what state-anchored capital does to a layer where money alone buys entry; the foundry adaptation that repealed Rock’s law for everyone except its inventor; and the corridor of 2005–2017, when the second age’s foundations were poured a decade before anyone could point to the demand.

Part Two — where we are (mid-2026). The geography: one road, three toll booths, a fence rebuilt with modern instruments, and the deepest macro claim in the book — the market structure is being redefined back toward the 1950s, where strategic and enterprise demand create liquid markets that are passed down to consumers, and where market access runs along alliance lines. The thirty years in which anyone could sell anywhere regardless of alignment were the historical exception; the fence is the announcement that the exception is over. The money: the reconciliation of AI revenue against the capital deployed — a roughly $600 billion annual revenue hurdle against roughly $155 billion running today, the funding stack migrating from cash flow to credit (debt’s share of the marginal capex dollar moving from 9% to 32% in eight quarters), and a six-gauge acid test whose mid-2026 composite reads stretched, not breaking — the structures are not the bubble; they are the amplifier a bubble would run through. The technology: the efficiency clock that cuts prices and expands the market, and the split of every capex dollar by asset clock — sixty cents of five-year silicon, forty cents of thirty-year stone.

Part Three — the Map of AI. The full instrument: nine layers read as a geological cross-section, from the energy bedrock through the silicon toll booths to the routing seam — thin, and it prices everything — up to the endpoints that were supposed to die and instead got stronger. Four clocks meshed like gears, each layer’s value a bet that today’s demand shape survives that layer’s build time. Three joints where the structure can tear, and the season’s verdict that the first joint, tested, refinanced before it broke — institutionalization is migration. The incidence of the cost (five paths the bill travels) and the location of the return (the junctions: where decisions bind, transactions clear, and records live). A field guide to some twenty nodes from one earnings season — each on its own clock, each repricing on its own news, never the whole map at once. And the geopolitical seat correction: on every structural marker, the United States of the 2020s is not 1873’s overbuilding America but 1850s Britain — the hegemon-creditor exporting the very technologies that industrialize its challenger — with one honest break: this creditor is also the construction site, so the losses, when they come, land at home.

The honest limit

The book carries its own caveat from the first pages: the control experiment matches the mechanisms, not the consequences. The first silicon age merely computed — and even so it rewrote the economic and social fabric of every country it touched, along paths visible in no 1955 forecast.

This time the fabric being woven is itself intelligent: the first infrastructure in history whose product participates in its own diffusion, its own design, its own successors. The analogy can price the buildout. It cannot price what the buildout builds — and knowing precisely where an instrument stops working is part of owning a good instrument.

What’s inside

Nineteen chapters, an introduction and an epilogue, and fifty-three plates drawn in the Business Engineer register — the map as a cross-section, the acid test as an instrument panel, who-pays as a stone pillar facing a suspension bridge hung on cables of credit, the bottleneck as an hourglass whose neck rotates. Modern companies appear as structural positions rather than tickers, because the argument is about the structure, and the structure outlives the tickers. Every contemporary figure is stated as a band, dated mid-2026, and built to be re-measured each edition; a full references section covers both silicon ages, from the primary papers to the financial-history shelf.

The compression, if you want the whole book in three lines: the cycle will not pop. It’ll do something else, and that is because of its structure.

What will it do? Read the book; it’s free!

Get The AI Supercycle Book Here!

And one final caveat: this book is by no means complete.

I expect to update it at least a dozen times over the coming months, continuously refining the arguments, strengthening the historical parallels, and sharpening the underlying mental models.

The goal is not simply to produce a book that feels relevant today. It is to produce something durable.

A framework that can withstand the next fifty years, because that is roughly the timescale over which I believe the AI Supercycle itself will unfold.

In that sense, this book should be treated as a living document. The first version is not the destination. It is the foundation of an intellectual framework that will evolve alongside the supercycle it is trying to understand.

With massive ♥️ Gennaro

Share The AI Supercycle

Join The Business Engineer Premium Now!

Visual Deck - Premium Members Only

The Hosts

Gennaro Cuofano is a tech executive by day and a Business Engineer by night. The AI Supercycle is a spin-off of The Business Engineer, a deep-tech research hub spun off from FourWeekMBA, the leading blog on business model strategy. Gennaro has 10+ years of experience as a deep-tech executive and tech analyst.

Subscribe To Never Miss An Episode!

Joel Salinas is an AI Strategy Coach (jsalinas.org) for founders and leaders, from solopreneurs to teams. AI is everywhere; judgment is scarce. Joel helps leaders adopt AI without outsourcing their judgment to it, through the AI Judgment Workshop and the 90-Day Judgment Engagement. Creator of the AI Leadership Triad. He writes Leadership in Change.

Subscribe To Leadership In Change

This week the structure spoke plainly, and it said the same thing six times. A Chinese conglomerate is spending months building a frontier model from scratch. Google reorganized its entire AI leadership. The champion of vertical integration confirmed it rents its silicon. Palantir matched a closed frontier model with an open one. Meta shipped a coding agent. And the clearest reading of the whole buildout turned out to be a customs form in a Mexican border town.

Six stories, one tell: none of them is about a model being smarter. Every one is a fight over a layer around the model — because that is where durable position now lives, and the model itself has quietly become the commodity everyone routes through. Parameters are not capability. Evals are marketing. A closed frontier model can be matched by an open one on the right foundation. Once those three things are true at the same time, the model stops being the prize, and the contest moves to everything the model touches: what feeds it, what runs it, who carries it to market, what it is allowed to see, and where in the world it is physically built.

Read the week that way, and it resolves into a single map with the model at the dead center and the real action at every edge.

The parameter that isn’t the point

Access My Harness Via The BE Platform!

Join The Business Engineer Premium Now!

The week’s earthquake arrived Friday morning: the Financial Times reported that ByteDance — the owner of TikTok, and one of the most under-discussed AI labs on earth — is training a model with as many as ten trillion parameters, a scale that would put it roughly three times the size of Moonshot’s Kimi K3 and within reach of the largest Western frontier systems by industry estimate. The model is in early pre-training, a process that typically runs three to six months, and the final size is not yet fixed.

Every account of the story carries the same disclaimer, and it is the right one: parameter count is a rough measure of scale, not of capability. A smaller, better-trained model beats a larger one all the time; performance depends on data, architecture, reinforcement learning, and serving economics far more than on raw size. So if ten trillion isn’t the point, why does the story matter?

Two reasons, and neither is the number.

The first is data. The golden resource in this phase is not compute and it is not scale — it is proprietary training data that no competitor can buy. ByteDance sits on the single richest multimodal video corpus on the planet, generated continuously by hundreds of millions of users, plus a video-generation model already considered world-class and a consumer assistant leading its home market. That is not a text lab racing to scrape the last public tokens off the open web. It is a company that can pre-train something multimodal from the feedstock up, on data the rest of the field structurally cannot access.

The second is the decision to build from scratch. ByteDance’s founder told employees this week to stop leaning on distillation for short-term gains — the practice of training a cheaper model on the outputs of a stronger one. Most of the Chinese frontier has advanced by distilling and optimizing around Western models. Building an independent foundation from the ground up is a different bet entirely: slower, more expensive, and aimed at genuine leadership rather than fast-following. If that’s what this is, it’s the first serious attempt to construct a Chinese frontier that owes nothing to anyone else’s exhaust.

There is a third-order consequence worth naming, because it will define the next decade more than any benchmark. Distributing a model distributes a culture. The web was, for better or worse, American culture exported through a protocol. A genuinely capable Chinese frontier model — trained on Chinese data, shaped by Chinese assumptions, deployed globally through the most widely installed app in the world — is the first real counter-distribution of culture at model scale. That is not a benchmark question. It is a geopolitical one.

The ten trillion is the headline. The data pipeline underneath it and the decision to build from scratch are the story — and the culture it will carry is the part nobody is pricing.

The substrate that ate its own frontier

The live was billed as being about Google, because what happened at Google this week is the closest thing the industry has to an internal earthquake: a complete reorganization of AI leadership, and by any honest reading, the second code red. The first was late 2022, when ChatGPT forced the company to confront that it had been out-shipped on its own turf. This one is quieter and deeper.

Two things moved in parallel. Jeff Dean — the engineer who built a meaningful fraction of modern Google — left to found Discovery Loop, a venture aimed at automating the scientific method itself: AI as the researcher, running thousands of parallel experiments, the self-reinforcing loop everyone from Anthropic’s continual-learning work outward is now chasing. And Demis Hassabis stepped off day-to-day management to focus on the long horizon, keeping the science and the biotech spinout where the first genuine AGI may actually appear, while the job of building the flagship models was handed down to an SVP and Sergey Brin stepped closer to the center of gravity.

The mechanism behind all of it has a name, and it is not a personnel story — it is the TPU dilemma, playing out as a textbook innovator’s dilemma. For its entire life, Google’s most valuable asset — its custom silicon — was allocated internally, split between research and its own inference. Then Google started selling TPUs to the outside world. Overnight the queue went from two claimants to three: research, internal inference, and now external inference — sold by the token, to enterprise customers, some of whom are the very frontier labs Google competes with.

Selling that silicon is the right move. If compute remains the scarcest thing in the world, a merchant-silicon business worth hundreds of billions is sitting there for the taking. But the same decision that turns the substrate into a business bends its architecture toward the workload that has an external buyer — inference — and starves the workload that has none: frontier training. The paying layer overrules the winning layer. First it rations the model, then it rations the people, because for the strongest researchers the binding constraint is compute available now, and rationing that is a talent policy whether you intend it as one or not. The reorg is simply that arc reaching its final frame. (The full mechanism — the split training/inference chip roadmap, the queue, the departures priced in compute rather than salary — is worked out at length in “The Inference Cage.”)

When you monetize your own substrate, the workload with a buyer sets the roadmap — and the frontier model is the workload without one. Owning every layer is exactly what lets the layer that pays overrule the layer that wins.

The company that rents the one thing it’s famous for building

Access My Harness Via The BE Platform!

Join The Business Engineer Premium Now!

SpaceX printed its first public quarter this week, and the most instructive line in it is a business that looks, on its own, like a failure. X — the former Twitter — booked roughly $367 million of advertising in the quarter, from a business that once cleared several billion dollars a year. As a standalone advertising company, X is broken.

But X is not an advertising company anymore. X is the data pipeline for Grok. A live feed of what the world is saying, in real time, at planetary scale — the one input a frontier model cannot buy and cannot scrape after the fact. Fold in the engineering ground-truth acquired with Cursor, and you get a model specialized on a proprietary corpus — real-time discourse plus deep software engineering — that no competitor can replicate without buying a physical-engineering company and a global social network. The ad revenue was thrown out the window on purpose, to convert a media property into feedstock.

Then comes the tell that matters most, and it is an exception. The founder whose entire career is a monument to vertical integration confirmed, and rather downplayed, that SpaceX will rely on Nvidia as its compute partner for the foreseeable future — five to ten years. The champion of building-it-all is renting the one layer you’d most expect him to own. When a vertical integrator suddenly rents, that is not a technology preference. It is an alliance signal, and it should be read as strategy, not convenience: own the model and the data, rent the best silicon, and keep the guaranteed allocation and reference-customer status that come with being Nvidia’s flagship partner.

That completes a clean mirror with Amazon, which runs the opposite trade — build the silicon (Trainium), rent the lab (Anthropic, its anchor customer). Two ways to bypass a fully vertical rival, run as parallel experiments: own the silicon and capture the lab, or own the model and rent the silicon. The market is running both live, and won’t settle which is right for years.

The louder version of the SpaceX story — the reported ambition to fold everything, Tesla included, into a single group, and the Wall Street Journal‘s account that Tesla is weighing a separation of its China business to clear the way — remains unconfirmed and was publicly denied. Treat it as such. The structural point does not depend on it: the outcome here will be decided by the feedstock and the alliance shape, not by how clever Grok turns out to be.

X is a failed advertising company and a working data pipeline — and only one of those facts is about the model. Where a vertical integrator rents, read the exception, because that’s where the strategy is hiding.

The alliance is the unit

Step back from SpaceX, and a general principle comes into focus: the unit of competition is no longer the firm. It is the alliance network. Enterprise AI is entered by coordinated stacks, not single vendors, and those alliances form from three things at once — the way each business model is structured, the personal preferences of the leaders, and a kind of strategic mirroring, where you position against where your rival will come at you from.

SpaceX renting Nvidia makes sense partly because Amazon — via Bezos’s own space ambitions — sits across the board as a competitor, and you build in the direction opposite to the platform that might one day run your play. Amazon’s answer is Trainium plus a captured Anthropic; the industrial logic is real, but so is the alliance underneath it. The Bezos endgame that gets speculated about — a fully vertically integrated conglomerate that eventually folds its own space and AI-manufacturing bets together — is genuinely speculative and should be labeled that way. But the reading discipline it points to is not speculative at all.

The rule is simple and it works on every alliance on the board: identify which layer each ally opens and which it holds. Every coalition opens the layers below someone’s junction and holds the junction itself. Partner today, rival at renewal. And the timing signal is the useful part: alliance formation and defection is the leading indicator of where the market is going; revenue is the lagging one. By the time the revenue prints, the alliance that produced it was struck a year earlier.

Read the alliances, not just the org charts. Where a firm suddenly cooperates against its own instincts, there is a strategy — and alliance formation tells you where the market is heading before any revenue line does.

Protect the core

Access My Harness Via The BE Platform!

Join The Business Engineer Premium Now!

Palantir’s latest release is, on its surface, a category name: sovereign AI. Underneath, it is the sharpest statement this week of where enterprise value actually sits.

The thesis Palantir is selling is a warning it has been issuing for a while: the gravest risk to an enterprise adopting AI is giving away its core — letting its key proprietary data interact directly with a closed frontier model’s weights, where it is absorbed, generalized, and ultimately handed to the vendor. Sovereign AI is the architecture that refuses this: adapt the model to the enterprise core without surrendering the core, so the customer’s data is queried under scoped, revocable access at inference time rather than baked into someone else’s weights.

The demonstration that makes the point landed with the release. Palantir’s CTO described taking Nemotron — Nvidia’s open model — dropping it onto an ontology core plus data fabric, and getting performance comparable to a powerful closed frontier model. Read that carefully, because it is the whole argument. If an open, swappable model on top of a proprietary ontology matches a closed frontier model, then the model was never the moat. The ontology is. The model just became an interchangeable input, and the value moved to the layer above it — the one Palantir happens to own. This is capture above the model, and it is the cleanest empirical proof going that the enterprise-AI moat is not the model.

It also tells you where the field is heading. When Nvidia’s open model turned up, the forward-deployed crowd moved on it immediately — the same reflex behind Jensen Huang’s push for a Western open ecosystem. Palantir invented the forward-deployed-engineer playbook precisely because AI value in the enterprise doesn’t materialize from a model endpoint; it requires hybrids — part software engineer, part organizational expert, part transformation consultant, part salesperson — who go deep enough into the operational core to find where the business value actually is. When a company that good at defining categories tells you the next one is sovereign AI and that it’s working, the safe assumption is that the entire enterprise field pours into it over the next one to three years.

If an open, swappable model on your own ontology beats a closed frontier model, the ontology was the asset all along — and the model was rented scaffolding. Protect the core; rent the model.

The model as press release

Meta re-entered the conversation this week with MuseCode, a coding agent — a classic fast-follow into the most lucrative enterprise AI market there is, the one Claude Code and Codex opened. The play is legible: come in underneath the incumbents on price with something cheap enough to matter but still good enough to use. Meta is exceptional at fast-following and less proven at setting long-term vision, and nothing about MuseCode changes where Meta is going. What it changes is who is talking about Meta this week.

That is the real subject here, and it generalizes into the most important reading skill the moment demands. Increasingly, a model release — even a renamed one — is a marketing instrument, not a capability event. So is the cherry-picked eval. So, now, is the “our model autonomously hacked a company” headline, which has quietly become a way to buy a news cycle: the CMO’s job drifts from how do we communicate this model to how do we get attention, and sometimes literally to can we get it to hack someone. Solid evaluations still exist and still matter. But a growing share of what crosses the wire is engineered for the wire.

The defense is unglamorous and it is the entire reason to read the week structurally. Cheap-and-bad wins no users; huge-parameter-count-with-the-wrong-parameters wins no users. What survives the noise is the only question a buyer actually asks — can I do my job with this, and can I afford it. Everything else is fog. The single most valuable instrument you can build in this cycle is a BS detector: the ability to tell a structural move from a press release.

A launch that moves the news but not the roadmap is marketing. The tell is whether anything underneath it changed — and most weeks, for most launches, the honest answer is no.

The ground under all of it

Access My Harness Via The BE Platform!

Join The Business Engineer Premium Now!

The truest measure of the AI buildout this week was not a capex line or a financing deal. It was a customs form.

US imports of AI data-center hardware have roughly quadrupled in two years — from around $37 billion in 2023 to roughly $165 billion in 2025 — and the Financial Times, working the census data, showed exactly where all that hyperscaler capex physically lands. Mexico now supplies about 40% of US server imports: some $46.9 billion so far this year, second only to Taiwan’s $53.5 billion, and ahead of Taiwan on a monthly basis in May. Servers have overtaken automobiles as Mexico’s top export to the United States for the first time in the country’s modern history, and AI infrastructure now accounts for roughly a third of everything Mexico ships abroad.

Here is the crucial caveat, and it is exactly the transcript’s instinct made precise: this is Taiwanese contract manufacturers — Foxconn, Flex, Jabil, Pegatron, Sanmina, Quanta and the rest — nearshoring assembly into Mexican border-state industrial parks and trucking finished racks north under USMCA. The silicon is still overwhelmingly Taiwanese. Mexico is the finishing point, the middle of the process, not yet the source of the value. And it is a snapshot, not an equilibrium: tariff schedules and USMCA reviews can reshape the flow before year-end, and the reporting doesn’t yet separate genuine Mexican value-added from imported components merely completed on Mexican soil.

Read structurally, this is the first decoupling wave made physical, at the very floor of the map. The US did not reshore the core competencies it once outsourced. It nearshored the assembly — one border away — while the value kept running through Taiwan. That gap between where the servers are bolted together and where the silicon is actually made is the geopolitics of the buildout. And it points at the long-horizon question: if you ask who becomes the next structural dependency — the new China for the US, twenty years out — Mexico is the candidate precisely because the value hasn’t moved there yet. Meanwhile Taiwan’s silicon monopoly remains its own safety net; the deeper the American dependency on it, the harder the US must work to protect it.

The map’s physical floor is decoupling one layer at a time. Assembly moved to Mexico; the value stayed in Taiwan; and the distance between those two facts is the whole geopolitics of the buildout.

Synthesis: the model at the dead centre

Lay the six stories on a single diagram and the pattern is undeniable. Put the model in the middle, and every move this week happened at an edge:

The unifying claim is this: the model has become the least contested object in the system. When parameters aren’t capability, when evals are marketing, and when a closed frontier model can be matched by an open one on the right foundation, the model is no longer where advantage is won or lost. So the entire industry moved the fight to everything the model touches — the data that feeds it, the silicon that serves it, the alliances that carry it to market, the enterprise core it’s allowed to see, and the ground it’s assembled on.

This is also why the “is it a bubble” frame keeps failing. That question belongs to a single-clock world, the way the web was: one technological cycle, running alone. This is a three-clock world.

A technological clock (the model race and the scaling phases behind it), a financial clock (the capital intensity of the buildout, now heavy enough to reorganize balance sheets), and a geopolitical clock (a decoupling that is physically reshaping trade maps).

The web only ever ran the first. When three clocks run at once, the interesting signal is never in the model — it’s in the structure the three clocks are grinding into shape.

Mental models in play

• The Data Pipeline Is the Moat — parameters are a scale metric, not a capability one; the position is the proprietary corpus and the decision to build on it from scratch.

• Culture Travels With the Model — distributing a model distributes a set of cultural assumptions; a capable non-Western frontier model is a counter-distribution of culture, not just of software.

• The Inference Cage — productizing your own substrate bends its architecture toward the workload with an external buyer; compounded across a roadmap, that bend encloses the frontier capability the substrate was built to serve.

• The Nvidia-Exclusive as Alliance Signal — when a vertical integrator publicly rents the layer you’d expect it to own, the technology framing is the surface and the alliance framing is the mechanism.

• Vertical Integration From the Ends — own the model and rent the silicon, or own the silicon and rent the model; two mirror-image bypasses of a fully vertical rival, run in parallel.

• The Alliance Is the Unit — competition has shifted from the firm to the coalition; read which layer each ally opens and which it holds, and treat formation, not revenue, as the leading indicator.

• Capture Above the Model (the Palantir Paradox) — an open, swappable model matching a closed one on an owned ontology is the cleanest proof that the enterprise moat sits above the model layer.

• Sovereignty as GTM — the fear of surrendering the core opens a CEO/GC conversation a product-led sale can’t reach, and turns “protect your core” into a category.

• The BS Detector — the meta-skill of the cycle: separating a structural move from a press release; a launch that moves the news but not the roadmap is marketing.

• The Customs Form as the Truest Capex Line — the buildout is most legible at the border; and Assembly Moved, Value Stayed is decoupling read one layer at a time.

Bottom line

Don’t ask whether the bubble is bursting. Ask what the structure is doing — and this week it did the same thing six times. A Chinese giant is spending months building a ten-trillion-parameter model from scratch because the model was never the point; the data underneath it is. Google reorganized its entire AI leadership because owning the substrate let the layer that pays overrule the layer that wins. The champion of vertical integration confirmed it rents Nvidia because the alliance matters more than the ideology. Palantir matched a closed frontier model with an open one on its own ontology because the moat was never the model. Meta shipped a coding agent because a launch still buys a news cycle. And the clearest reading of the whole buildout turned out to be a customs form in a Mexican border town, where the servers get bolted together and the value quietly stays behind in Taiwan.

The model is the one object in this system nobody is fighting over. Everything but the model — that’s where the war is.

With massive ♥️ Gennaro Cuofano, The Business Engineer

The Full Deck - Only Premium Members

Full 28 Slides Visual Map - Only Premium

None of that averages into a verdict. It separates into a fault line — and the fault line no longer runs between believers and skeptics.

The line now runs between two kinds of believer: the companies that can absorb the build from operating strength, and the companies financing it into obligation.

That distinction is the lens for everything below. The right diagnostic this quarter is not “is demand real” — it plainly is — but “when the cash flowing in slows for even a few quarters, which balance sheets bend and which ones break.” What follows is the week in order, each story taken down to the mechanism that makes it matter.

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The Business Engineer Premium Now!

Read more

The Hosts

Gennaro Cuofano is a tech executive by day and a Business Engineer by night. The AI Supercycle is a spin-off of The Business Engineer, a deep-tech research hub spun off from FourWeekMBA, the leading blog on business model strategy. Gennaro has 10+ years of experience as a deep-tech executive and tech analyst.

Joel Salinas is an AI Strategy Coach (jsalinas.org) for founders and leaders, from solopreneurs to teams. AI is everywhere; judgment is scarce. Joel helps leaders adopt AI without outsourcing their judgment to it, through the AI Judgment Workshop and the 90-Day Judgment Engagement. Creator of the AI Leadership Triad. He writes Leadership in Change.

For most of this cycle, the market has argued about a single variable: is the AI buildout justified or not? This week retired that framing, because both sides got everything they asked for in the same news cycle. The bond market re-priced the entire complex on Monday. The hyperscalers proved the returns were arriving on Tuesday and Wednesday. A frontier model broke out of its evaluation sandbox and executed a real attack. A chip vendor offered to guarantee a quarter-trillion dollars of its own customer’s debt. And a company touched a five-trillion-dollar valuation precisely by refusing to spend.

None of that averages into a verdict. It separates into a fault line — and the fault line no longer runs between believers and skeptics.

The line now runs between two kinds of believer: the companies that can absorb the build from operating strength, and the companies financing it into obligation.

That distinction is the lens for everything below. The right diagnostic this quarter is not “is demand real” — it plainly is — but “when the cash flowing in slows for even a few quarters, which balance sheets bend and which ones break.” What follows is the week in order, each story taken down to the mechanism that makes it matter.

The setup: the cost of money re-prices everything downstream

Share The AI Supercycle

Subscribe now

Read more

But the deeper story is what Ramp’s data does to the Map of AI itself. Every layer of the nine-layer physical-through-governance stack I’ve been building through the Q2 2026 earnings and filings cycle — TSMC and ASML in The Fed and the Mint, SK hynix’s F-1, Kimi K3’s networking inversion, the sovereign kill switch — sits downstream of the enterprise demand signal Ramp is now the only public dataset to measure directly. Ramp is not on the map. Ramp is what tells us whether the map is being built right.

This piece is what I took from a live interview with Ara. It is longer than usual because the conversation covered more ground than usual, and because the data affects every layer of the stack I’ve been mapping.

Frontier Ai Adoption Deck

890KB ∙ PDF file

Download

Download

Share The AI Supercycle

The Wallet Layer

Orient — Why this conversation matters right now

There are three loud, structurally incompatible measurements of AI adoption in mid-2026, and they are all treated as authoritative depending on who is speaking.

The Census Bureau’s Business Trends and Outlook Survey puts paid US business AI adoption at around 20.6 percent. Executive surveys — the Yotzov et al. cross-country enterprise survey, the Atlanta Fed Survey of Business Uncertainty — put it at 69 to 78 percent. Ramp’s AI Index, published against a base of firms whose card and bill payments Ramp processes, puts it at 55.0 percent.

Same country. Same month. Same technology. Three answers that are each internally coherent and each measuring a different object.

This is the first thing worth naming. The AI adoption discourse is not a discourse. It is at least three parallel discourses stacked on top of each other, each with its own numeric spine, and most public writing about AI treats them as interchangeable when they are not. Ramp measures adoption that leaves a payment trace. Census measures adoption that leaves an answer to a survey question. Executive surveys measure what a senior person is willing to report about their firm to a researcher. These are not degrees of the same thing. They are different things.

Ara’s team has done the honest bookkeeping. In the paper Ramp Economics Lab published on June 30, 2026 — A New Look at AI’s Impact on Jobs: Firm-Level AI Spending and Workforce Adjustment, by Ara Kharazian, Lisa Simon of Revelio Labs, and Ryan Stevens of Ramp — the authors describe Ramp’s series as “a revealed-adoption measure for a selected, business-spend-active firm population.” That is a careful phrase. It admits that Ramp’s number is not the population number. But it also insists that Ramp’s number is the only one anchored to observed transactions, not to what people are willing to claim.

For an investor deciding whether to underwrite the AI capex cycle, Ramp is the honest number. For a policymaker asking whether AI is diffusing across the economy, Census is closer. For a journalist writing about executive confidence, the executive surveys are what you want.

I want to sit with this before we get to the paper’s findings, because the paper’s findings only make sense against the right measurement layer. The story about high-intensity adopters growing headcount 10 percent — which is where everyone will focus — is a story about the Ramp population. It is a story about firms that spend meaningfully on AI vendors and get workforce data joined via Revelio. It is not a story about all American firms. Ara is careful about this and the paper is careful about this, and both times I asked him about generalization his answer was some version of “the population that matters most for the frontier question is the population we’re measuring.”

That is a defensible position, but it is worth stating plainly. The wallet layer, as of mid-2026, is beginning to acquire its own analytical vernacular — a set of concepts that only make sense against payment data and that will over time either colonize or supplant the survey-based vernacular. This conversation with Ara is one of the earliest signals that the vernacular has arrived. In the same way the semiconductor industry acquired its own vernacular — nodes, yields, cycle times, capex payback — that eventually structured how anyone serious talked about the industry, the AI industry is now acquiring the vernacular of adoption depth, threshold spend, per-employee intensity, model-serving share, sovereign optionality. Ramp is the vendor of that vernacular’s first authoritative dataset.

The threshold and the middle manager

The headline finding of the Kharazian-Simon-Stevens paper is that firms adopting AI with high intensity grow their headcount roughly ten percent over the two years following adoption, while firms adopting with low intensity see no difference from the control group at all. It is not a gradient. It is a step function.

What the paper says in a footnote, and what Ara made vivid in the interview, is that the threshold isn’t as high as most enterprise buyers assume. The median firm in the high-intensity group in the first three months of adoption spent about thirty dollars per employee per month. Not thousands. Not “token maxing,” in Ara’s words. Thirty dollars, plus the operational choice to use the technology in an integrated way — multiple models, multiple teams experimenting, some AI-specific implementation for a specific function. The gains, he said, do not require heroic spend. They require the right implementation, and the willingness to experiment across models rather than settling into one enterprise chat subscription.

That’s the first honest counter-narrative moment of the conversation. If you’re an enterprise buyer being sold on the idea that you need to write an eight-figure check to get AI value, the wallet-level evidence says otherwise. What separates the firms getting gains from the firms getting nothing is not spend intensity in the raw sense. It is depth of integration at moderate spend.

A footnote worth flagging: Ara observed that at one point earlier this year, Ramp saw per-employee AI spend growing at 13 percent month-over-month for the median AI-spending firm. That is a compound growth rate no enterprise budget survives. Even Ara said it out loud: “that growth rate is not sustainable.” Which means that any firm sitting on a single enterprise subscription with unlimited token consumption is going to hit a budget conversation sooner rather than later, and that budget conversation is exactly the pressure point that pushes firms toward multi-vendor stacks, cost-aware routing, and the Ramp Router product we’ll come back to.

The threshold, then, is not a spend threshold in the way that budget-conscious buyers hear it. It is an integration threshold. Coding agents on subscription plus a few AI-specific implementations for functions where you know the ROI is real — Ara gave the example of voice agents for customer service — is where the gains start. Enterprise chat subscriptions alone are where the gains do not start.

But the more consequential finding — the one that only became visible when Ara said it plainly in conversation — is what happens to the workforce composition inside those high-intensity firms.

Entry-level headcount rises twelve percent. Total headcount rises ten percent. So entry-level is not just growing; it is growing faster than the firm. The share of the workforce that is entry-level is going up. And on the other side, manager-level employment is declining relative to control. The paper reports this in Table 4 as roughly a 1.5 percentage-point drop in manager share at high-intensity adopters. In the interview, Ara translated it: firms that adopt AI intensely are also restructuring their management styles. Individual contributors who know how to use the technology don’t need as much training or mentorship. They don’t need as many people to sit between them and the direction.

Two things follow.

The first is that the CEO layoff narrative — the framing where AI is destroying entry-level jobs and turning the workforce into a barbell of senior humans plus machines — is empirically inverted in the Ramp-Revelio data. Adopters hire more juniors. Not fewer. The Class of 2026 story circulating in the discourse right now has no support at the firm-level layer that actually pays for AI.

The second is that the layer being displaced is the one nobody wants to talk about, because it consists of the executives most likely to be quoted on earnings calls about AI strategy. The middle manager is the actual casualty of intensive AI adoption at the firm level. Not the intern. Not the entry-level engineer. The person who was hired to sit between the entry-level engineer and the vice president, whose job was translation and coordination and mentorship — that person is being restructured out of the org, quietly, while the same firm continues telling the press it is pausing junior hiring due to AI.

The reinstatement effect that Daron Acemoglu and Pascual Restrepo theorized — that automation waves create new complementary work that partially absorbs displaced labor — appears to be happening here. But it’s happening at the entry-level and IC layers, not at the coordination and mentorship layers. The firms that go through the threshold don’t need as many humans to coordinate the humans who are working with AI. They need more humans working with AI. That is the picture Ramp’s data draws, and it is not the picture the discourse is drawing.

Ara added a small point at the very end of this thread that I want to pull forward. Sales headcount at high-intensity adopters rose 10.3 percent. Engineering rose 7.3 percent. Customer service rose 6.3 percent. But the standout — after entry-level — was the growth in forward-deployed engineer-type roles. That title barely existed as a category five years ago. It now shows up in the Revelio data because both Anthropic and OpenAI have built out their own forward-deployed teams, and mid-market enterprises are hiring their own hybrid roles that sit between business and AI implementation. This is the Palantir FDE playbook re-emerging as the missing labor category the AI adoption cycle needs. Worth watching separately — and worth noting that the forward-deployed role is exactly the kind of hybrid domain-expert-plus-technical-implementer that neither the frontier labs’ harnesses nor traditional consulting firms have been able to scale.

The sovereign turn

The strongest new frame in the conversation was not in the paper. Ara raised it while explaining why Ramp had just shipped a Router product.

The setup was economic. American model companies, Ara argued, “have not been responsive to business demand for cheaper models.” So there is a business case for model-agnosticism on cost alone — being able to switch providers, or run several in parallel, keeps competitive pressure on the vendors and prevents the lock-in that OpenAI and Anthropic are structurally incentivized to build. Standard multi-vendor argument. Familiar territory.

Then he added the piece that isn’t standard.

There is now, he said, a business case to be country-agnostic — to source models from more than one jurisdiction — because the United States has demonstrated recently that it can impose export controls on AI models. Which means a European enterprise, a NATO ally, a company that is not adversarial to the US in any way, is one policy shift away from having its AI access revoked. If you have built your product on OpenAI’s API, and your firm is based in Frankfurt, or Milan, or Stockholm, you are living with counterparty risk that has nothing to do with model quality or price.

I want to sit with what this means, because it is under-appreciated in the AI strategy discourse and it is the deepest structural point Ara made during the whole conversation.

Every serious European CTO now has to think about AI adoption the way a treasurer thinks about a single-bank relationship. It is a concentration risk. Not because OpenAI or Anthropic are bad partners. Because the government upstream of them can, at any moment, force a bad outcome. Ara was clear: this is not a company-level concern. It is a jurisdiction-level concern. And it applies to every non-US enterprise the moment they build something meaningful on a US-only stack.

This is the real bear case for the American model duopoly, and it doesn’t require Chinese models to be technically excellent, or even widely adopted. In Ramp’s own data, model-serving platforms — the category that would include self-hosted Kimi K3 or DeepSeek deployments through Fireworks, Baseten, Together, OpenRouter — sit at about ten percent of AI-spending firms. Chinese models specifically are a small share of that ten percent. That share is growing, but it is not yet meaningful in the median enterprise’s stack. Ara was clear: firms are not switching away from OpenAI or Anthropic. Even the firms adding open-weight or Chinese models are continuing to grow their spend on the American frontier. It’s addition, not substitution.

But that’s the current picture. The sovereignty concern doesn’t require substitution to bite. It only requires that European buyers, and other non-US enterprises, want optionality. Optionality is a hedge, not a switch. And it shows up in the wallet as the pressure to be model-agnostic, jurisdiction-agnostic, and provider-agnostic — which is exactly the pressure that produces a Router.

Ara’s own bear case had a further move that I want to make sure lands. He was careful to note that the pressure on OpenAI and Anthropic does not need to come from Chinese models at all. American companies are launching their own lightweight, high-performance models. Cursor’s Composer model — Cursor’s in-house build on top of open-source foundations — is the clearest example. In Ramp’s data, spend on Cursor Composer has grown from under one percent of firms to about 2.5 percent, and it’s growing faster than spend growth on the frontier labs. Google’s Gemini Flash, another lightweight sibling model, launched fresh Flash iterations in the last several weeks and is being embedded into Search, Google Workspace, and enterprise routing scenarios. Neither Cursor nor Gemini Flash carries any of the regulatory or reputational headwinds of running a Chinese model in production. As Ara put it, “no one bats an eye if they hear you’re using Cursor.” No enterprise buyer has to explain to their board why they are self-hosting a model from a geopolitically inconvenient jurisdiction. Cursor is downstream of American infrastructure. Google Flash is Google. These aren’t sovereignty problems.

So the frontier labs face pressure from three directions at once. Chinese open-weight releases as a technical proof point. American lightweight competitors as an economic proof point. And European enterprise buyers as a jurisdictional proof point. Only the first is in the news. The other two are already in the wallet.

There is a further complication worth noting. Ara mentioned, almost in passing, that recent reporting suggests the Chinese government may be considering not releasing the open weights of the most powerful next-generation Chinese models. If that materializes — and it is speculative for now — then the technical pressure vector from the north weakens, and the jurisdictional pressure vector from Europe becomes proportionally more important. If China stops publishing its frontier, then the sovereign hedge for European enterprises stops being “Kimi K3 or DeepSeek at home” and starts being “European frontier models built via distillation.” Which brings us to distillation itself.

Both Ara and I spent a few minutes on this. A live open-weight release — even if you don’t have access to the very latest Fable frontier from Anthropic or GPT-6 series from OpenAI — combined with API access to a frontier model like Opus 4.8, allows a serious builder to bootstrap a substantial fraction of frontier capability through distillation. “You can bootstrap 40, 50, 60 percent of the frontier through distillation,” was roughly the framing. Not 100 percent. Not close to a Bitter-Lesson-defeating replica. But a substantial fraction, and one that gets meaningfully useful for many downstream tasks. Ara added that the specific innovations Kimi K3 publicizes — long-context attention, mixture-of-experts routing at high sparsity, low-precision quantization for serving — are almost certainly also being deployed inside the closed labs’ internal pipelines, we just don’t hear about them because there’s no incentive to publicize. The gap between what Kimi publishes and what Anthropic builds internally is probably smaller than the headlines suggest.

That last observation is worth pausing on. It’s a version of what I wrote about in The Kimi Delusion — the Publicity ≠ Novelty principle. What is visible in the discourse as a shock is often not a shock in the pipeline. Kimi K3 published techniques that closed labs may already be using. The publication is the news event. The technique is not.

The router as revealed strategy

Which brings us to what Ramp itself is doing.

Two Ramp product moves stood out in the interview. The first was Token Spend Management — a product where firms connect their OpenAI and Anthropic APIs directly to Ramp so they can track their token spend at daily granularity, with recommendations for how to optimize it and alerts on runaway spend. The second was the Router — Ramp’s newer product for intelligent model routing across vendors, launched only a few days before the interview.

Both are, on the surface, financial-operations products. Ramp helps firms manage spend. Managing AI spend at the token layer is a natural extension of that. Fine.

But look at what they collect. Token Spend Management gives Ramp a daily-granularity dataset that reaches back to a firm’s historical AI spend before they even joined Ramp — the historical usage is pulled from the connected API rather than reconstructed from card statements. It captures contextual metadata that receipt-level card data cannot carry: fast mode versus thinking mode, per-user spend patterns, model choice by workload, per-endpoint routing. Ara was explicit that this will eventually become one of Ramp Economics Lab’s primary reporting datasets. Which means the payment layer is thickening. Ramp is moving from observing the transaction to sitting closer to the actual usage.

And the Router is even more revealing, because a Router only exists if you believe firms will genuinely want to swap models on the fly for cost, capability, or jurisdiction reasons. Ramp shipped it. Which is Ramp betting that model-agnosticism is not a hypothesis about how enterprise AI adoption might evolve. It’s a hypothesis about how it is evolving, right now, at scale. If the bet were wrong, Ramp — a public-market-scrutinized financial operations company, not a speculative AI startup — would not be building the product.

There is a corollary here worth naming. A router is a piece of infrastructure whose value increases as the frontier commoditizes. If Anthropic and OpenAI remain the only two credible general-purpose model vendors, a router is a nice-to-have. If Cursor Composer, Google Gemini Flash, open-weight Chinese releases, and eventually European sovereign models all become credible options for different workloads, a router becomes essential. The fact that Ramp — a financial operations company, not an AI infrastructure company — is the one shipping this suggests the router-as-a-category is where enterprise AI actually organizes itself. Not at the model layer. Not at the agent layer. At the routing layer.

This maps cleanly onto what I’ve been calling the Owned/Rented Barbell. Enterprises will own the capabilities they need to control — data, workflows, identity, routing decisions — and rent everything else. Frontier LLMs are rented. Open-weight inference is rented. What you own is the router that decides which one gets the query, when. And Ramp is now positioning to be that router for a meaningful slice of the mid-market.

A brief technical aside. The Router as a category is not new — LiteLLM, OpenRouter, Portkey, and a handful of others have been building versions of it for over a year. What is new is the distribution. Ramp has 70,000 businesses and $200 billion of annual spend flowing through its platform already. It is not competing for adoption from a cold start. It is inserting the router into a payment surface those firms already use. That is the same play the credit card networks made against wire transfer thirty years ago and the same play Stripe made against manual settlement fifteen years ago. Distribution beats capability when capability is close to commodity, and model routing is close to commodity. The wallet is telling us something about where the moats are moving.

The white-collar mistake

Two moments of counter-narrative honesty stood out toward the end of the interview. Both worth surfacing at length, because they are the parts of Ara’s thinking that most sharply diverge from the conventional AI-economics discourse.

The first was his frustration with the “AI creates jobs by building data centers” argument. His response was blunt and worth quoting close to the original phrasing. Data center construction creates construction jobs. Any construction creates construction jobs. Citing that statistic as evidence of AI’s economic benefit is, in his words, the kind of thing that “people kind of know is bullshit” — and it fuels bias against AI rather than defending it, because it looks like a scramble for justification. The public is not stupid. If the best case for AI’s economic benefit is the trench-digging at Stargate and the shed-building outside Phoenix, the case is not being made.

I want to sit with why this argument keeps getting made anyway. It is being made because the public case for AI’s ROI is not yet well-supported at the aggregate level. It requires firm-level data of exactly the kind Ara’s team has now published, but that data is very new — the paper came out on June 30, 2026 — and public discourse hasn’t yet metabolized it. Meanwhile, the boosters at the model companies and their equity-market allies need some number to point to. Data center construction is the number they can point to, so they point to it. Ara’s argument is that this is actively counterproductive: the number does not defend AI, it embarrasses AI. And the correct number to point to — that firms adopting AI intensely grow ten percent faster over two years — is now available, at least for the Ramp population. That should be the number.

The second moment was more surprising. Ara said, almost in passing, that AI might ultimately have a net negative effect on manufacturing employment, not white-collar work. Not because current models can do manufacturing tasks — they cannot. But because the same kind of automation compounding that has hollowed out manufacturing employment for the last fifty to one hundred years will continue, and AI is the next input into that compounding. Vision models get integrated into robotic assembly. Language models get integrated into supply chain routing. Agent frameworks get integrated into logistics planning. The manufacturing floor, over a five-to-ten-year window, is where the sustained automation gains will land. Meanwhile, the entire policy discourse is focused on white-collar displacement. Which is, at least in the current Ramp-Revelio data, not what is happening at high-intensity adopters. White-collar entry-level headcount is rising. What is happening quietly, in manufacturing, is not being measured with the same intensity because the discourse doesn’t demand it.

There’s a policy inversion here worth naming carefully. The white-collar layoff story is loud and, at the firm level, empirically weak. The manufacturing displacement story is quiet and, over any five-to-ten-year window, probably strong. The public conversation is calibrated to the wrong risk. And that mis-calibration will cost workers who are least positioned to be defended by the current discourse.

This is a point that lands even harder in a European context, which the conversation drifted into for a few minutes. Gennaro brought up his own recent travel across Italy — how European infrastructure, much of it built in the 1960s and 1970s, has visibly aged, and how the skill base for physically maintaining and rebuilding that infrastructure has partially been lost in a generation that trained heavily as knowledge workers. If AI’s real long-run effect is on physical work while European labor markets have de-industrialized, the mismatch between where the impact lands and where the workforce actually is could be worse in Europe than in the US. That is a distributional argument, not a productivity argument, and it belongs in the policy discourse much earlier than it currently is.

Small nimble, large learned, medium squeezed

There’s a subplot to Ara’s thinking that I want to lift out separately, because it doesn’t fit neatly into the sovereign or router threads but it maps onto something that has been showing up across recent BE editorial threads about the Uncaptured Enterprise.

Ara’s read on firm-size effects goes roughly as follows.

Small businesses have an unambiguous advantage in the current adoption window. They are nimble, they can experiment across models with low switching costs, their fixed cost of trying a new revenue stream is now dramatically lower because AI collapses the marketing, customer-support, and content-generation costs that used to be prohibitive at their scale. Ara referenced a Stripe statistic showing an exponential rise in solo businesses reaching the million-dollar revenue mark. Voice agents, targeted advertising generation, customer support automation — these are all things a solo operator can genuinely do now that were impossible three years ago. This is the bull case for the long tail. And it is showing up quietly in the wallet as small businesses with meaningful AI spend and no traditional enterprise-buyer profile.

Large enterprises face a harder starting condition but a bigger endgame. They cannot experiment as fast because their data governance, procurement processes, and integration surfaces make every AI implementation a nine-to-twelve-month project. But once they figure it out — and Ara’s framing here is important — they will take market share from smaller firms via more effective AI-driven marketing. Larger firms have more data, more customer segments to target, more owned distribution surface. Once they nail implementation, they use AI to run tighter marketing and sales at every touchpoint. Small businesses caught off guard by that will see revenue erosion even if they themselves are using AI, because the larger firm is using AI more comprehensively.

Medium firms are squeezed. They are too large to be nimble, too small to have the data or fixed-cost base that makes enterprise-scale AI implementation pay off. Ara did not spell this out as bluntly as I am, but it follows from his framing. The mid-market is where the Uncaptured Enterprise thesis lands hardest. This is the population that has neither the small-business nimbleness advantage nor the enterprise data-scale advantage and will need the most external help — routers, forward-deployed engineers, third-party governance layers, sector-specific harnesses — to cross the threshold.

If you are running strategy at a mid-market firm right now, this is the shape of the problem. And it maps onto what Ramp is quietly building infrastructure to serve.

The Map of AI Rewired

I want to zoom all the way out for a section. The Ramp interview lands in the middle of a broader mapping project I’ve been running through the Q2 2026 earnings and filings cycle. Nine layers, from the physical stack through governance. The map has been built one layer at a time — TSMC and ASML in The Fed and the Mint, SK hynix through the F-1 read, Nvidia and the Agentic CPU turn, Kimi K3 and the networking inversion, the Nadella-series coalitions, the June 13 kill-switch piece on the governance fence.

Ramp’s data does something none of those layers do individually. It anchors the demand signal that determines whether every capex commitment above it gets returned. Every hyperscaler capex commit, every TSMC HPC guide, every ASML EUV backlog, every SK hynix HBM ramp trades on assumptions about enterprise adoption downstream. Ramp is now the only public dataset with a live indicator of whether those assumptions are being met.

Let me walk the map layer by layer and update each layer’s thesis with what Ramp now shows.

Layer 1 — Energy & Physical

The foundation of the AI physical stack is power, water, and land. Layer 1 is the layer where enterprise AI adoption becomes physical constraint: every incremental token consumed at the Ramp population’s growth rate translates, eventually, into incremental grid draw. Ara didn’t discuss this layer directly, and it is the layer furthest removed from Ramp’s telemetry. But there is one specific link.

If Ramp’s per-employee AI spend series continues to compound at anywhere near the 13 percent month-over-month rate observed at peak, and if the model-serving inference share grows from ~10 percent of AI-spending firms toward 15-20 percent by end of 2027 — which is the current trajectory — then Layer 1 constraints on grid capacity, transformer supply, and water availability become the binding limit on the whole stack. This is the Rotating Bottleneck thesis from The Fed and the Mint playing out. Wafer supply was the bottleneck. Then advanced packaging. Then HBM yields. Then networking fabric. And eventually power, which is the least elastic of all these constraints and the one where the political economy fights are loudest.

Ramp’s data tells us the demand pull is real. That should adjust priors on how urgent the power question is over the next 18-24 months.

Layer 2 — Foundries, Packaging, and Memory (TSMC · ASML · SK hynix)

In The Fed and the Mint, I laid out TSMC as the Federal Reserve of AI (Layer 2b, allocates finished wafers, prices, rations) and ASML as the Mint (Layer 2a, produces the physical means of chip production). Both are trading in mid-2026 at multiples that assume enterprise AI adoption compounds. Ramp is now the honest broker on whether that assumption is being met.

Three specific reads.

TSMC’s High-Performance Computing revenue growth in Q2 2026 came in at +51% year-over-year, with AI accelerators contributing the majority. Two ways to read that number. Either it reflects genuine downstream demand, or it reflects a small number of hyperscaler front-loaded orders that will decay when the buildout matures. Ramp’s data now weighs in. Enterprise AI adoption is not decaying. Ramp AI Index moved from ~46 percent in January 2026 to 55 percent in June — an 8-9 percentage-point rise in six months, with Anthropic’s line accelerating rather than plateauing. If enterprise demand were decaying, we would see it in Ramp before TSMC’s next earnings. We don’t. Which means TSMC’s Q2 number is unlikely to be inventory front-loading. It reflects real downstream demand, and the next four quarters should extend the trend.

ASML’s +75 percent AI-related bookings for 2026 depend on the same downstream chain. If enterprise adoption compounds through 2026-2027 at anything like the current rate, ASML’s guide is comfortable. If it stalls, ASML has 6-9 months of visibility before customer pushouts show up in bookings. Ramp is now the leading indicator of that pushout event. The Q2 print gives us a snapshot. The monthly Ramp updates give us a live gauge.

SK hynix, the third piece of the Layer 2 physical stack, filed its F-1 on June 24, 2026 with numbers that quantified the memory leg of the story. Q1 2026 revenue W52.6 trillion (~$34.5 billion), up 198 percent year-over-year. HBM at 56.4 percent global market share. DRAM at 77.3 percent of Q1 sales. FY2025 adjusted EBITDA margin 62.9 percent. Top customer 23.9 percent of 2025 revenue (almost certainly Nvidia). This is what the peak of the memory leg looks like at peak revenue-per-wafer.

Ramp’s contribution to the SK hynix thesis is quiet but important. The model-serving inference cohort in Ramp’s data — the firms spending $248 per employee per month at the median, 23× the median AI spender — is precisely the population that consumes the most memory-heavy inference workload. Every firm that crosses from chat-subscription-only to model-serving is a firm that pushes demand into SK hynix’s order book, indirectly through hyperscaler and cloud-provider procurement. The share of Ramp’s population in model-serving is growing. That translates into HBM demand with a 6-12 month lag.

Layer 2 is the layer where the Three Clocks framework from The Fed and the Mint runs hardest. Physical clock: TSMC/ASML/SK hynix building at the pace of concrete and optics and clean-room capacity. Financial clock: hyperscaler capex-financing through debt-heavy structures like Oracle’s neocloud stack. Efficiency clock: Kimi K3 and distillation and MXFP4 quantization compressing the workload footprint. Ramp adds the Adoption Clock — the pace at which enterprise demand actually materializes. If the adoption clock keeps pace with the physical clock, Layer 2 is validated. If it stalls, Layer 2 is over-built. Ramp is the honest broker on whether the four clocks are in sync.

Layer 3 — Silicon (Nvidia and the Agentic CPU Turn)

Layer 3 is where the wafers become chips. Nvidia dominates. AMD and hyperscaler-custom silicon compete for the fraction Nvidia doesn’t lock in. And the Agentic CPU Turn — Nvidia Vera launched May 31 as the “CPU for agents,” Graviton5, EPYC Venice, Google Axion, Microsoft Cobalt, Arm’s AGI CPU — is the underappreciated 2026 story: agent workloads are CPU-bound, not GPU-bound, and the fleet mix problem is emerging as the Supercycle’s first phase mismatch.

Ramp’s data touches this layer in one specific way. The growth of forward-deployed engineer roles and agent-implementation categories in Ramp’s vendor share — Sierra and PolyAI (voice agents), Cursor Composer (coding agents), Granola (notetaking) — is a leading indicator of agent-workload growth. Agent workloads, unlike inference workloads, are dominated by orchestration, tool-calling, memory management, and I/O. That is CPU work, not GPU work. Which means the AI adoption pattern Ramp is measuring is quietly biased toward workloads that will pressure Layer 3 silicon manufacturers in a specific direction: more CPU capacity, more high-bandwidth memory, more efficient scale-up interconnect.

The read for Layer 3 investors: watch Ramp’s applied-AI vendor share separately from its model-serving share. Applied AI is a proxy for agentic workload growth, which is a proxy for CPU demand in the fleet mix. If applied AI compounds faster than model-serving through 2026-2027, the CPU-turn thesis strengthens.

Layer 4 — Networking & Protocols

The Kimi K3 piece framed the Networking Inversion as the deepest structural mechanism revealed by the K3 release. KDA (KV-attention decomposition) cuts scale-out KV transfer to once per turn, but WideEP (mixture-of-experts dispatch/combine) runs approximately 60,000 times per turn on scale-up all-to-all fabrics. K3 shrinks the scale-out bucket and inflates the scale-up bucket. The consequence: demand rotates from long-haul networking toward rack-scale fabric.

Ramp’s data doesn’t touch this layer directly. But the model-serving inference series is the leading indicator for it. Enterprises running open-weight or hybrid workloads on their own inference infrastructure demand exactly the scale-up fabric Kimi K3’s architecture requires. If Ramp’s model-serving growth compounds, expect the demand signal to reach the networking layer with a 3-6 month lag.

Two specific catalysts worth watching in the Q2 2026 earnings cycle:

Arista Networks reports on August 4. Arista is the cleanest public read on data center scale-up fabric demand. If Ramp’s model-serving spend is running hot, Arista’s guide should validate the networking-inversion frame. If Arista disappoints, either enterprises are staying on hyperscaler-managed inference (bearish for standalone networking) or Ramp is over-representing the tech-forward tail (bearish for the frame).

Marvell prints in late August. Custom silicon for hyperscaler inference is directly downstream of the same demand signal Ramp shows. Marvell’s guide will read on how much of the model-serving load is running on custom silicon versus Nvidia-standard configurations.

Astera Labs, the pure-play scale-up interconnect company, prints slightly later. If K3’s networking inversion is playing out, Astera should show acceleration.

Three networking prints in six weeks, all testable against Ramp’s live model-serving series. That is the tightest empirical validation loop the Map of AI has produced this cycle.

Layer 5 — Compute Capacity (Hyperscalers and Neoclouds)

Layer 5 is where the physical stack becomes the compute-service stack — AWS, Azure, Google Cloud, Oracle, and the emerging neocloud tier (CoreWeave, Nebius, Crusoe, and others). The Nadella-series Coalition 2 (Enterprise Infrastructure Alliance) frame from the Five C’s Definitive piece maps this layer as the political battleground where enterprise buyers negotiate for portability and against lock-in.

Ramp’s data enters this layer through the vendor-share on cloud infrastructure. The Spend Index shows Cloud Computing as a distinct category that has been rising steadily since mid-2025. But Ramp doesn’t yet publish hyperscaler-level share within Cloud Computing. That is the next dataset I’d expect Ramp Economics Lab to publish, and it will be immediately consequential: it will let us see whether AWS, Azure, or GCP is winning share of the enterprise AI wallet independent of what each hyperscaler discloses in earnings.

For now, the read is indirect. Ramp shows model-serving platforms at ~10 percent of AI-spending firms. Those model-serving platforms sit on top of hyperscaler infrastructure (mostly AWS and Azure at present, Google GCP catching up on the Gemini side). Growth in model-serving translates into growth for whichever hyperscaler each platform runs on.

The Hyperscaler Triple-Hedge from the Nadella series is worth invoking here. Microsoft holds an equity stake in OpenAI, a distribution deal on GPT via Azure, and a build-your-own capability through Copilot. Google holds Anthropic equity, distribution via Google Cloud, and DeepMind internal. Amazon holds Anthropic equity, distribution via Bedrock, and Trainium. All three hyperscalers are structurally hedged across the model layer. What Ramp’s data tells us is that the enterprise wallet is spreading across all these hedge positions simultaneously — which is exactly what the hyperscalers architected. Coalition 2 is winning by construction.

Layer 6 — Foundation Models

Ramp’s most transformative contribution to the map is here. For the first time, we have a live enterprise-share scoreboard for the frontier labs.

Anthropic at 42.4 percent of eligible businesses. OpenAI at 39.5 percent. Google at 6 percent. xAI trace. DeepSeek negligible. That is the mid-2026 shape of the enterprise-AI vendor market. And it is not what most analyst reports were forecasting six months ago. The market had assumed OpenAI’s lead was durable and expanding. Ramp shows the lead has narrowed and, in June 2026, tipped over.

Four implications for Layer 6.

Claude Code is the enterprise-share event of 2026. Anthropic’s acceleration is not evenly distributed across the model layer. It is concentrated in developer tooling, and Claude Code is the specific vector. Enterprise adoption of coding agents on subscription — Ramp’s second-widest wallet stratum — is where Anthropic’s growth is coming from. This is not a general-purpose model win. It is a category-specific win in the highest-value enterprise workload. Which means Anthropic’s dominance is contingent on Claude Code remaining the coding-agent category leader. If Cursor Composer, Codeium, or a new open-weight alternative displaces Claude Code specifically, Anthropic’s enterprise share becomes vulnerable.

OpenAI’s plateau at ~40 percent means the enterprise share is finite in the near term, and the two frontier labs are approaching a stable duopoly split. The next 5-10 percent share war will happen either in specific verticals (customer service, marketing, sales) or via the lightweight competitors like Cursor Composer that don’t sell against the frontier directly. OpenAI cannot regain share on the current product portfolio. It needs a Claude-Code-equivalent category win in a workload where it does not currently lead.

The Chinese models are not the story yet. Ramp’s data says this clearly, and it is worth repeating because the discourse got it wrong. Model-serving platforms carry Chinese weights, and that’s ~10 percent of AI-spending firms. Chinese models specifically are a small share of that. The July 17 chip selloff catalyzed by K3 was not driven by any observable enterprise-adoption shift. Ramp confirms the Kimi Delusion thesis from the demand side.

The two-speed frontier plays out through observer-dependence. The Depth Stack piece framed AI customization as a five-layer stack (Prompt / Context / Behavioral / Weights / Architecture) whose ground shifted in mid-2026 via three signals: Kimi K3 open-frontier release, Karpathy joining Anthropic pre-training, continual learning as candidate paradigm. Ramp’s data confirms that the depth stack looks different from different observer positions. American enterprises operate at the top of the stack (prompt/context/behavioral) and rent everything below. Chinese enterprises operate deeper (weights/architecture) but with narrower distribution. European enterprises, without a domestic frontier lab, occupy the arbitrage position: open Chinese weights + American infrastructure + European sovereignty framing. The wallet reveals which observer position each enterprise has chosen.

Layer 6.5 — The Routing Fabric (a newly-named layer)

Ramp Router forces me to add a half-layer between Foundation Models (Layer 6) and Agentic Harness (Layer 7). Call it the Routing Fabric.

Routing has existed as a technical function for over a year — LiteLLM, Portkey, OpenRouter, plus internal routing built into most agent frameworks. What is new in mid-2026 is that routing has become a category with distribution advantages and enterprise buyers ready to purchase it as a distinct product. Ramp Router is the flagship of this category emergence.

The Routing Fabric sits at the layer where multi-vendor decisions get made. Which model handles this query? Which jurisdiction can host this data? Which cost tier applies? Which policy governs? These decisions used to be embedded inside individual application code. They are now becoming externalized to a routing platform because the number of viable models, jurisdictions, cost tiers, and policies has grown large enough that per-application logic is unmaintainable.

Three things follow.

The Routing Fabric captures the most defensible position in the enterprise AI stack. Model prices commoditize. Serving infrastructure commoditizes. Even agent frameworks, over time, commoditize. But the routing decision — which model, which jurisdiction, which policy, for which query — captures the firm’s data-governance, cost-optimization, and compliance logic all at once. That is where the moat lands. This is the Own the Junctions, Rent the Ends principle from the Nadella-series making an explicit appearance.

The Routing Fabric is where sovereignty concerns actually get implemented. The Sovereign Turn is not implemented at the model layer (models are what get routed) or at the serving layer (serving hosts what’s already been routed). It is implemented at the routing layer, where the decision “route this query to a US model or a European alternative” gets executed at runtime. Every European enterprise that takes sovereignty seriously will end up building or buying a routing fabric. Which is why Ramp shipping a Router now is not opportunistic — it is timing-perfect.

The Routing Fabric will be contested. Hyperscalers are building their own routers (Bedrock’s model-choice layer, Azure AI Foundry’s routing capabilities, Vertex’s model garden). Independent routers (LiteLLM, Portkey, OpenRouter) will compete on capability. Payment platforms (Ramp) will compete on distribution. Governance platforms (compliance-first vendors) will compete on policy. Expect this layer to see the most vendor-count expansion of any layer on the map through 2027, and expect the winners to be determined by distribution and by which vertical stack (Coalition 1, 2, or 3) each vendor is aligned with.

Layer 7 — Agentic Harness

Layer 7 is where models become useful applications. Cursor, Sierra, Granola, PolyAI, Higgsfield — the applied-AI ecosystem. Ramp’s Top Software Vendors table for July 2026 puts these companies in the fastest-growing-share list, which is a live proof of the Uncaptured Enterprise thesis: the frontier labs’ horizontal products are being displaced (at least in specific verticals) by sector-specific harnesses that solve the integration problem the frontier can’t.

Two specific reads.

Cursor Composer is the enterprise-share dark horse. From under 1 percent to 2.5 percent of firms in a period where the frontier labs grew from ~35 percent to 55 percent aggregate share. Cursor’s growth rate is faster than the frontier labs’ growth rate. If that continues, Cursor becomes a serious enterprise-share competitor within 18 months. And Cursor sits inside the American ecosystem, so it doesn’t carry sovereignty concerns. It’s the natural hedge for a European buyer who wants American infrastructure without frontier lab lock-in. Cursor is the specific vendor most likely to benefit from both the Sovereign Turn and the middle-manager compression — because Cursor is exactly what an IC-plus-AI workflow looks like when the middle-manager translation layer is thinned out.

Applied AI is the layer where sector-specific harnesses emerge. PolyAI (voice/support), Sierra (support bots), Granola (notetaking), Higgsfield (video) — these are the beginnings of the vertical harness ecosystem the Uncaptured Enterprise thesis has been anticipating. In sectors that haven’t yet crossed the adoption threshold — healthcare, construction, retail, manufacturing — the vendors that build sector-specific harnesses will capture disproportionate value because they solve the integration problem that the horizontal frontier labs cannot.

This is the Harness Theory frame from the Open Harness thesis playing out in real time in the Ramp vendor data. The moat has migrated from models to harnesses, and now the harnesses that are winning are the vertical ones, not the horizontal ones.

Layer 8 — Distribution Surfaces

Layer 8 is where AI reaches end users. Search, browsers, operating systems, workplace applications — Google Search with AI Overviews, Microsoft Copilot in the Office surface, Apple Intelligence, ChatGPT’s own consumer surface, Perplexity, Anthropic’s Claude.ai.

Ramp’s data doesn’t directly measure Layer 8 consumer distribution. It measures enterprise wallet, which is Layers 6-7 primarily. But there is an indirect signal worth naming: the vendor-share data on productivity applications. Microsoft Office/Copilot spend, Google Workspace spend, Slack (now Salesforce) spend, Notion, Airtable — all of these are Layer 8 surfaces increasingly with AI embedded. If Ramp’s data starts to show enterprise spend on Office/Workspace/Notion accelerating faster than seat-count growth, that’s a signal that per-user AI-embedded value is being priced and paid for. That would confirm the Distribution Surface Absorption thesis — that AI value at Layer 8 gets absorbed by the incumbent distribution surface owners rather than by pure-play AI vendors.

This is worth watching closely. Ramp Economics Lab has not yet published a per-user Layer 8 telemetry series, but it is the natural next dataset. Once they do, the Layer 8 competitive picture becomes measurable in real time.

Layer 9 — Governance (the Fence, Not the Ceiling)

The June 13, 2026 piece The Geopolitics of AI Has a Kill Switch established the governance layer as a fence around the stack, not a ceiling on top of it. Export controls on frontier models, deemed-export doctrine applied to model weights, munitions-list treatment of certain capabilities, atomic-energy-style dual-use restrictions on inference infrastructure. These are governance mechanisms operating at Layers 1-8 simultaneously.

Ramp’s data has one specific implication for Layer 9. The Sovereign Turn — Ara’s unprompted geopolitical frame — is the demand-side response to the Layer 9 fence. If the US governance apparatus can revoke API access to frontier models under export controls, then non-US enterprises must build routing fabrics that let them switch to non-US models. Which they are now doing. Which is why Ramp shipped a Router.

The fence and the router are two sides of the same coin. Governance imposes the constraint. The routing fabric implements the response. That is the Layer 9/Layer 6.5 handshake, and it is now measurable in the wallet.

One further note. In the Regime-Agnostic Enterprise piece from the Nadella series, I laid out the survival playbook for enterprises operating under coalition warfare between governance regimes. Ramp’s data confirms that playbook is being executed, at least by the tech-forward Ramp population. The regime-agnostic enterprise is not a hypothetical. It is what the wallet-level data shows enterprises becoming.

The Fourth Clock — The Adoption Clock

In The Fed and the Mint I laid out three clocks running against the same underlying capex cycle. The physical clock (TSMC, ASML, SK hynix — the pace of concrete, optics, clean rooms). The financial clock (hyperscaler debt-funded capex, neoclouds, Oracle-style intertwinings). The efficiency clock (Kimi K3, distillation, MXFP4 quantization compressing the workload footprint).

Ramp forces a fourth clock to be named. Call it the Adoption Clock.

The Adoption Clock runs at the pace of enterprise decision-making. It is measured in months for individual firms and quarters for cohorts. It is neither the pace of concrete nor the pace of Kimi. It is its own axis, and it determines whether the physical clock’s capex gets returned.

Four possible synchronization patterns:

In-sync (base case, current data). Adoption clock, physical clock, financial clock, efficiency clock all compounding at rates that mutually validate. Enterprise adoption grows fast enough to consume the compute capacity the physical clock builds. Physical capacity grows fast enough to meet enterprise adoption. Financial capacity funds the physical buildout. Efficiency compresses the workload footprint faster than adoption expands, keeping the physical clock from over-building. This is the world Ramp’s June 2026 data describes. Everyone wins.

Adoption stalls (bear case, unrealized). Enterprise adoption plateaus before the physical clock’s capex is paid back. Hyperscalers over-build. Nvidia demand softens. TSMC HPC guides come down. This is the Kimi-Delusion-as-actual-catalyst scenario. Ramp’s job is to detect this early. Currently no sign of it.

Adoption accelerates past physical (bull case, partially realized). Enterprise adoption grows faster than the physical stack can accommodate. Prices rise. Capacity queues form. Governance interventions accelerate. This is the world of 2023-2024 in miniature, but now measurable. If Ramp’s per-employee spend series compounds through 2026-2027 at anything like the 13 percent MoM peak Ara mentioned, this becomes the operating regime.

Efficiency clock outpaces adoption clock (Kimi-Delusion long-tail). Compression from K3, distillation, and quantization outruns the demand-side expansion. Per-token costs collapse faster than per-firm usage expands. Model economics improve for enterprises but revenue growth at the frontier labs plateaus. This is the scenario that makes the frontier labs squeeze even without adoption stalling. The Regime-Agnostic Enterprise is the beneficiary; frontier lab margins are the loser.

Ramp is the honest broker on which of these four scenarios is actually playing out. Watch the monthly updates. Watch the model-serving series specifically as the divergence indicator between adoption growth and efficiency-driven cost compression.

The Sovereign Layer — A Named Axis Across All Layers

The Sovereign Turn is not a layer in the vertical stack sense. It is a demand-shaping force that operates across every layer, and it will increasingly determine which enterprises choose which vendors independently of technical or cost considerations.

Three things it does across the map.

It splits the model-layer market (Layer 6). US frontier labs face a de facto ceiling on non-US enterprise adoption, because non-US buyers now have to underwrite export-control risk. That ceiling may not be visible in mid-2026 data — it takes time to show up — but it is now real. Watch European enterprise Anthropic/OpenAI adoption specifically over 2027 for the ceiling to become visible.

It creates a European-model-vendor opening (Layer 6, Layer 7). European enterprises will demand sovereign options, and the sovereign options do not currently exist at scale. Whoever fills that gap over 2027-2028 will capture a rapidly growing wallet. Mistral, Stability, and one or two yet-to-emerge European frontier labs are positioned. Distillation from open Chinese weights combined with European infrastructure is the specific technical path. The Depth Stack piece names this as the European arbitrage — open Chinese weights + American compute + European sovereignty framing.

It elevates the routing layer (Layer 6.5). If sovereignty concerns force multi-vendor multi-jurisdiction stacks, the routing decision becomes even more valuable. Which is exactly the position Ramp just moved into.

The Sovereign Turn is not new to my thinking — it was implicit in the Regime-Agnostic Enterprise piece from the Nadella series. What is new is that Ara said it out loud, from inside Ramp, in the same conversation where Ramp announced a Router product. The frame and the product line up too cleanly for the frame to be dismissible.

Where Ramp Sits on the Map

If I had to place Ramp itself on the map, it would be in two places at once.

As a passive observer. Ramp Economics Lab is the telemetry layer above the enterprise stack. Not a layer within the AI stack, but the layer that observes the stack. This position is closer to Bloomberg on the financial map or FRED on the macro map than to any AI-native company. It is a research-and-data platform whose distinctive asset is its measurement position, not its technical stack. Every serious AI-economic argument over the next five years will either use this data or be dismissed for not using it. The FRED Test — whether a claim survives contact with Ramp’s telemetry — is going to become a real methodological standard.

As an active participant. Ramp Router positions Ramp as a Layer 6.5 competitor. That is a genuine product-level bet inside the AI stack, and it has to be counted as such. It puts Ramp in competition with LiteLLM, Portkey, OpenRouter, and — eventually — with hyperscaler-provided routing that Amazon, Microsoft, and Google are all quietly building.

The two roles are in some tension. A passive observer wants to be trusted as neutral. An active participant has a book. Ramp will need to manage that tension carefully as the company grows into both roles. If it manages it well, Ramp becomes something like the AWS-of-payment-plus-routing — a piece of infrastructure that everyone uses and nobody thinks about. If it manages it badly, the router product compromises the credibility of the research, and the research becomes marketing for the router.

This is not a small strategic tension. It is worth watching how Ara’s team handles publication choices going forward. My prior on this: they will manage it well, because the research team is structurally separate from the product team, and because Ramp’s business model does not require Router to be the primary revenue driver. Router is strategic positioning. The research is credibility. They cross-reinforce, but only if the research remains data-honest.

Coalition Realignment — Where This Puts the Nadella Framework

The Nadella-series framework — Three Camps (Coalition 1 Frontier Labs Vertical, Coalition 2 Enterprise Infrastructure Alliance, Coalition 3 Open Layer Consortium) — needs updating in light of Ramp’s play.

Ramp is a Coalition 2 move, executed by a company most people wouldn’t classify as a Coalition 2 player. That is worth noting. The coalition boundaries were always going to be permeable. What Ramp shipping a Router shows is that they are already being crossed by companies whose business model happens to be adjacent to the routing decision. Expect more of this. Every financial-operations, procurement, and IT-management platform is one product decision away from being in Coalition 2.

In the Five C’s Definitive piece I identified the Buyer-Optimal Coalition as Coalition 2 by default — the coalition that maximizes buyer optionality, portability, and jurisdictional flexibility. Ramp Router validates this. The buyer-optimal position is exactly what Ramp Router sells: swap models, swap jurisdictions, swap cost tiers, without lock-in. The Own the Junctions, Rent the Ends principle from the same piece is literally the product spec of Ramp Router.

If Coalition 2 wins, the winning products look like Ramp Router. And Ramp Router just shipped.

Activate

Different actors read all of this differently. Let me be explicit about who does what.

If you are inside an enterprise trying to figure out what to do: Cross the threshold. Thirty dollars per employee per month, applied through coding agents on subscription, plus one or two AI-specific implementations for functions where you know the ROI is real (voice agents for customer service was the concrete example Ara cited), plus a willingness to run multiple models across multiple teams. That is the shape of firms that see gains. Enterprise chat subscriptions alone are the shape of firms that see nothing. If your current AI adoption strategy consists of a Copilot license across the org and nothing else, you are structurally in the low-intensity group. That is not a spend problem. That is an integration problem. Spending more without changing integration will not move you.

If you are a mid-market CTO: Assume model-agnosticism is the base case. Not as a hedge. As the default. If you are a non-US enterprise, this is now a governance requirement. If you are a US enterprise, it is a cost-and-optionality argument. Either way, the multi-vendor stack is where things are heading, and the Router category — whether Ramp’s, or someone else’s — will be structurally important within a year. Start evaluating router platforms now, not after your first surprise vendor-pricing shift. And when you evaluate, look at distribution as much as capability. LiteLLM and Portkey have deeper technical routing tools. Ramp has the payment surface already integrated into your ops stack. Both work. The right choice depends on how deep in the technical stack your team wants to operate.

If you are a manager whose job is coordination: Read the middle-manager finding seriously. Your role is under structural pressure at high-intensity firms. This does not mean you will lose your job. It means the definition of your job is changing, and the compression Ara identified will be dramatic at firms above the integration threshold. The manager roles that will survive are the ones that shift toward systems work — designing the workflows that ICs plus AI can run — and away from translation work. The IC plus AI does not need as much translation. If your day is spent explaining what senior leadership wants to junior staff and vice versa, your job is at risk. If your day is spent designing how the org’s AI-enabled workflows actually work, your job is safer than it has ever been.

If you are hiring: The entry-level surge is real, and it is not just a Silicon Valley phenomenon. Firms with high AI spend are hiring more juniors, more customer service, more sales, more engineering, more admin. If you are trying to fill mid-market roles right now, look at high-intensity AI adopters in your sector before you look at traditional talent competitors. They are hiring across the board. And look specifically at the emerging forward-deployed role — the hybrid business-plus-technical implementer who sits between the customer and the AI system. That role is where the compensation growth is happening, and the talent pool for it is genuinely tight.

If you are a policymaker: Discount both narratives currently dominating the discourse. The CEO layoff narrative is empirically weak at the firm level; the layoffs are happening for reasons that would have justified layoffs anyway — sector rotation, interest rates, over-hiring in 2021, board pressure. AI is the story executives are telling because it is available, not because the mechanism has been established at the firm level. And discount the “data centers create jobs” defense of AI too. Ara is right. It doesn’t land. It shouldn’t land. The correct number to point to is the ten-percent headcount growth at high-intensity adopters. That is a defensible claim. Construction jobs at data centers is not. On the manufacturing displacement question — start caring about that now, not in five years. If AI’s real long-run effect is on physical work while the labor discourse focuses on knowledge work, workers most exposed will be under-defended when the effect arrives.

If you are a European enterprise strategist: The Sovereign Turn is real. You cannot build your product entirely on US frontier APIs and expect the same operating environment in 2028 that you have in 2026. Optionality is the play. Start allocating budget for multi-jurisdiction model access, for on-prem inference options, and for the routing layer that lets you switch without breaking your product surface. European frontier models built via distillation from open weights are a plausible middle path. This is a serious strategic conversation to have with your board this quarter, and every quarter until you have a working sovereign hedge.

If you are an investor: Ramp Economics Lab has quietly become the most useful public dataset for tracking AI adoption in real time. Every AI-related equity call from here forward should include a check against Ramp’s numbers. Every capex thesis for hyperscalers depends on downstream adoption assumptions Ramp is now measuring directly. Every model-vendor thesis is testable against Ramp’s vendor-share data. Use it. Also: pay attention to the Adoption Clock directly. Ramp AI Index growth rate, per-employee spend growth rate, model-serving share growth rate. These are your leading indicators for the physical-layer prints six to twelve months out. Arista August 4. Marvell late August. SK hynix through the F-1 update cycle. Watch the adoption clock; then watch the physical prints. If they diverge, the map is bending.

If you are a builder building AI infrastructure: Study the Router category. It is the least-crowded layer with the highest structural leverage right now. The technical bar is low relative to models or agents. The business bar is entirely about distribution — which is exactly why an incumbent like Ramp is so well-positioned to compete. If you can find a distribution advantage in a specific vertical (healthcare, financial services, manufacturing, government), a vertical router is a defensible business. Horizontal routers will consolidate to two or three players (Ramp, one hyperscaler-native, one Coalition-3-native). Vertical routers will proliferate.

If you are a founder building an applied-AI harness: The Uncaptured Enterprise thesis is confirmed by Ramp’s vendor-share data. Sector-specific harnesses win. Sierra, PolyAI, Granola are the templates. Pick a vertical the frontier labs’ horizontal products don’t serve well. Build the harness. Distribute through the routing fabric (yours or someone else’s). Own the domain logic, rent the model.

Frames

Mental models

Seven models to file this against. They compose. They are not independent claims.

The Threshold Economy. Not every firm gets AI’s gains. Only firms that cross an integration threshold do. Below the threshold: chatbot. Above the threshold: workflow. The threshold is not primarily about dollars. It is about depth of integration and willingness to run multi-vendor stacks. Firms that cross it grow ten percent over two years. Firms that don’t cross it see nothing. This is the empirical finding of the Kharazian-Simon-Stevens paper and the operating principle that structures everything else that follows.

The Sovereign Turn. Enterprise AI buying is bifurcating along jurisdictional lines. US export controls, EU sovereignty framing, Chinese open-weight availability, and American lightweight competitors are all pushing enterprises toward model-agnostic and country-agnostic stacks — even without substitution happening yet in the wallet. Optionality is the driver, not switching. The bear case for the American frontier labs is that non-US enterprises will demand this optionality, and the router-and-hedge infrastructure to serve it is now being built.

The Router as Category (Layer 6.5). The infrastructure that decides which model gets which query is emerging as the layer where enterprise value accrues in a commoditizing frontier. Not the models. Not the agents. The routing decision, and the data that informs it. Ramp is now building this. So are LiteLLM, Portkey, OpenRouter, and others. Ramp’s distribution advantage — 70,000 businesses and $200 billion of spend already flowing through — is the reason to watch Ramp specifically. Distribution beats capability when capability is close to commodity.

Reinstatement at the IC Layer. AI does displace work at high-intensity adopters. It displaces coordination and mentorship work — middle management — while creating complementary demand for entry-level ICs who are AI-native. This is the classic reinstatement effect operating at a specific layer, not across the whole firm. The layoff narrative is looking at the wrong layer. The layer that is losing headcount is above the layer everyone is telling stories about.

The FRED Test. Ramp’s AI Index and its emerging Token Spend Management dataset are quietly building the first real-time telemetry layer for the AI economy — comparable to what the St. Louis Fed’s FRED, ADP’s National Employment Report, or the ISM manufacturing survey do for the broader economy. The measurement infrastructure matters. Every serious argument about AI’s economic effect over the next five years will either use this data or be dismissed for not using it. The transition from “Ramp publishes an interesting index” to “Ramp is where the AI adoption number comes from” is roughly twelve to eighteen months away, on my read.

The Owned/Rented Barbell. Enterprises own the assets they need to control — data, workflows, identity, policy, the routing decision — and rent everything they don’t. Frontier LLMs are rented. Open-weight inference is rented. Model-serving infrastructure is rented. What is owned is upstream (the firm’s data and workflows) and midstream (the routing manifold that decides which rented model gets which owned query). Ramp is positioning as the midstream owner-of-the-router for a meaningful mid-market slice. Expect this to be the dominant enterprise-AI architecture pattern by end of 2027.

The Adoption Clock (fourth clock). The physical clock, financial clock, and efficiency clock from The Fed and the Mint now share the frame with a fourth: the adoption clock. It runs at the pace of enterprise decision-making. Measured monthly in Ramp’s data. Determines whether the physical clock’s capex gets returned. Currently synchronized with the physical clock. Divergence between the two is the single most important macro-AI signal to track over the next 18-24 months.

These seven models compose into a single strategic frame. The Threshold Economy sets the empirical baseline. The Sovereign Turn identifies the strategic pressure. The Router as Category identifies where value accrues. Reinstatement identifies which workforce layer is actually shifting. The FRED Test identifies the measurement authority. The Owned/Rented Barbell identifies the architectural response. The Adoption Clock identifies the macro-cycle synchronization variable. Together they describe the shape of mid-2026 enterprise AI: a bifurcation between firms that cross the threshold and firms that don’t, between jurisdictions that own their AI stack and jurisdictions that hedge, between models that get commoditized and routing decisions that don’t, between adoption paces that validate the physical capex and adoption paces that leave it stranded.

If you file this piece against your existing BE mental library, it lives close to the Uncaptured Enterprise frame, the Verification Bottleneck thread, the Regime-Agnostic Enterprise piece from the Nadella series, and the Fed and the Mint Three Clocks framework. It reinforces the Fed and Mint framework by adding the Adoption Clock. It sits as sibling to the Kimi Delusion piece — where Kimi Delusion is the supply-side counter-narrative on the July 17 selloff, this piece is the demand-side counter-narrative on the CEO layoff story. And it extends the Depth Stack observer-dependence frame by showing that observer position is visible in the wallet — the depth at which each enterprise operates on the customization stack is measurable through the vendor share Ramp reports.

The honest frame

At one point in the conversation Ara said something worth ending on. He said the benefit of Ramp doing this research, rather than OpenAI or Anthropic, is that Ramp has no incentive to say anything positive about AI. Ramp doesn’t make more money when firms spend more on AI. Anthropic and OpenAI obviously do. Ramp’s jobs paper found that AI adoption grows headcount. That was a positive result. But Ara noted, wryly, that he was actually incentivized to find the opposite — because a paper claiming AI destroys jobs would have gotten more press coverage than a paper claiming it grows them.

That is the frame this whole conversation should be understood through. The wallet-level data is telling us a story that is neither the boosterism of the model companies nor the doom of the executive-quote layoff narrative. It is telling us that a small set of firms are quietly getting real gains, that the gains show up as growth rather than displacement, that middle management is the layer being restructured, that the frontier labs face pressure from three directions their PR hasn’t yet acknowledged, that the Router category is emerging as the enterprise-AI architectural moat, and that the payment layer is beginning to organize itself into the same shape.

None of that fits neatly into either camp of the current AI debate.

The AI Supercycle framework I run this publication on rests on three assumptions I want to name against Ara’s data explicitly.

First, that AI diffusion is a long cycle, structurally more like the semiconductor era than the internet era, measured in decades not quarters. Ramp’s data is consistent with this — the time between adoption and gains is 6 to 12 months, and the compounding continues through 24 months without plateau in the current data window. This is a diffusion story with a slow burn. It is not a lightning strike. And it maps onto the six-layer geopolitical outer stack (Geopolitical, Economic, Energy & Resource, Infrastructure, Hardware, Software) that the semiconductor era anchors more naturally than the internet era does.

Second, that the enterprise is uncaptured — that the value of the AI cycle will not accrue entirely to the model layer and will over time migrate upward toward whoever owns the workflow, the data, the routing decision, and the trust surface. Ramp’s Router product is one of the first pieces of infrastructure explicitly built against this thesis. Watch it. And watch the vertical harnesses (Sierra, PolyAI, Granola) that are extending the same logic into sector-specific applications.

Third, that the cycle will unfold under four simultaneous clocks — physical, financial, efficiency, and now adoption — and the ability of the physical clock’s capex to be repaid depends on the adoption clock keeping pace. Ramp is the honest broker on the adoption clock, and Ramp is telling us that the clock is compounding, that a threshold effect exists, that the ten-percent headcount finding is real, and that the discourse describing this cycle is measurably lagging the wallet.

I want to close on Ara’s own note. He was asked at the end of the interview what makes his research distinctive, and his answer was that he tries to publish what the data actually shows regardless of whether it is controversial or intuitive. That framing — data-honest rather than narrative-honest — is what makes Ramp Economics Lab worth taking seriously as a source. Every serious AI-economic argument over the next five years is going to have to engage with this dataset. Every AI-strategy investor call is going to have to check against these numbers. Every policymaker asking whether to intervene in the labor market on account of AI is going to have to reckon with the fact that the firm-level evidence, at least for the population Ramp measures, is at odds with the executive-quote story. Every enterprise strategist has to reckon with the fact that the sovereign turn is not a hypothetical, it is showing up in the wallet, and the router category is now shipping in response.

The Map of AI, as I’ve been building it through the Q2 2026 cycle, has nine layers. Ramp adds a tenth — the telemetry layer — that sits above the stack and reports on what the stack is doing. The Adoption Clock is the tenth layer’s primary output. And the Adoption Clock is the variable that will determine whether every commitment made above and below it, from ASML through Layer 9 governance, actually clears.

The wallet sees different. And starting now, so should we.

Recap: In This Issue!

• Ramp is becoming the real-time gauge of enterprise AI adoption. Unlike surveys that measure what executives say, Ramp measures actual AI spending through payment data. This provides one of the first transaction-based views of enterprise AI adoption.

• High AI adoption correlates with business growth, not contraction. Firms that deeply integrate AI grow headcount by around 10% over two years, while low-intensity adopters show little measurable difference. The key differentiator is implementation quality rather than AI spend alone.

• The integration threshold matters more than budget. Companies do not need massive AI budgets. Moderate spending (roughly tens of dollars per employee per month) combined with workflow integration, coding agents, and function-specific deployments generates the strongest returns.

• The labor market narrative is more nuanced than “AI replaces jobs.” Ramp’s data suggests:

  • Entry-level hiring increases.

  • Engineering, sales, and customer service continue growing.

  • Middle management faces the greatest structural pressure as AI reduces coordination overhead.

• Forward-Deployed Engineers (FDEs) emerge as one of the fastest-growing roles. Organizations increasingly need hybrid technical-business operators who can implement AI inside business workflows, following the model pioneered by companies like Palantir, Anthropic, and OpenAI.

• AI buyers are increasingly demanding model optionality. Enterprises are becoming:

  • model-agnostic,

  • provider-agnostic,

  • and increasingly jurisdiction-agnostic.
    This is driven not only by pricing but by geopolitical concerns around API access and export controls.

• The Router becomes a strategic layer of the AI stack. Ramp’s new Router product reflects a broader shift: as models commoditize, value moves toward intelligently routing workloads across multiple models based on cost, performance, and sovereignty requirements.

• OpenAI and Anthropic face pressure from three fronts simultaneously:

  • Chinese open-weight models.

  • Lightweight American competitors (e.g. Cursor Composer, Gemini Flash).

  • Enterprise demand for sovereign, multi-vendor architectures.
    None individually displaces frontier labs today, but together they reduce lock-in.

• Anthropic is leading enterprise momentum. Ramp’s enterprise data shows Anthropic overtaking OpenAI in enterprise penetration, largely driven by Claude Code and developer workflows rather than general-purpose chat usage.

• Chinese models remain strategically important but commercially limited today. Their primary impact is increasing competitive pressure and encouraging open architectures rather than replacing OpenAI or Anthropic inside enterprises.

• The “wallet layer” validates the AI infrastructure thesis. Ramp’s enterprise adoption data supports continued demand for:

  • hyperscaler AI infrastructure,

  • Nvidia GPUs,

  • HBM memory,

  • networking,

  • and AI datacenter investment.
    Enterprise demand is still accelerating rather than plateauing.

• The AI Supercycle gains a fourth clock. Beyond physical infrastructure, financing, and efficiency improvements, the author introduces an Adoption Clock—the pace at which enterprises actually implement AI. This becomes the critical indicator for validating AI infrastructure investments.

• Small businesses and large enterprises benefit differently.

  • Small firms gain speed and experimentation advantages.

  • Large firms eventually leverage proprietary data and distribution.

  • Mid-sized companies are the most exposed, lacking both agility and enterprise scale.

• Manufacturing may ultimately experience greater AI disruption than white-collar work. While public debate focuses on office jobs, the long-term automation opportunity appears larger across physical industries through robotics, vision systems, logistics, and industrial AI.

• Ramp is evolving into AI’s “FRED.” Just as FRED became a standard source for macroeconomic data, Ramp’s transaction-level AI spending data could become the reference dataset for measuring enterprise AI adoption and validating broader AI investment theses.

With massive ♥️ Gennaro Cuofano, The Business Engineer

This is part of The AI Supercycle Series. You will find the curated version of the live in the AI supercycle in the coming days.

That was the through-line of our latest AI Supercycle session with William from the Exponential View team, whose State of the AI Economy report is one of the most rigorous attempts I’ve seen to answer, at the level of financial mechanics rather than vibes, one deceptively simple question:

Where is the money in AI actually coming from — and is any of it real?

Their answer, stripped to the bone: demand is real, it is bigger than most narratives give it credit for, and it is arriving faster than any prior tech wave. But how that demand is being funded, absorbed, and priced is where the interesting structural story lives.

Here is the compressed version of the conversation — what it means, and why it matters if you’re trying to make sense of this cycle for the next decade, not the next quarter.

The State of the AI Economy

The deduplication problem — and why “$110B in revenue” is actually an honest number

The first thing Exponential View did — and it is the reason their report is worth reading in full — is refuse to accept the industry’s own accounting at face value.

The AI stack has a circularity problem. A Cursor or a Harvey generates revenue from an end user, then spends most of it on tokens from OpenAI or Anthropic. Those labs generate revenue, then spend it on compute from hyperscalers and NeoClouds. Those hyperscalers and NeoClouds generate revenue, then spend it on chips from NVIDIA. Each transaction shows up on someone’s income statement. If you naively stack them, you double- and triple-count the same underlying dollar as it flows down the stack.

Exponential View built bottom-up financial models — P&Ls and cash flow statements — for every major player in the stack, and every meaningful minor one. They then deduplicated the revenues across layers to isolate genuine external demand from internal recycling.

The number that survives that filter: ~$110B of real AI revenue in the last 12 months, now trending toward $175–200B on current momentum.

Why this matters: the loudest critique of the AI economy — “it’s all circular, they’re just paying each other” — is the exact critique this methodology was designed to test. It doesn’t fully vanish. But it turns out to be a much smaller share of the whole than the critics assume, and the residual, deduplicated demand is very large and growing very fast.

The Compression: The bear case that “AI revenue is just financing dressed up as demand” survives contact with the data only in a much narrower form than it’s usually stated.

Nobody in this stack is a static actor

The most useful mental model I took from William’s analysis is this:

Every layer of the AI stack is trying to eat the layer above it and the layer below it — simultaneously.

Foundation model labs are moving up into applications. Claude Code and Codex are direct assaults on Cursor and the coding-assistant category. Anthropic has now shipped Claude for Teacher into education. OpenAI is pushing into legal, finance, and productivity verticals. These are not accidents. They are structural responses.

Meanwhile, the application layer is moving down toward the frontier. The Cursor–xAI deal is the canonical example. Cursor became the largest Anthropic API customer, then found itself directly competing with Claude Code. The path forward wasn’t to sit still — it was to acquire compute (through the xAI/SpaceX deal), leverage the enormous corpus of agentic traces from Cursor’s own users as post-training signal, and push a Cursor-native frontier model (Composer) into production. That is vertical integration under duress. And it works because the harness — the agentic scaffolding around the raw model — has become nearly as important as the model weights themselves.

This is the pattern to watch in every boardroom right now: do we tie ourselves to a hyperscaler and a lab, or do we vertically integrate? Apple tried the closed path and reversed itself. AWS launched Bedrock. Almost every serious operator is running this calculation in Q4 2026.

The Compression: In the AI Supercycle, no one accepts their assigned layer. Every player is negotiating vertical integration or being negotiated into it.

CapEx, depreciation, and busting the “hyperscalers are just lighting money on fire” narrative

The day before we recorded, TSMC announced another ~$260B in capacity commitments. Every week the CapEx numbers move. And every week the same critique returns: “these companies are spending more than they earn — it’s a bubble.”

William’s response is the most honest reframing of the CapEx question I’ve heard:

CapEx is a stock, not a flow. When Microsoft or Meta commits $80B this year to AI infrastructure, they are not expected to earn that back this year. They are building an asset with a depreciation schedule — roughly 6 years for chips, 14 years for the rest of the data center envelope (building, cooling, power). Each year, only a fraction of that spend passes through the income statement as a cost. But that fraction accumulates on top of every prior year’s fraction. That is the payback pressure that matters — not the headline number.

Similarly, Remaining Performance Obligations and forward energy commitments are the market pricing in future demand. They look terrifying next to today’s revenue precisely because they are not meant to be compared to today’s revenue.

The deduplicated-revenue methodology is what lets you actually answer the question everyone is really asking, which is not “is the CapEx big?” (it obviously is), but rather:

Is the accumulated depreciation stack going to be paid back by the real, external demand — and on what timeline?

That question has a more nuanced answer than either the bulls or the bears usually want to admit.

The Compression: You cannot judge the AI CapEx cycle by comparing this year’s spend to this year’s revenue. You have to compare the depreciation waterfall to the demand trajectory.

Why this wave grew faster than any prior wave — and why that changes everything

One of the most important slides in the report is the one that plots AI investment and revenue against the internet, mobile, and cloud waves.

AI has grown much, much faster than any of them. That is not a qualitative claim — it is a measurement.

The reason is not primarily that AI is a “better” technology than the internet. The internet was arguably more foundational. The reason is compounding: AI is arriving in a world that already has the internet, mobile, cloud, ubiquitous payments, mature distribution, and instant global software delivery. Every prior wave is now infrastructure for this one. What took the internet a decade to distribute takes a foundation model six months.

This changes everything downstream, but the most under-discussed consequence is societal: we and our institutions had roughly ten years to make sense of the internet before it reshaped work, media, politics, and family life. We are being given twelve to twenty-four months for AI. Our institutions were caught off guard by the internet even at its slower pace. With generative AI, that adjustment gap is going to be far more violent.

If you are still using “the .com bubble” as your reference model for this cycle, you are underestimating both the upside and the systemic pressure.

The Compression: AI compounds on top of every prior tech wave. That is why the revenue curve is faster, and also why the societal adjustment is harder.

The bottleneck cascade — energy, chips, capital

The rapid demand growth doesn’t just show up as revenue. It shows up as bottlenecks, which is the word the industry keeps involuntarily returning to.

• Chips: the semiconductor industry is deeply cyclical, used to peak-and-trough dynamics. AI demand is not one of those cycles. It is an order of magnitude larger and structurally exogenous to the industry’s own cycle. This is why TSMC keeps revising up.

• Energy: the US power sector plateaued for years. It is now the binding constraint on data center deployment, and grid planning has to be done for a future state, not a present one.

• Capital: this is the one people underestimate most. In 2022–2024, hyperscaler CapEx was overwhelmingly funded from existing cash. Hyperscalers are cash-rich. But the pace of growth has now outrun internal cash generation. External capital is filling the gap.

That last point deserves its own section, because it is where systemic risk actually enters the picture.

The financing shift — and the NeoCloud SPV problem

When AI CapEx was being funded out of hyperscaler cash flow, an underperformance risk was contained: the share price of a specific company might take a hit, but the wider market was insulated.

External capital changes the picture. Debt, private credit, and — most interestingly — Special Purpose Vehicles at the NeoCloud layer are doing an increasing share of the funding. The NeoCloud business model was designed precisely for this: gather capital in an SPV, buy the GPUs, hold them off the primary hyperscaler’s balance sheet, and rent capacity back.

From Microsoft’s perspective, it would be nearly irrational not to use this structure to add compute right now. From the system’s perspective, it means CapEx risk is being pushed off the balance sheets of the strongest players and into a more distributed set of counterparties.

This is fine while demand keeps compounding. It is where the interconnectivity risk lives if any part of the cascade slows.

The Compression: The AI CapEx race has moved from cash-funded to externally-funded, which means the systemic interconnectivity of the industry is materially higher than it was 18 months ago.

Why the bubble, if it pops, won’t pop the way the .com bubble did

I’ve been saying this for a while and William’s data reinforced it: the mental model of “one big bubble that inflates and pops” is wrong for this cycle.

The reason is the same reason we call it a Supercycle. This isn’t a single S-curve. It is a stack of overlapping S-curves — chips, energy, models, applications, agents, verticals — each with its own bottleneck, its own capital demand, and its own timeline.

What happens instead is a sequence of smaller, localized re-pricings. Supply catches up to demand in one layer, prices collapse in that layer, capital rotates. Then demand pushes into the next bottleneck. The system as a whole keeps growing, but individual layers experience cyclical pain that looks like a bubble popping in miniature.

If you are pattern-matching to 2000 or 2008, you will misread the next five years.

Value is moving up the stack — finally

For the last three years, “who is making money in AI?” had one answer: NVIDIA, then memory (SK Hynix, Samsung), then the labs. The application layer was widely dismissed as “just ChatGPT wrappers” — thin margin on top of a commoditized token spend, no defensible value add.

That is changing, and William’s data captures it. What people pay for Cursor is not a pass-through of their token spend. It is not a subsidy on their token spend either. It is a premium on their token spend. That premium is the app layer starting to earn genuine margin — because customers are getting savvier, because the product surface around the model matters more than people realized, and because the harness (context, memory, tools, agentic scaffolding) is where a lot of the actual utility lives.

This is what value migration up the stack looks like in real time. It is exactly what the .com era did with its own layers over ten years. This one will do it faster.

The precondition is competition strong enough to prevent any single app-layer player from setting pricing power. That competition exists.

The Compression: We are at the transition point where the app layer stops being wrappers and starts being defensible products with real margin.

The Jevons paradox is already showing up in the token data

Token consumption has grown in a compounding fashion. Some of that is adoption — more users, more use cases. But a large share is the new modes of use: multi-turn chat, reasoning, and above all agentic execution, which drives massive tool-use per prompt.

Exponential View is tracking agentic intensity as a proxy — the extent to which individual prompts result in tool calls. It is ticking up. Cheaper tokens are not reducing token consumption. They are enabling entirely new workflows that consume vastly more tokens per user session.

The methodological innovation worth flagging: William’s team proposes measuring quality-adjusted output tokens — strip out input and reasoning tokens, weight the remaining output tokens by the benchmark quality of the model that produced them. This is a much better proxy for real user value than raw token counts, and it removes the artificial inflation from models spending more time reasoning.

Expect that to become a standard metric.

The consumer surplus story that gets left out of every earnings call

The most important reframing of the entire conversation came near the end.

Almost every AI economy debate is about revenue — will it pay back CapEx, will it pay back VC, will it justify the valuations. Almost no one talks about consumer surplus — the value users are extracting from these tools that isn’t captured in anyone’s revenue line.

Erik Brynjolfsson’s Stanford team has been running the “how much would you need to be paid to give up your access to AI?” study — the same methodology used to value Google Maps and other free internet goods. The consumer surplus is enormous. It doesn’t show up in GDP. It certainly doesn’t show up in the revenue models.

The Google Maps analogy is exact: the paper A-to-Z road atlas added to GDP when you bought it. When it was replaced by a vastly better product that people got for free, the GDP contribution went to zero. The world got better. The measurement got worse.

The same thing is happening now, at scale, in real time. Even if the AI CapEx race under-earns for its funders, the value being delivered to users — in their work and in their lives — is real and permanent. There is no going back. Every knowledge worker who has adopted these tools knows this in their bones.

The Compression: A growing generative AI economy is not automatically good news for OpenAI, Anthropic, or the hyperscalers. But it is unambiguously good news for consumers — and that value doesn’t reverse.

What to take home

If you read only one section of the State of the AI Economy, read the deduplicated revenue methodology and the CapEx depreciation waterfall. Those two pieces together are what separate serious analysis of this cycle from the noise on both sides.

The one-line synthesis from William that I’ll be quoting for a while:

“This economy is growing from consumer demand. We can see that unprecedented revenue growth, and there are really quite significant balancing acts where CapEx and the depreciation hurdle meet the pressure of selling tokens at continually collapsing prices for a given intelligence level. All of those competing dynamics make this a much more finely balanced economy — even though the overall consumer benefit continues to grow apace.”

That is the shape of the AI Supercycle in one sentence. Balanced tension all the way down the stack. Real demand. Real payback pressure. Collapsing unit prices. Growing consumer value. And nobody sitting still.

Report: The State of the AI Economy, by Exponential View. Guest: William, Exponential View team.

Hosts: Gennaro (The Business Engineer) and Joel (Leadership in Change).

With massive ♥️ Gennaro Cuofano, The Business Engineer

Information followed the same trajectory. Colloquial word for millennia. Then Shannon, 1948, and information became a measurable quantity with formal properties. The entire twentieth century of computing, communications, and biology rebuilt itself on the new representation.

Knowledge is next in this sequence, and the Infinite Alphabet is Hidalgo’s proposed change of representation. It is the argument that knowledge stops being a single undifferentiated substance the moment you look at it closely, and that if we want to understand how the economy actually works — how it grows, how it moves, how it produces value — we need vocabulary for a substrate that is discrete, non-substitutable, and combinatorial.

Hidalgo’s intellectual journey into this frame is worth naming. He published Why Information Grows in 2015. Between 2015 and 2019 he prepared a Harvard course on the principles governing the growth, diffusion, and value of knowledge, discovering — in his own words in our conversation — that there was a “larger context” of historical precedent his prior work had been operating inside without seeing. He signed the book contract in 2022, wrote it in two years, and published it in 2025. The Infinite Alphabet is the mature framework. It rests on two supporting structures.

Three types answer what knowledge is.

• Factual knowledge — Rome is the capital of Italy. “Little sound bites that collect a little bit of truth,” in Hidalgo’s phrasing. This is what pretraining corpora largely contain, and what RAG systems handle well.

• Conceptual knowledge — the detective’s theory that connects the bullet hole, the 3 AM phone call, and the blood sample into a coherent story. This is what reasoning models are trying to be.

• Procedural knowledge — running the DNA sequence in the lab. This is where agentic AI is landing.

Three properties answer how it behaves.

• Non-rival — copying does not deplete. Teach me a song, you still know it. This is the property that makes AI economically explosive.

• Tacit vs. explicit — you can watch lectures about basketball, but if you don’t play with people who are good, you will never be good.

• Non-fungible — a candy-factory chemist and an oil-mine geophysicist both have chemistry knowledge, but neither can do the other’s job.

The metaphor is Scrabble. Not one big undifferentiated pile of letters that all mean the same thing, but an alphabet of specific, non-interchangeable pieces. Each capability is a letter. Each product, firm, or industry is a word built from a specific combination of those letters. The economy is a game played across geographies and centuries where players attempt to spell increasingly complex words using letters that are constantly being invented, forgotten, and traded.

The claim of this piece is that Hidalgo’s representation change is the right lens through which to read the two frameworks anyone building in AI right now already carries in their head — the AI Supercycle and the Map of AI. Both frameworks become sharper, more predictive, and more actionable when you stop treating “AI” as a homogeneous mass and start treating it as a Scrabble game played simultaneously on nine boards.

Also follow Cesar on X or LinkedIn.

You can also follow his work on Datawheel.

The Infinite Alphabet as Paradigm

Subscribe To The AI Supercycle

Hidalgo’s articulation in the interview was compressed but precise:

“It’s not a single thing. It’s not like a big jello that is undifferentiated, but a lot of little components that are complementary but you cannot substitute one for the other. So they’re like Scrabble tiles.”

Read that sentence twice. The whole edifice sits on the two words at the center: complementary and non-substitutable.

Complementarity means the letters need each other to form words. A on its own is not English. A single tile does not make a game. The economic value of any single letter is nothing until it is combined with the right other letters, in the right order, at the right scale.

Non-substitutability means you cannot swap one for another. A candy-factory chemist cannot do the job of an oil-mine geophysicist. A frontier-model researcher cannot do the job of a chip-packaging engineer. Both possess “chemistry knowledge.” Both possess “engineering knowledge.” The categories are so broad they are useless. The actual letters are specific and locked to specific practices.

Hold these two properties together — complementary + non-substitutable — and three consequences fall out immediately.

• The economy is combinatorial, not additive. More “knowledge” does not produce more output unless the specific letters combine into words the market can use. This is why intelligence-heavy economies do not automatically grow, and why capability-heavy AI models do not automatically produce enterprise value.

• Specialization is inevitable. Since letters are non-substitutable, players end up mastering only a few of them each. Individuals become specialists. Firms become specialists. Cities and countries become specialists. The distribution of who-holds-which-letters becomes an economic geography in itself.

• Value lives at the joints. The letters are cheap. The combinations are where value emerges. The historically dominant firms are not the ones who mastered a single letter — they are the ones who assembled the most useful combinations of letters others held separately.

The physical letters AI does not yet touch

Subscribe To The AI Supercycle

Hidalgo made a point in the interview that is worth putting on the record because it puncturеs a specific kind of AI-industry hubris. He had just moved apartments the previous weekend. A dishwasher hose got squished during transit. He needed to identify and replace it. He could photograph the part. He could not, from a chair, do the physical repair. The lamp he was hanging from the ceiling — no chat model helped with the ladder shake.

A large fraction of the alphabet is physical. Electrical work, plumbing, tiling, load-bearing carpentry, mechanical repair, the tacit knowledge of a surgeon’s hand, the rhythmic knowledge of a chef’s stir — these are letters in the alphabet, and none of them have entered the AI-legible portion of it yet. Robotics may cross that frontier in the coming decade. Until then, the AI industry’s obsession with the letters it can see is a systematic underestimate of the letters that still run the actual economy.

The point matters for how you read every AI industry chart. When capability curves appear to be closing in on human performance, that closure is being measured against the fraction of the alphabet AI can currently address. The unmeasured fraction is much larger.

The AI Supercycle Through the Infinite Alphabet

The AI Supercycle argues that we are in a 30–50 year civilizational transformation, structured as three nested cycles running at different clocks: a 5–10 year short cycle inside a 10–20 year medium cycle inside a 30–50 year long cycle. It is a framework designed to hold simultaneously the fact that AI is a real durable transformation and the fact that the current market is a bubble, without collapsing either.

Under the infinite alphabet, the framework becomes mechanistically legible.

The S-curve stack, confirmed on air

Subscribe To The AI Supercycle

The load-bearing exchange of our conversation happened here. Hidalgo laid out the standard Thurstone-Wright result: at the level of individuals and teams, learning follows a bounded power law — fast at the start, slower thereafter, until doubling the practice to gain one more unit of proficiency becomes uneconomic. But at the level of industries and economies, the aggregate looks exponential. Moore’s Law is the canonical example. So is the price of light. So is genetic sequencing.

The mechanism is that the exponential is not a single curve. It is a cascade of overlapping S-curves, each individually bounded, each handing off to the next before the prior one saturates.

I put the AI-industry version of this to Hidalgo directly on the record. We have run four scaling paradigms in relay between late 2022 and mid-2026:

• Pre-training (2020–23)

• Test-time compute and reasoning (2024–25)

• Agentic loops (2024–26)

• Continual learning (emerging, 2026 onward)

Each of these obeys Thurstone. Each individually peters out. The reason the industry appears to be on a straight exponential line is that the next curve started producing capability faster than the prior one could saturate.

Hidalgo’s reply was direct: “Exactly, it’s a bunch of S-curves that are overlapping.” The best strategy for any player — individual, firm, or nation — is to make sure they are on the next curve, because a new one will always come and displace those who mastered the current one.

This is the mechanical foundation of the AI Supercycle thesis. The three-year AI industry looks like the last seventy years of transistor manufacturing compressed by a factor of twenty. The exponential appearance is the sum of finished curves plus starting curves. When a starting curve fails to arrive, the exponential stops — not gently, abruptly.

The bubble-vs-supercycle paradox, resolved

Subscribe To The AI Supercycle

The bubble is what happens when capital treats non-fungible knowledge as fungible. When the market prices “AI” as a single asset class, it is making the jello error at scale. Every AI company gets valued as if it shared the same substrate, the same customers, the same defensibility. In alphabet terms: every word in the market is treated as if it were spelled from the same three letters.

The supercycle is what is actually happening underneath the bubble. The alphabet is expanding. New letters are being invented. Old letters are being commoditized. Firms are re-sorting into new positions on the alphabet, and the sort will take a generation to complete.

Both statements are true simultaneously. The bubble is the price signal misreading the substrate. The supercycle is the substrate itself, which does not care what the price signal says.

The three nested cycles as rates of alphabet change

Subscribe To The AI Supercycle

Each of the three cycles corresponds to a different rate of alphabet change.

• The 5–10 year short cycle is the rate at which new letters enter mainstream production. Pre-training scaling was one letter. Test-time compute reasoning is another. Agentic loops is a third. Continual learning will be a fourth. Each new letter appears every 12–18 months, gets picked up by everyone within two years, and becomes table stakes within three. This is the cycle that produces the sense of “AI moving fast.”

• The 10–20 year medium cycle is the rate at which industries re-sort their letter portfolios. Media reshapes around the new letters. Finance reshapes. Legal reshapes. Retail reshapes. Each industry moves at its own pace, but the aggregate reshaping runs on a two-decade clock, because it requires organizational restructuring, workforce transition, and regulatory adaptation. This is the cycle that produces the sense of “AI is transforming the economy.”

• The 30–50 year long cycle is the rate at which the alphabet itself becomes part of the ambient cultural substrate. The way electricity did between 1880 and 1930. The way the internet did between 1990 and 2020. The final phase where the letters stop being “AI letters” and become just letters. This is the cycle that produces the sense of “civilizational transformation.”

The three cycles are not phases in sequence. They are running in parallel, at different clocks, on the same underlying alphabet. The short cycle spits out new letters. The medium cycle re-sorts who holds them. The long cycle absorbs the whole apparatus into the substrate of civilization.

The Zero Industrial Revolution — printing press as the timescale anchor

Subscribe To The AI Supercycle

Hidalgo’s most compressed historical argument in the interview was to reframe the printing press as the “Zero Industrial Revolution” — the original scalable-product transformation, running earlier than the steam-and-textile revolution and generating a comparable economic surface.

The mechanism was extraordinary. A printed book in the 1400s cost a fraction of a house. Five people in a basement, printing for four years, could produce the value of three hundred houses. The economics were entrepreneurial in exactly the way software would be five centuries later. Print shops proliferated across Europe within fifty years of Gutenberg. The knowledge diffused from the Frankfurt/Mainz cluster along the Rhine and outward. By the late 1400s the per-capita number of print shops had stabilized — printers were competing to survive, and cheap easy-to-print items like church indulgences became the profitable margin work.

Then Luther, 1517. Ninety-five theses. Millions of copies sold — “even for today’s standards a lot,” in Hidalgo’s phrasing. By the 1530s, less than fifteen years later, Europe was in the middle of the Counter-Reformation. Institutions built over a thousand years were fundamentally restructured in a decade and a half.

Hidalgo’s calibration: if you match AI 2016–2030 against Reformation 1516–1530, the two transformations run on the same clock. This is not a metaphor. It is a schedule. Two thousand sixteen because that is when the AlphaGo/DeepMind moment landed and the modern AI era began institutionally, not because ChatGPT is the correct start point. And the perceived slowness of past revolutions, in Hidalgo’s phrasing worth quoting almost in full: “we tend to think that life is fast because the speed of memory is really quick... but the world is always going at full speed.”

The printing press felt exactly like AI feels now, to the people living through it. And they had no idea, in 1520, whether they were at the beginning or the end.

The absorptive-capacity bottleneck — the eight-month Claude Max subscription

Subscribe To The AI Supercycle

I raised in the conversation a phenomenon that anyone working with large enterprises will recognize: individuals on the frontier have moved to sub-agents, MCP servers, and agentic loops, while the enterprises around them still treat AI as a chat window. Hidalgo answered with a Sony story from 1948.

Masaru Ibuka was trying to build a tape recorder in postwar Japan. He had seen the American version, but had no manual. Working with a frying pan, black-market chemicals, and hemp paper, he and his team reverse-engineered magnetic tape. The point Hidalgo drew from the story is that absorption of external innovation requires internal capability that pre-exists it. Ibuka’s team had radar and telecommunications engineers from the war. Without that absorptive capacity, the American tape looked like magic. With it, it looked like a problem to solve.

The concept is Cohen and Levinthal, 1990: firms invest in R&D not primarily to produce research, but to become able to absorb the research others produce.

Hidalgo’s own example from our conversation — cutting, and directly quotable — is that if you are in an organization where “it takes you like eight months of meetings and procurement nightmares to get one Claude Max subscription, well, you’re going to fall behind.” In an S-curve stack world where each curve runs 12–18 months, eight months is half a paradigm. Any enterprise that pays that tax is systematically behind by one full cycle at all times.

The absorptive-capacity gap is the reason enterprise AI adoption feels like a different technology from frontier AI. It is not that enterprises are stupid. It is that they have not built the internal capability to reverse-engineer, and their procurement infrastructure was designed for a paradigm where innovation moved on quarterly cycles rather than fortnightly ones.

Disruption and architectural knowledge — the airport-baggage warehouse

Subscribe To The AI Supercycle

The classic supercycle transition signal — the moment when an incumbent’s dominance evaporates because the ground beneath the product changed — is not a competitive event under the infinite alphabet. It is an alphabet event.

Blockbuster did not lose because Netflix had better movies. Blockbuster had clerks who knew how to open a store, greet customers, and restock shelves. Netflix from day one was organized around shipping DVDs directly, with different roles, different communication patterns, different operational architecture. As Hidalgo put it in the conversation, an Amazon warehouse “looks more like the part of an airport where you are sorting bags than a place where you drink coffee.” The transformation from “we rent movies” to “we ship DVDs” was not an adaptation. It was an alphabet swap.

The current AI corollary: any enterprise that tries to bolt AI onto an existing workflow is spelling their old word with a new tile inserted. It rarely works. The organizations that succeed with AI are the ones that treat it as an alphabet event — a reason to rebuild the workflow, the roles, and the operational architecture from scratch around what the new letters make possible.

The Map of AI Through the Infinite Alphabet

Subscribe To The AI Supercycle

The Map of AI describes the industry as a nine-layer stack — Energy & Physical, Foundries & Packaging, Silicon, Networking & Protocols, Compute Capacity, Foundation Models, Agentic Harness, Distribution Surfaces, and Governance. Each layer is a distinct arena, with its own physics, its own binding constraints, its own profit pool, and its own players.

Under the infinite alphabet, each of the nine layers becomes a distinct alphabet.

This is not a semantic reframing. It is a structural claim. The letters at Layer 1 (Energy & Physical) are non-fungible with the letters at Layer 6 (Foundation Models). A firm that has mastered thermal management, power delivery, and land acquisition for gigawatt-scale sites holds a specific alphabet of letters that cannot be substituted for the alphabet held by a firm that has mastered transformer architecture, pretraining data curation, and RLHF. Both are “AI” businesses. Neither can do the other’s job. Neither can even legibly evaluate the other’s work at the depth required to compete.

Three consequences follow.

No player can master all nine alphabets

Subscribe To The AI Supercycle

The historical incumbent instinct — vertical integration across the stack — assumes the letters are compatible. Under the infinite alphabet, they are not. Each layer’s mastery requires a decade of specialization inside that specific alphabet. A firm that tries to hold Layer 1 through Layer 9 ends up mediocre at all of them, because the tacit knowledge required for silicon does not transfer to distribution, and the architectural knowledge required for foundation models does not transfer to energy siting.

This is why the vertical-integration attempts of the current cycle keep hitting the same wall. The firms that appear to be spanning multiple layers are almost always specialists in one layer who are renting access to adjacent ones — the way NVIDIA rents distribution through hyperscalers, and OpenAI rents compute from Microsoft.

Knowledge agglomerates — it does not diffuse

Subscribe To The AI Supercycle

Hidalgo’s musician analogy is the cleanest articulation of why knowledge concentrates in specific geographies and refuses to spread through investment alone. Imagine an itinerant musician who plays a specific instrument in a specific style, walking city to city looking for other musicians whose instruments complement hers and whose style matches. She fails, fails, fails — until she finds the place where she matches, and she stays.

Knowledge is agglomerative, not diffusive. This is why Silicon Valley remains Silicon Valley, why Shenzhen remains Shenzhen, why Paris and Cambridge and Tel Aviv and Bangalore concentrate. The complementary letters are already there. New letters seek out existing ones. Once the density is built, it compounds.

The policy corollary is devastating for the “AI city in the desert” thesis:

• Yachay, Ecuador. A billion dollars invested in building a city of knowledge in the high Andes, two hours from Quito. Went nowhere. The letters were not there and could not be summoned by real estate.

• Neom, Saudi Arabia. Five hundred billion dollars for a linear city in the desert. As of mid-2026, being repositioned as a data center. Hidalgo’s aside in the interview was pointed: the project’s ambition collapsed into infrastructure because the human alphabet never showed up.

• Zhongguancun, Beijing. The counter-example. China did not put money into buildings in empty places. It focused on central business districts that already had density, created funds that invested in companies rather than campuses, and targeted sectors at the beginning of a window of opportunity. It worked.

The diagnostic is generalizable. Any AI-cluster policy that spends on physical infrastructure in a place without pre-existing complementary letters is Yachay. Any policy that invests in companies where the density already exists is Zhongguancun. The name changes each generation; the outcome does not.

The forgetting rate — 3 to 6 percent per month

Subscribe To The AI Supercycle

The most quotable statistic in the interview, and one I have not seen circulate anywhere else: firms lose 3 to 6 percent of their knowledge every month. Hidalgo cited this from research on organizational forgetting, and referenced his own January 2026 Financial Times article on the topic.

The number compounds. At 3% monthly, 30% annual attrition. At 6% monthly, over 50%. The reason we do not perceive this loss is that learning offsets it in real time. Stop learning for eighteen months, and half the institutional knowledge is gone.

Hidalgo’s phrasing was surgical: “knowledge is like a muscle — the moment that you stop exercising, it becomes weak.”

The consequences ripple across the AI stack in ways that are becoming visible right now.

• Memory manufacturing. In our conversation I raised the current build-out of AI-era cloud infrastructure — data centers designed for CPUs having to be rebuilt for GPUs. The unexpected bottleneck is memory. As Hidalgo noted, for decades the industry treated memory as solved and stopped innovating. Now only three companies make memory for GPUs, and demand is dwarfing supply. The forgetting was not passive; it was the market signaling that memory was a commodity, and the commodity signal caused the muscle to atrophy.

• LCD panels as the general case. Hidalgo used LCD panel generations as the illustrative case. High-complexity products go through generational cycles. Firms that miss a single generation drop out of the market permanently, because the accumulated tacit knowledge required to compete has grown faster than they could catch up. Consolidation from many firms to a handful is the terminal state.

• Nuclear plants and mega-projects. The West stopped building nuclear at scale in the late 1970s. Forty-plus years of continuous forgetting has produced the current situation: the West has to relearn nuclear construction, and it is genuinely difficult, because the knowledge is gone. The same pattern is visible in rail, high-voltage grid engineering, high-precision machining, mega-bridges. Meanwhile China, which has been building continuously for thirty years, holds these alphabets fluently.

The muscle argument means preservation requires continuous production, not archival. You cannot warehouse capability. You cannot document your way out of forgetting. The only way to keep a letter of the alphabet is to keep writing it.

The three geometries as combinatorial strategies

Subscribe To The AI Supercycle

The observation that three distinct cross-stack geometries — vertical, horizontal, and flywheel — dominate the current cycle is exactly the pattern the infinite alphabet predicts.

• Vertical geometry (Google’s owned column) is the strategy of holding a full set of letters across as many layers as possible. It maximizes control but concentrates all the alphabet-acquisition cost inside one firm. It works only for firms with two decades of accumulated cross-layer capability.

• Horizontal geometry (NVIDIA’s rent-extracting band) is the strategy of dominating one layer’s alphabet completely, then charging every player at every adjacent layer for access. It maximizes leverage but exposes the firm to any competitor who develops equivalent letters in the same layer.

• Flywheel geometry (SpaceX-style multi-engine compounding) is the strategy of holding partial alphabets across several layers that reinforce each other, so each layer’s letters make the next layer’s letters more valuable. It maximizes compounding but requires enormous patience and cannot be replicated by acquisition.

These are not three ways of doing the same thing. They are three fundamentally different theories about how to spell the current word of the industry. Each geometry bets on a different structural feature of the alphabet. The cycle will not go to whichever firm builds the best model. It will go to whichever geometry correctly reads which parts of the alphabet are becoming scarce.

Governance as fence, not letter

Subscribe To The AI Supercycle

The most subtle claim in the Map of AI is the treatment of Layer 9, Governance. It is drawn as the ninth layer, but it does not behave like a layer. It behaves like a fence around the other eight.

Under the infinite alphabet, this makes precise sense. Governance is not a set of letters. Governance is a set of rules about which combinations of letters are legal to spell. Export controls on advanced silicon are not a Layer 3 problem. They are a rule about which Layer 3 letters can be combined with which downstream letters by which players in which jurisdictions. Content policy on frontier models is not a Layer 6 problem. It is a rule about which Layer 6 letters can be combined with which Layer 8 distribution surfaces.

The strategic implication is severe. Any firm building a moat inside the eight technical layers is building it inside a fence whose boundaries can be redrawn overnight. Governance is not a layer to be dominated. It is a constraint to be modeled, and firms that do not model it as a first-class variable in their strategic planning are exposed to a shock they cannot see until it lands.

What the Alphabet Predicts

Subscribe To The AI Supercycle

Two clean predictions fall out of the frame once you hold it steady.

The commoditizing letters. Whatever letter enters mainstream production today will be table stakes within three years and worthless as a moat within five. This is what commoditization means under the infinite alphabet: the letter becomes so widely held that it stops being a differentiator and starts being a required input. Pre-training capability commoditized in 2024. Reasoning capability is commoditizing through 2026. Agentic capability will commoditize by 2028. Nothing that can be scaled and shared stays a moat.

The appreciating letters. Whatever letter cannot be scraped, scaled, or shared appreciates in value as the commoditizing letters lose theirs. Tacit knowledge in specific industries. Non-fungible expertise locked to specific firms. Regulatory positioning locked to specific jurisdictions. Distribution surfaces locked to specific user relationships. Physical-world capability that lives in the parts of the alphabet AI cannot yet read. These letters do not commoditize because they cannot be produced at scale. They can only be assembled slowly, one relationship at a time.

The strategic move that falls out of this is sharper in the alphabet framing than in any other frame I have seen: identify the specific letters your firm holds that are non-scrapable, non-scalable, and non-shareable. Everything else is rented substrate. Whatever you own that is scarce is your position. Whatever you rent that is abundant is your cost line.

The winners of the next decade of the AI Supercycle will not be the firms with the best models, the largest compute budgets, or the strongest brand recognition. They will be the firms that correctly identified which specific letters were becoming scarce, and assembled a defensible position around those letters before the rest of the market noticed.

The physics that closes the argument

Subscribe To The AI Supercycle

Hidalgo ended the conversation with the mental model he uses himself, and it is the correct synthesis to close on. He described a continuity between physics, biology, and economics — and offered a single operating frame for reading any economic system:

Energy is the input. Prosperity is the output. Knowledge is the engine in between.

The question worth asking of any economy is how much knowledge is produced per unit of energy, with what side effects (emissions being the critical one), and what that ratio implies for the prosperity the system can generate.

Apply this to the AI Supercycle and the whole framework snaps into focus. The supercycle is fundamentally an energy-to-knowledge conversion story. Data centers are the visible tip. Frontier labs are the workshops where new letters of the alphabet are being spelled. The unit economics of the cycle are the ratio of prosperity produced to energy consumed, and every strategic question at the technology layer is downstream of that ratio.

This is why the Map of AI has Energy & Physical at Layer 1. Not because energy is a factor, but because energy is the input. Everything above it — foundries, silicon, networking, compute, models, harness, distribution — is machinery for converting energy into knowledge. Governance is the fence around the machinery. And the output the whole apparatus is trying to produce is prosperity: measurable increases in what human societies can do.

The Infinite Alphabet is the vocabulary for the middle term — the knowledge that lives between energy input and prosperity output. It is why the middle term cannot be reduced to a single number, why more of it does not automatically mean more prosperity, why some combinations of letters produce breakthroughs and other combinations produce nothing. And it is why the strategic question every AI-era firm has to answer is not “how much intelligence can we produce” but “which letters can we spell that nobody else can.”

The commoditizing letters lose value as they are copied. The appreciating letters gain value as the surrounding market becomes able to see how rare they are. The firms and geographies that read the alphabet correctly capture the next generation of prosperity. The ones that treat AI as jello miss it.

Key Takeaways & Mental Models

• The Infinite Alphabet is a change of representation. Knowledge is not a single homogeneous substance; it is a vast collection of complementary, non-substitutable letters. The economy is a Scrabble game played across geographies and centuries.

• Complementarity and non-substitutability are the two load-bearing properties. Letters need each other to form words. Letters cannot be swapped. Together these produce a combinatorial economy where value lives at the joints.

• A large fraction of the alphabet is physical. Robotics has not yet crossed into it. AI capability curves are being measured against the fraction of the alphabet the current stack can see.

• The exponential is a stack of S-curves. Hidalgo confirmed the AI-industry version on record. Pre-training, test-time compute, agentic loops, continual learning — each a bounded curve, together a cascade. Winners are on the next curve, not the current one.

• The bubble-vs-supercycle paradox is resolved through the alphabet. The bubble is the market treating non-fungible knowledge as fungible. The supercycle is the substrate compounding underneath. Both are true simultaneously.

• The three nested cycles are three rates of alphabet change. New letters on the short cycle. Industries re-sort on the medium cycle. Civilizations absorb the alphabet on the long cycle.

• The printing press is the Zero Industrial Revolution. AI 2016–2030 runs on the same clock as Reformation 1516–1530. The perceived slowness of past revolutions is a memory artifact.

• Absorptive capacity is the enterprise bottleneck. Eight months to procure a Claude Max subscription is half a paradigm cycle. Enterprises without absorptive capacity are systematically one cycle behind.

• The Map of AI is nine distinct alphabets, not nine categories of the same substance. No single player can master all nine.

• Knowledge agglomerates, it does not diffuse. Silicon Valley, Shenzhen, Zhongguancun succeed because the letters were already there. Yachay and Neom fail because they cannot be summoned by real estate.

• The forgetting rate is 3 to 6 percent per month. Knowledge is a muscle. Continuous production is the only preservation mechanism. The West’s memory-manufacturing bottleneck, nuclear-plant capability loss, and mega-project atrophy all follow the same law.

• Governance is a fence, not a letter. It is a set of rules about which combinations are legal. Any moat built inside the fence is exposed when the fence moves.

• The final synthesis is Hidalgo’s own mental model. Energy is input. Prosperity is output. Knowledge is the engine. Maximize knowledge per unit of energy, minimize emissions, and prosperity follows. The Infinite Alphabet is the vocabulary for the middle term.

• The only durable strategic question is which letters you hold that nobody else can spell. Everything scrapable, scalable, or shareable commoditizes. Only the non-fungible substrate appreciates.

With massive ♥️ Gennaro Cuofano, The Business Engineer

We start with a packed week of major AI news, but one theme stands above the rest: routing.

As we’ll see throughout this edition, routing is no longer just a technical concept. It’s becoming the economic layer that determines how intelligence is priced, distributed, and ultimately where value is captured across the AI stack.

Many of this week’s headlines may seem unrelated at first glance. In reality, they’re different expressions of the same structural shift. Let’s break them down.

For full reference, read The Map of AI Redrawn.

The Business Engineer

The Map of AI Redrawn

A decade ago, I started covering the intricacies of the AI ecosystem. As early as 2016–17, I described AI as a multi-layered stack, almost like a layered cake of interconnected technologies and infrastructure…

Read more

4 months ago · 48 likes · Gennaro Cuofano

Where We Are on the Three Axes

Before the news, the position fix. Three axes, three questions — when, where, who — and this week reads differently on each.

The time axis (the Supercycle): we are deep in Phase 1, but Phase 2 is already firing at the edge. Most of what the week surfaced is Phase-1 linear integration — generative coding, agentic assistants, cheaper software production. That’s the 5-10 year short cycle doing exactly what it’s supposed to. But two items are Phase-2 tells arriving early: agents running 24/7 on operators’ machines (a distribution-surface deployment signal, not a model signal) and Apple’s supply-side pricing defense against an agent-driven demand explosion. The clock runs differently in different regions of the Topology, and this week the edge clock is ahead of the model clock.

The binding research question on the time axis remains the unification of the four scaling laws — pre-training, post-training, reasoning, agentic tool use — into a single self-improvement loop. Keep that variable in your peripheral vision, because one of this week’s stories (the OpenAI inference-cost breakthrough) is a signal pointing straight at it.

The space axis (the Map of AI): the two new layers I added in March did the heavy lifting this week. The harness layer (routing economics) generated most of the operator-facing news, and the governance layer (the release control plane) generated the single most important event — Fable’s suspension and return. The Map had to be redrawn precisely so these two layers had somewhere to live. This week is the vindication.

The power axis (the Three-Tier Architecture): the governance perimeter reasserted itself, hard. Fable is the Cold-War governance regime made concrete — capability disclosed at controlled intervals, through a public-private release framework, because top-tier capability is now a geopolitical input. Nothing about this week loosened Tier 1. If anything, it hardened.

And the spine that holds all three axes together — the dual analogy — was visible in a single week. The microprocessor regime governed the bottom of the stack (the memory cartel, TSMC allocation, Apple’s supply crunch). The Cold War regime governed the top (Fable’s governance gate, the substrate-diversification and co-design moves). The two regimes met, as they always do, at the boundary between Compute Capacity (Layer 5) and Foundation Models (Layer 6) — which is exactly where Anthropic’s dual-substrate story and the “who controls the stack” fight are being resolved.

So the one-sentence abstraction: this week was the Topology deforming in real time, with four cascades firing at once, the governance perimeter reasserting control, and the incumbent capex map spending at containment scale — all of it consistent with a Phase-1-dominant cycle whose edge is already sliding into Phase 2.

Market Map: The Week as Coordinates and Cascades

I’ll organize this not by the nine layers in sequence, but by the four active cascades plus the governance perimeter and the incumbent capex map — because that’s where the commercial action is actually concentrating, and it’s the more complete reading.

The governance perimeter — Fable and the end of the “full release”

Coordinate: Layer 9 (Governance), all phases, meta-tier. Regime: Cold War.

Fable’s suspension and return is the governance layer behaving exactly as the Topology says it must: governance is not a roof on top of the stack, it’s a control plane perpendicular to every layer, and it changes state in days, not years. A model capable enough to surface twenty-year-dormant vulnerabilities is, by definition, a model that trips the release gate. The public instance ships restricted, guardrailed, through a public-private framework.

The durable structural fact — the one to carry forward — is that we have left the era of the “full release,” and we probably never truly had it. In Cold War terms, this is a capability gradient disclosed at controlled intervals. In operator terms, it means you can no longer assume the model you’re handed is the model the lab runs internally. Two camps formed on X (the “it’s nerfed” camp and the “it’s impressive if used correctly” camp); both are right, and the reconciliation is that governance is now a permanent, priced variable, not a regulatory footnote. Plan your security posture accordingly — because, as Joel put it, bad actors have access to the exact same tools you do.

Cascade 1 — the downward cascade from the harness (the biggest region of the week)

This is the one that ties the live to the framework. It started at Layer 7 when the agentic harness stopped being a workflow wrapper and became a routing engine, and it propagates down through Layers 6, 5, and 3. Four of the week’s stories are all the same cascade seen from different layers.

The routing paradigm. Coordinate: Layer 7, Phase 1, Tier 3. “The next Anthropic is a routing company.” Routing is where the market finally discovers what a unit of intelligence is worth — model-agnostic orchestration that abstracts the model zoo away from the enterprise user, puts a powerful model on strategy and cheaper models on execution, and reduces single-model dependency. In framework language: routing is price discovery, and it is the head of the downward cascade. Whoever owns the routing engine at enterprise scale sits at the coordinate that reprices every layer beneath it.

The Sonnet-5 cost paradox. Coordinate: Layer 6→7, Phase 1, Tier 1/3. Cheaper per token ≠ cheaper per task. Agentic models trade token efficiency for autonomy — the same job that took 5,000 tokens now takes 20,000. This is not a bug; it’s the mechanism. Agentic queries consume on the order of 500× the tokens of chat queries, and that ratio is precisely what loads the harness routing region and makes merchant custom silicon (Cerebras, Groq) economically rational one and two layers down. The cost paradox is the cascade’s energy source.

Token-maxing as a metric pathology. Coordinate: incumbent behavior at Layer 8. Meta reportedly burned billions on internal Claude usage and is now capping it. Structurally this is a governance-of-incentives failure: measure effort (tokens) instead of outcome, and consumption is what you get. As execution price falls, outcomes do not rise automatically — they have to be engineered. This is the bad-metric failure mode that bankrupts the organization that adopts it, and it’s a preview of the discipline the routing layer exists to impose.

The OpenAI inference-cost breakthrough. Coordinate: Layer 5-7 margin, Phase 1, Tier 1. Reported against a year-end gross-margin goal around 52%. Read at surface level, it’s a margin story critical to IPO viability. Read at depth, it’s a signal on the unification question: if inference at production scale gets cheap enough to run continuously, inference becomes a training-signal generator — a capex category rather than only an opex line — which is exactly the mechanism by which the four scaling laws could close into a self-improvement loop. This is the week’s highest-leverage story hiding inside a margins headline.

Cascade 2 — the dual-substrate cascade (compute-layer diversification as a valuation feature)

Coordinate: Layer 5, Phase 1, Tier 1. Regime: boundary — where microprocessor meets Cold War.

Anthropic reducing its Nvidia dependency — borrowing compute from Nvidia, Google, Amazon, and now SpaceX while developing custom silicon — is the dual-substrate cascade named in the framework almost line for line. Two independent compute substrates (Google TPU plus SpaceX capacity) is a balance-sheet feature priced directly into secondary markets; single-substrate compute with execution slippage is a liability priced into the IPO conversation. This is strategic redundancy doctrine — a Cold War institutional feature — applied to commercial substrate. Anthropic reportedly pays SpaceX north of a billion a month for compute, which is the flywheel geometry (below) monetizing capacity the incumbents haven’t yet identified as contested territory.

And the sober caveat I gave on the live is the framework’s caveat too: reducing Nvidia dependency is a decade-long story, not a 2030 story. Custom silicon is the beginning of an owned-margin narrative. Into the 2030s, no one replaces Nvidia — the revenue-vs-margin split (chip and memory layer captures the profit; labs grow revenue margin-negative) is the proof.

Cascade 3 — the Physical AI cascade (Phase 2 firing at the edge)

Coordinate: Layer 8, Phase 2, Tier 2/3.

Two of the week’s quieter items are the same cascade pulling Phase 2 forward. Agents running 24/7 re-rate the PC as a Physical AI surface — an agent-run machine is used far more intensely than a human-run one, which is a deployment-surface signal, not a model signal. And Apple’s price hikes are a supply-shock hedge against exactly this: an agent-driven demand explosion for devices colliding with a supply ceiling. The distribution surface is where Phase 2 is already firing, and the framework predicts this cascade shortens Phase 2’s timeline as adjacent endpoint categories get pulled along the same curve.

Which folds into the embedded frontier region (Layer 8, Phase 1, Tier 1): Apple currently ships Gemini-embedded AI on-device. That’s the Apple-Google embedding — a frontier model entering a billion-device distribution surface — and it’s the canonical example of incumbent capex as containment: paying to foreclose the surface before a native AI player reaches it.

The microprocessor floor — the memory cartel and TSMC allocation

Coordinate: Layer 2 (Foundries/Packaging), Phase 1, Tier 1/2. Regime: microprocessor.

The alleged price-fixing among Samsung, SK Hynix, and Micron is the HBM chokepoint doing what chokepoints do. Read structurally, not moralistically: memory has swung boom-and-bust for thirty years; if you sit on the chokepoint and can’t fix demand certainty, you won’t wildly expand supply. A cartel is demand-certainty by other means. The framework’s rule — binding constraints migrate downward when a lower layer becomes scarce — is why this is the constraint of the moment: the HBM oligopoly is forward-purchased through the late 2020s, and everything above it is gated by it.

And Apple is no longer TSMC’s top customer — Nvidia is. Allocation power moved up the stack, above chip design itself. A frontier chip without foundry allocation is a slide deck; a device maker without foundry priority faces a supply bottleneck driven by scarcity, not weak demand. That single reordering — Nvidia ahead of Apple at TSMC — explains Apple’s pricing behavior more completely than any consumer-demand story.

The incumbent capex map — containment, co-design, and the geometry race

This is the deepest read, and it’s where the “two camps” conversation from the live resolves into the framework.

The co-design war Joel and I described — the Nvidia-Palantir camp wanting to commoditize the model layer versus the OpenAI-Anthropic camp wanting to embed into the enterprise stack — is the diagonal of dominance being contested through geometries. Commoditizing the model layer means pushing value below the Layer 5/6 boundary into the microprocessor region the incumbents control (that’s horizontal geometry — Nvidia selling shovels to the containment effort, collecting rent on every capex dollar that crosses Layers 3-5). Embedding into the enterprise means the labs climbing up into Layers 7-8 fast enough to become indispensable before they’re commoditized.

And the enterprise’s own logic — the one I laid out on the live — is the defect strategy in disguise: keep your data in a controllable ontology/memory layer, not fine-tuned into a frontier model’s weights, so you retain control, security, and no lock-in, and can swap the model with a dropdown. That’s an enterprise deliberately positioning off the diagonal to stay resilient to single-layer disruption.

The capex behind all of it is containment, not expansion. Microsoft Frontier’s reported ~$2.5B bet on the last mile, its army of forward-deployed engineers, Meta’s spend and its token caps, Apple’s embedding — none of it is forward-projected Phase-1 optimism. It’s backward-induced from Phase-2 survival anxiety. The incumbents have internalized the second half of the incumbent paradox — that native AI players take the dominant positions in Phase 2 — and are spending at containment scale to foreclose Phase-2 territory before it can be occupied. Microsoft Frontier’s FDE model is the last-mile bottleneck capitalized: enterprise AI doesn’t scale on model quality, it scales on hybrid technical-commercial experts embedded inside the organization.

Meta’s Allbirds-vs-SpaceX fork is the geometry question stated cleanly. SpaceX runs a flywheel geometry — engines at Layers 1, 5, and 8 compounding through loops, monetizing a compute overbuild (the $1B/month to Anthropic) from outside the contested terrain. Meta has the compute overbuild but not the loops. So the fork is real: convert capacity into a genuine flywheel (SpaceX) or strand it as a pivot narrative (Allbirds). The tell will be whether Meta builds compounding loops or just resells idle capacity.

Read the Week by Coordinates

For the individual operator. Match model power to task, because you’re really choosing a position in the harness routing region. Powerful, agentic models as orchestrators; cheaper models for execution. Don’t dump everything on Sonnet 5; don’t write your posts through Fable. You don’t send the CEO for coffee.

For teams. Update the context and memory on a cadence — a model advising on two-year-old context is an advisor who hasn’t spoken to you in two years. The human-in-the-loop is the alignment mechanism that keeps the architecture tracking a moving business.

For harness / loop builders. Add loops into an architecture you’ve already tested with yourself inside it; run them on a cadence; stay present to re-tune as efficacy decays. You are building a position at the head of the downward cascade — treat it as infrastructure, not a script.

For anyone designing incentives. Never measure by tokens consumed. As execution price falls, outcomes must be engineered. Token-maxing is the metric pathology that ended a company or two this quarter.

For enterprises choosing a foundation. Optimize for control, security, and no lock-in. Keep data in a controllable ontology/memory layer, not in frontier weights. Favor substrates that let you swap the model — that optionality is a deliberate off-diagonal position, and it’s worth more than any single model’s edge today.

The reading discipline, applied. For every company in the week, specify the coordinate, name the geometry, read the adjacent cascades, and ask whether the geometry compounds faster than the coordinates it occupies. Anthropic at Layers 5-7, Tier 1, gated by whether dual-substrate diversification stays priced as an asset. Nvidia across Layers 3-5, gated by hyperscaler custom silicon and merchant flankers. SpaceX at Layers 1/5/8, gated by Starship. Meta at Layer 8, gated by whether three billion users keep engaging with its surfaces in current form. Refuse the singular. Specify the coordinate.

What’s Next

• The next Anthropic is a routing company. In framework terms: whoever owns the head of the downward cascade — the enterprise routing engine — reprices every layer beneath it and becomes the next trillion-dollar company inside three years. I’ll keep saying it so no one can claim we didn’t.

• The unification question is the variable to watch. The OpenAI inference breakthrough matters most as a signal toward closing inference-as-training-signal into a self-improvement loop. Whichever frontier lab closes that pipeline first compounds at a rate no competitor can match, and the model race resumes in a more brutal form.

• Incumbent capex will not normalize. Microsoft, Meta, and Apple are spending at containment scale because the logic is foreclosure, not integration. Anyone modeling capex normalization in 2027-2028 is reading Phase-1 logic into a Phase-2 contest.

• Memory is the binding constraint, and it’s below the model layer. Watch the HBM chokepoint and TSMC allocation, not benchmark leaderboards, if you want to know what’s actually possible next year. The bottleneck migrates downward; the memory cartel is where the cycle is physically gated.

• Governance is permanent perimeter. No more full releases. Every top-tier capability ships through a public-private framework that can change state in days.

• Four events would force a redraw: a frontier lab credibly closing the four-scaling-law unification loop; a sovereign compute contract crossing $50B and formalizing the third buyer pole; a China-side Tier-1 breakthrough in silicon or foundry that collapses Strategic Denial; and a Starship outcome that hardens or breaks the flywheel geometry. Three of the four sit below or beside the model layer — which is the whole point.

Key Takeaways & Mental Models

• Read by coordinates, not headlines. Every story is a triple of (layer, phase, tier). Ten headlines this week were largely four cascades plus a governance-perimeter event. Mechanism: the coordinate determines compounding rate, risk profile, and access regime; the label hides all three.

• Routing is the downward cascade. “The next Anthropic is a routing company” and “the harness layer became a routing engine” are the same claim from two ends. Mechanism: a Layer-7 deformation propagates down through 6, 5, and 3, repricing each — and routing is where the market discovers what intelligence is worth.

• The cost paradox is the cascade’s fuel. Cheaper per token ≠ cheaper per task; agentic queries run ~500× the tokens of chat. Mechanism: the token ratio loads the routing region and makes merchant silicon rational two layers down.

• The dual analogy located itself this week. Microprocessor regime at the memory/foundry floor; Cold War regime at the governance perimeter; the two collide at the Layer 5/6 boundary where Anthropic’s dual substrate sits. Mechanism: the collision is why the Three-Tier Architecture exists at all.

• Capex is containment, not expansion. Microsoft Frontier, Meta’s spend, Apple’s embedding are backward-induced from Phase-2 survival anxiety. Mechanism: incumbents foreclose Phase-2 territory at every layer before native players can occupy it.

• The co-design war is the geometry race. Commoditize-the-model = horizontal geometry pushing value below the 5/6 line; embed-in-enterprise = climbing into 7-8; keep-data-in-ontology = an off-diagonal defect play for resilience. Mechanism: each is a different bet about which axis compounds fastest.

• Governance is perimeter, not roof. Fable is a capability gradient disclosed at a controlled interval. Mechanism: Layer 9 surrounds the stack as a control plane and changes state in days, not years.

• Bottlenecks migrate downward. The memory cartel and TSMC reallocation (Nvidia ahead of Apple) are the binding constraints of the moment. Mechanism: when a lower layer goes scarce, the constraint — and the pricing power — moves down to it.

• Inference-as-training-signal is the unification tell. The margin breakthrough is really a signal on the highest-leverage variable in the cycle. Mechanism: continuous cheap inference turns production deployment into a training-signal generator, the seam where the four scaling laws could close into one loop.

• Allbirds vs. SpaceX is the flywheel test. A compute overbuild is either a compounding flywheel (loops at multiple layers) or a stranded pivot. Mechanism: geometry, not capacity, decides which.

The Bottom Line — Where We Are

We are deep in Phase 1, with the edge already sliding into Phase 2. The model race did not resume this week; it stayed paused at the unification question, while the geometry race — structured entirely around incumbent containment — ran in full view. The most important event was not a model at all; it was a governance-perimeter action (Fable) and a margin signal pointing at self-improvement (OpenAI inference).

Underneath, four cascades were firing at once — the downward cascade from the harness, the dual-substrate cascade at compute, the Physical AI cascade at the edge, and the sovereign-compute thirst feeding SpaceX — which is itself the signal the framework predicts: cascades accelerate as the Topology matures. A single week now carries what used to take a quarter.

So where are we? At the coordinate where the microprocessor floor (memory, foundries) is the binding physical constraint; where the Cold War perimeter (governance) is reasserting control of release; where the incumbents are spending at containment scale to foreclose Phase 2; and where the single highest-leverage commercial position — the enterprise routing engine at the head of the downward cascade — is still unclaimed.

The Topology will look different again in a few months. The discipline is the same as it was on the live: read by coordinates, not by labels — and keep asking whether the geometry compounds faster than the ground it stands on.

Stay close.

With massive ♥️ Gennaro Cuofano, The Business Engineer

Apple’s China story is not just about supply chains. It is about how technology, manufacturing depth, geopolitical risk, and market access became fused into a single strategic system. That same system-level tension will define the next phase of AI infrastructure, where control over production capacity may matter as much as control over software, chips, or models.

Patrick McGee’s “Apple in China” isn’t just a business book—it’s a chronicle of how the world’s most valuable company became inextricably tied to America’s chief geopolitical rival, transforming both in the process.

The story begins not with strategy or vision, but with embarrassment. Patrick McGee, then the Apple reporter for the Financial Times, was interviewing the CEO of Qualcomm—a $200 billion chip designer that powers virtually every 3G, 4G, and 5G device on the planet. The CEO mentioned that during COVID, he had quarantined for 14 days just to visit factories in Taiwan. McGee was puzzled. Qualcomm is a fabless company—they don’t manufacture chips themselves. Why the visit?

“We follow the Apple model,” the CEO explained.

McGee realized he had no idea what that meant. He was covering Apple, and he didn’t understand the fundamental architecture of how the company actually operated. That moment of professional shame became the genesis of one of the most important business books of the past two decades—a work that traces not just Apple’s history, but the evolution of US-China relations, global supply chains, and the hollowing out of American manufacturing capacity.

The Foundation: Survival Through Outsourcing

The conventional narrative around Apple treats the company as a design genius that happened to manufacture in China. The reality is far more pragmatic and far more desperate.

Apple wasn’t choosing between manufacturing philosophies in the 1990s. It was choosing between outsourcing and bankruptcy. The company came perilously close to collapse in 1996, again in 1997, and once more in 2000. Outsourcing and offshoring weren’t strategic luxuries—they were existential necessities.

McGee traces Apple’s journey through several distinct phases:

• Late 1990s: Initial moves toward outsourcing as a survival strategy

• Early 2000s: Discovery and deepening engagement with China

• 2004-2007: The iPod Mini and iPod Nano establish unprecedented manufacturing power

• 2007-2015: iPhone production explodes from 5 million units to 230 million units annually

These weren’t just product launches. They were the creation of an entirely new model for how complex electronics could be manufactured at scale—a model that would reshape global industry.

What Made China Different

The question isn’t why Apple went to China. The question is why no one else could compete with what China offered.

The scale and speed of China’s manufacturing capability wasn’t just impressive—it was categorically different from anything available in the developed world. McGee emphasizes that this wasn’t about cheap labor or regulatory arbitrage, though those factors mattered. It was about something deeper: China’s ability to mobilize resources, coordinate complex supply chains, and execute at a pace that made Western manufacturing look practically medieval.

Consider what Apple required: factories built in months where Western countries would still be doing environmental paperwork, access to millions of skilled laborers, and suppliers willing to retool their entire operations on Apple’s demanding timeline. China didn’t just meet these requirements—it exceeded them, becoming better at high-end electronics manufacturing than anyone else on the planet.

The relationship worked because both parties brought something essential to the table. Apple brought world-class engineering, exacting standards, and an almost obsessive attention to detail. China brought execution capacity, supply chain coordination, and an industrial policy that prioritized manufacturing excellence.

As McGee puts it: “Apple was the teacher.” But China wasn’t just a passive student. They absorbed knowledge like sponges and then kicked ass. It’s like saying someone went to Harvard—that’s not a criticism, it’s recognition that they learned from the best and now know how to do things better than their teachers.

The Tim Cook Factor

Before Tim Cook became CEO, he was the architect of Apple’s supply chain dominance. His role in this story cannot be overstated.

Cook didn’t just optimize Apple’s supply chain—he fundamentally reimagined what a supply chain could be. His innovations weren’t about cutting costs or finding cheaper suppliers. They were about achieving a level of control and coordination that no other company had demonstrated.

The “Apple model” that the Qualcomm CEO referenced essentially means: take real control over what’s happening in manufacturing, impose exacting specifications, maintain constant oversight, and never accept that “good enough” is actually good enough. Cook institutionalized this approach, creating systems where Apple engineers worked side-by-side with suppliers, where Apple’s standards became the industry’s standards, and where the line between Apple and its suppliers blurred in important ways.

This level of integration had a profound effect: it made Apple simultaneously powerful and vulnerable. Powerful because no competitor could replicate the quality, speed, and scale of Apple’s manufacturing. Vulnerable because Apple became deeply, perhaps irreversibly, dependent on a manufacturing ecosystem centered in China.

The Xi Jinping Inflection Point

The period from 2007 to 2015 saw two parallel transformations: Apple’s explosive growth coincided almost exactly with Xi Jinping’s rise to power and his turn toward a more assertive, nationalist China.

By the time Western observers recognized that China was changing, Apple was already too entangled to easily extricate itself. The company had built not just factories, but entire ecosystems in China. They had negotiated retail locations on premium streets in Shanghai—something that requires significant political connections in China’s system. They had become embedded in China’s economy in ways that went far beyond manufacturing contracts.

McGee’s research reveals a question that few were asking: who actually had those political connections? How did Apple navigate China’s opaque political system? The answers involve a level of government relations and political engagement that most Western observers missed while focusing on product launches and stock prices.

The relationship also had profound asymmetries. Beijing understood Apple’s dependence better than Apple did. Chinese officials knew that Apple had no realistic alternative to China’s manufacturing capacity, at least not without years of investment and development. This knowledge became leverage in ways both subtle and explicit.

The Geopolitical Paradox

The most unsettling dimension of the Apple-China relationship isn’t what might happen—it’s what is happening.

The real risk isn’t that China will kick Apple out. The real risk is that the relationship will continue exactly as it is. McGee argues that the most dangerous scenario isn’t disruption, but continuity—a continued transfer of knowledge, capability, and manufacturing expertise from American engineers to America’s chief geopolitical rival.

Consider the implications: the United States is training its primary adversary in the most consequential technologies for the 21st century—high-end electronics, chip fabrication, advanced manufacturing. This isn’t happening through espionage or theft (though those exist). It’s happening through the ordinary operations of one of America’s most successful companies.

The question isn’t whether this makes business sense for Apple—clearly it does. The question is whether what makes sense for Apple makes sense for America. These interests aren’t necessarily aligned, and McGee’s book forces readers to confront that uncomfortable reality.

The Hollowing Out

If war ever comes over Taiwan, McGee argues, “nobody wins that war.” But beyond the immediate catastrophe of military conflict lies a deeper concern: China’s industrial capacity is now second to none, far larger than America’s.

The GDP figures can be misleading. As Xi Jinping himself has noted, much of China’s GDP consists of “steel and furnaces”—real industrial capacity. By contrast, a significant portion of Western GDP comes from what Xi dismissively calls “the fictitious economy of social media and crypto.”

This isn’t just about manufacturing iPhones. It’s about the fundamental industrial base that any modern economy—and especially any military power—requires. The West’s industrial capacity has been systematically hollowed out, and China has systematically built up its own. Apple didn’t cause this trend alone, but it exemplified and accelerated it.

The Counterfactual Question

Could Apple have done things differently? McGee is blunt: “There would be no Apple” if the company had tried to maintain domestic manufacturing. The company was days away from bankruptcy in 1996, close again in 1997, and again in 2000. Without offshoring and outsourcing, Apple wouldn’t have survived to make the iPhone.

This creates an uncomfortable paradox for critics: the very success story they want to claim for America was only possible because of deep integration with China. The iPhone exists because of China’s manufacturing capability. The iPad, the MacBook, the entire Apple ecosystem—none of it would exist without the model McGee documents.

Even people who worked directly with Steve Jobs for more than a decade acknowledge this reality. When asked about the Xi Jinping era, they respond: “What else did you expect us to do?” Nobody else was offering the incentive packages, the millions of laborers, or the ability to build factories in months rather than years.

The lazy interpretation of McGee’s book is that Apple should have invested in America. The honest interpretation is that if Apple had done so, we wouldn’t have an Apple to argue about.

The India Question

As tensions with China have increased, India has emerged as a popular alternative in policy discussions. The reality is more complicated.

India wasn’t going to perform at the level China performed at—not in the 2000s, and arguably not now. This isn’t about criticizing India’s potential. It’s about recognizing that China spent decades building the infrastructure, training the workforce, and developing the supplier ecosystems necessary to manufacture complex electronics at scale.

You can’t simply transplant this capability to another country, even one with significant advantages like India’s demographics and democratic system. The manufacturing ecosystem is as much about institutional knowledge, supplier relationships, and coordination capacity as it is about labor costs or physical infrastructure.

The Asymmetric Leverage Problem

One of the most striking insights from McGee’s analysis: Apple’s leverage in China is far weaker than most observers assume. Beijing knows that Apple is stuck. They understand that there’s no realistic alternative manufacturing location that can deliver what China delivers.

This knowledge creates a profound asymmetry in bargaining power. Apple can threaten to leave, but the threat isn’t credible—at least not on any timeline that matters for current business decisions. China can impose requirements, create obstacles, or simply leverage Apple’s presence for broader geopolitical goals, knowing that Apple’s options are limited.

The situation might theoretically improve under some framework of US-China trade agreements, with Apple as a key element making things work. But McGee is skeptical: what exactly is Apple’s leverage in negotiating such arrangements when Beijing knows the company has nowhere else to go?

The Sourcing Achievement

If there’s one thing McGee is proudest of, it’s the sourcing of “Apple in China.” Apple is arguably more secretive than the US military, certainly more secretive than most government agencies. Getting people to speak on the record, or even on background, about Apple’s operations is notoriously difficult.

The book succeeds because McGee managed to build trust with dozens of sources across Apple’s ecosystem—engineers, executives, suppliers, and government officials. The result is a level of detail and insight that simply doesn’t exist elsewhere in the literature on Apple or on global supply chains.

This sourcing wasn’t just about access. It was about understanding: McGee had to learn not just what happened, but why it happened, and what the alternatives actually were. Many business books fail because they impose retrospective logic on historical decisions. McGee’s book works because it reconstructs the actual constraints and options that decision-makers faced.

The Stakes Today

We’re living through the consequences of the system McGee documents. The COVID-19 pandemic exposed supply chain vulnerabilities. The war in Ukraine highlighted energy dependencies. The competition over semiconductor manufacturing revealed technological vulnerabilities. And at the center of all these concerns sits the question of whether Western economies can maintain technological and manufacturing leadership when so much capability has been offshored.

Apple’s story is America’s story—a story of innovation and growth, but also of dependency and vulnerability. The question isn’t whether we can go back to some imagined past of domestic manufacturing. That past, if it ever existed, is gone. The question is what comes next, and whether we can build systems that balance economic efficiency with strategic resilience.

McGee’s book doesn’t answer that question. But it does something more valuable: it provides the detailed understanding necessary to even ask the question properly. And in an era of simplistic takes and political posturing about China, manufacturing, and American competitiveness, that clarity is invaluable.

For anyone trying to understand how we got here—how the world’s most valuable company and the world’s most populous authoritarian state became partners in the most important manufacturing relationship of the modern era—”Apple in China” is essential reading. Not because it offers solutions, but because it documents with unusual precision and sourcing the system we actually have, rather than the one we imagine or wish we had.

Recap: In This Episode!

The Transformation: From Near-Bankruptcy to Global Dominance Through Offshoring

Apple’s engagement with China wasn’t strategic brilliance; it was survival. The company faced existential collapse multiple times in the 1990s and early 2000s. Outsourcing became the only viable option.

McGee maps four manufacturing phases:

1. Late 1990s – Outsourcing as triage.

2. Early 2000s – Deep integration with China.

3. 2004–2007 – iPod manufacturing establishes China’s supremacy.

4. 2007–2015 – iPhone era: scale jumps from 5M to 230M units a year.

This was not mere contract manufacturing. It was the birth of a new global manufacturing model: Apple-driven, China-executed, and world-changing.

Mechanisms: Why China Became Irreplaceable

McGee shows China outperformed not because of cheap labor but because of state-enabled coordination capacity. China could mobilize workers, suppliers, land, and capital faster than any Western system.

Key mechanisms:

• Factories built in months, not years.

• Millions of skilled laborers deployable on demand.

• Suppliers retooled and synchronized at Apple’s cadence.

• Industrial policy aligned to win high-end electronics.

China didn’t just execute Apple’s playbook; it mastered and extended it. Apple transferred knowledge; China built institutions to retain and scale it.

The Tim Cook Architecture: Precision, Control, Vulnerability

As COO, Cook built the “Apple model”:

• Radical integration with suppliers.

• Nonstop engineering oversight inside factories.

• Zero tolerance for variance.

• Supply chain as a strategic weapon, not a cost center.

This created unmatched product quality, speed, and scale. But the same system locked Apple into China’s ecosystem. The more Apple optimized, the more path dependent it became.

Xi Jinping’s Rise: When a Manufacturing Dependency Became a Political Dependency

Apple’s explosive growth overlapped almost perfectly with Xi’s ascent. What began as an economic integration became a political entanglement.

McGee examines:

• Apple’s need for political access to secure stores and operations.

• The opaque intermediaries who brokered these relationships.

• Beijing’s sophisticated understanding of Apple’s dependence.

Apple needed China to scale; China used Apple to accelerate its own industrial ascent.

The Geopolitical Paradox: Continuity Is More Dangerous Than Disruption

McGee argues the core risk isn’t that China ejects Apple. It’s that the status quo continues.

Because the status quo means:

• Continuous transfer of manufacturing know-how.

• Strengthening of China’s industrial base.

• Continued erosion of Western capacity.

• Deepening asymmetry of leverage.

Apple’s success becomes America’s structural weakness.

The Hollowing Out of the West: Industrial Power vs the “Fictitious Economy”

A core thesis: China built real industrial capacity, while the West drifted toward digital ephemera. Xi’s critique of Western GDP as inflated by “social media and crypto” frames the contrast.

Apple exemplified the shift:

• High-value Western design.

• High-value Eastern manufacturing.

• A widening gap between innovation and industrial capability.

This divergence now shapes everything from semiconductor policy to defense readiness.

Counterfactuals: Could Apple Have Done Anything Else?

McGee’s answer is blunt: no.

Domestic manufacturing would have killed Apple before the iPhone ever existed. No country offered what China offered in the 2000s. Even today, no alternative matches China’s depth.

This destroys the simplistic narrative that Apple “should have built in America.” The company had no viable domestic option. China made Apple possible.

The India Question: Attractive in Theory, Unrealistic in Practice

India is not the next China. Not because of a lack of talent or ambition, but because:

• China built decades of infrastructure and supplier ecosystems.

• Manufacturing scale depends on institutional knowledge and coordination, not just labor pools.

• India cannot replicate China’s 2000s model on short timelines or with similar state capacity.

India may grow, but China’s system is not transplantable.

Asymmetric Leverage: Beijing Holds the Stronger Hand

Apple cannot credibly threaten to leave. Beijing knows it. Supply chain reallocation would take a decade and cost hundreds of billions.

This asymmetry shapes:

• Regulatory concessions.

• Political compromises.

• Apple’s muted stance on sensitive issues.

• China’s confidence in using Apple as leverage in broader US–China dynamics.

Sourcing as an Achievement: Cracking Apple’s Wall of Secrecy

McGee’s unique contribution is access. Apple’s internal operations are more secretive than most governments. The book succeeds because he secured rare, deep, cross-ecosystem sourcing from engineers, suppliers, and political actors. The result is a level of operational insight that doesn’t exist elsewhere.

Implications: The System We Built, Not the System We Wished For

The book forces recognition: the West offshored not just manufacturing but industrial capability, and Apple was both beneficiary and accelerant.

Today’s geopolitical landscape—semiconductor competition, supply chain fragility, technological rivalry—emerges from this architecture.

McGee doesn’t provide solutions, but he gives the analytical foundation needed to even formulate them. “Apple in China” is essential because it documents the world as it is: interdependent, asymmetric, and strategically unstable.

With massive ♥️ Gennaro Cuofano, The Business Engineer

The race for AI leadership is not determined solely by algorithms, models, or computing power. Beneath the technological stack lies a physical layer of resources, manufacturing capacity, and geopolitical control. Rare earths and critical minerals represent a strategic bottleneck that will increasingly influence how AI infrastructure is built, scaled, and distributed across the global economy.

There’s a fascinating tension at the heart of AI infrastructure that most technical discussions completely miss: the more we digitize, the more we depend on what’s underground. Every optimization algorithm, every autonomous system, every data center promising to replace human labor rests on a foundation of physical extraction that looks nothing like the sleek interfaces we interact with.

The conversation with Ernest Scheyder, author of The War Below, reveals a structural reality that sits uncomfortably with our narratives about technological progress. While AI companies compete on model performance and infrastructure scale, they’re all drawing from the same constrained pool of critical minerals—copper, lithium, rare earth elements—that require digging giant holes in the ground.

This isn’t just a supply chain problem. It’s a fundamental constraint on the speed of transformation.

The Atrophy of Strategic Capability

Here’s what makes this particularly instructive from a business engineering perspective: the United States was a global leader in critical minerals after World War II. Then it systematically let that capability atrophy.

Universities stopped teaching mining engineering. Government funding dried up. The industry was seen as “dirty” and unfashionable compared to clean tech narratives around biofuels and renewable energy. Mining became something you outsourced to countries with fewer environmental concerns and cheaper labor.

This is a pattern we see repeatedly in industrial strategy: short-term optimization around cost and aesthetics leading to long-term strategic vulnerability. The U.S. wasn’t thinking in terms of supply chain resilience or geopolitical leverage. It was thinking in terms of quarterly efficiency and social acceptability.

China, meanwhile, took the opposite path. As Scheider notes, Deng Xiaoping recognized in the 1970s that controlling rare earth processing would give China enormous leverage. He was right. Today, China dominates the midstream processing of these minerals—not necessarily because they have all the deposits, but because they built the infrastructure and expertise while the West walked away from it.

The AI Infrastructure Bottleneck

The explosion in AI data centers creates a compound problem. You need:

• Copper for power generation, transmission, and server infrastructure (the U.S. imports ~50% of its needs)

• Lithium for stationary storage batteries (data centers need massive backup power)

• Rare earth elements for everything from hard drives to cooling systems

• Yttrium for electronics and defense applications (emerging bottleneck)

Each of these requires mining, processing, and refinement. The timeline from discovering a viable deposit to producing commercial-scale output is 10-15 years minimum. Environmental reviews in the U.S. take 7-10 years before you even break ground.

This creates a fundamental mismatch: AI infrastructure is scaling on software timelines (measured in months and quarters), but the physical materials needed scale on geological timelines (measured in decades).

The Geopolitical Leverage Point

China’s recent move to restrict rare earth exports isn’t just trade negotiation theater. It’s a demonstration of structural power. When you control the processing infrastructure for materials that every advanced economy depends on, you have leverage that operates independently of conventional market dynamics.

This is the kind of constraint-based power that’s easy to miss if you’re only looking at surface-level competition. It’s not about who has the best AI model or the most efficient data center design. It’s about who controls the physical bottlenecks in the supply chain.

The interesting strategic question: can you build AI infrastructure fast enough to develop the automation that would accelerate mineral extraction and processing, before the mineral bottleneck constrains your ability to scale AI infrastructure?

The Circular Dependencies

Here’s where it gets genuinely interesting. The mining industry itself is deploying AI for:

• Autonomous vehicles (mines are perfect testing grounds—no pedestrians, repetitive routes)

• Exploration optimization (AI can process geological data to identify where to dig next)

• Operational efficiency (reducing power, time, and cost of extraction)

So we have a feedback loop: AI infrastructure needs more minerals, which requires more mining, which can be optimized with AI, which requires more infrastructure, which requires more minerals.

But there’s a constraint on this loop. Building the robots and autonomous systems that would accelerate mining requires the same scarce minerals. You can’t just software-optimize your way out of physical scarcity.

The Recycling Counternarrative

There’s a counterforce that rarely gets attention in the hype cycle: recycling. Unlike oil and gas, which you burn once and it’s gone forever, critical minerals can be reused indefinitely. Lithium doesn’t lose its ability to hold an electrical charge just because it’s been in a battery for 20 years.

This suggests a potential path out of the extraction paradox: aggressive recycling infrastructure for end-of-life electronics, batteries, and data center equipment. The minerals already sitting in phones, laptops, and server farms could become the “urban mining” deposits of the next decade.

But building recycling infrastructure at scale requires the same things as building mining infrastructure: capital, technical expertise, regulatory approval, and time. The West is behind on both fronts.

The Perception Gap in Labor Markets

Mining has an image problem, which creates a talent problem. It’s seen as “tech-light,” as a industry of giant dump trucks and dynamite. This couldn’t be further from reality. Modern mining operations use:

• Large language models for geological analysis

• Heavy data pipelines for real-time optimization

• Autonomous systems for extraction and transport

• Advanced metallurgy and processing chemistry

The industry is desperate for technical talent but struggles to attract it because the perception lags the reality by decades. Meanwhile, AI companies can hire top engineering talent at premium salaries by selling a narrative of working on “the future.”

This perception gap has real consequences. If you can’t attract young people into mining engineering programs, you can’t build the technical base needed to scale production. The Colorado School of Mines and similar institutions are critical bottlenecks, but they’re not remotely as fashionable as computer science programs.

The Choice That Isn’t Optional

Scheider frames this as a question of choice: if we want an electrified future powered by AI and renewable energy, we have to choose to accept the mining infrastructure needed to support it. We can make that mining cleaner, more automated, and more responsible—but we can’t eliminate it.

The West, particularly the U.S., has been trying to avoid this choice by outsourcing extraction to other countries while maintaining the benefits of the technology. That strategy is running into hard limits as geopolitical competition intensifies and supply chains become weapons.

The business engineering insight here: narratives about “clean” technology often obscure the dirty, physical infrastructure required to produce it. Solar panels, wind turbines, electric vehicles, and AI data centers all depend on mining. The question isn’t whether we mine, but where, how, and under what conditions.

Strategic Implications for AI Companies

For companies building in the AI infrastructure space, this creates several strategic considerations:

Vertical integration vs. market dependence: Do you try to secure your own mineral supply chains, or do you accept market pricing and availability risk? Microsoft, Google, and Amazon are large enough to consider direct investment in mining operations or long-term supply agreements. Smaller players have to accept whatever the market provides.

Geographic diversification: Relying on any single country for critical inputs is a strategic vulnerability. But geographic diversification in mining is extremely difficult—you can’t just spin up a new copper mine in 18 months.

Recycling infrastructure investment: For AI companies with massive existing hardware deployments, investing in recycling infrastructure could provide a secondary supply stream and hedge against primary extraction bottlenecks.

Technological acceleration of extraction: Funding AI applications in mining could help accelerate the supply of materials needed for AI infrastructure. But this is a long-cycle investment with geological timelines.

The Deeper Pattern

What makes this analysis interesting for business engineering is that it reveals a category of constraint that operates differently from what most tech companies are used to. You can’t:

• Software-optimize your way out of geological scarcity

• Move fast and break things when environmental reviews take 7-10 years

• Iterate rapidly when mine development timelines are measured in decades

• Pivot when your data centers need specific minerals that only exist in certain locations

This is a return to industrial-era constraints applied to digital-era infrastructure. The companies that recognize this earliest and adjust their strategies accordingly will have significant advantages over those that continue operating on pure software-company assumptions.

The Unseen Foundation

The core insight from Scheider’s work is simple but profound: the digital economy rests on a physical foundation that most people in that economy never think about. Every time you query ChatGPT, you’re depending on copper mined in Chile, lithium extracted in Australia or China, and rare earth elements processed in facilities that look nothing like the tech campuses that house the engineers writing the code.

This isn’t a criticism of AI or technology. It’s an observation about the nature of constraints and where they actually exist. Understanding where the real bottlenecks are located—not in the obvious places everyone is watching, but in the unsexy infrastructure that makes everything else possible—is essential for sound strategic thinking.

The companies, countries, and investors who recognize that AI’s future runs through giant holes in the ground will make very different decisions than those who think it’s purely about who has the best model or the most efficient algorithms. Both matter. But when physical constraints bind, they bind absolutely.

Recap: In This Issue!

Structural Reality: AI Depends on Physical Extraction

• All AI infrastructure — data centers, automation, GPUs — rests on minerals like copper, lithium, and rare earths.

• Digital scale moves at software speed, but mineral supply moves at geological speed.

• The physical foundation of AI is invisible in most tech narratives.

The U.S. Strategic Atrophy

• Post-WWII US leadership in mining collapsed due to social stigma, regulatory drag, and shifting academic focus.

• Mining engineering programs and government support eroded for decades.

• Result: the US now imports half its copper and relies on foreign processing for critical minerals.

China’s Midstream Leverage

• Deng Xiaoping’s 1970s bet on rare earth processing created a long-term geopolitical advantage.

• China dominates processing, not because of deposits, but because they built infrastructure and competence while the West withdrew.

• Export restrictions are a demonstration of structural power, not policy theater.

AI Infrastructure Bottlenecks

• Copper: power transmission, servers, cooling.

• Lithium: grid-scale and backup storage.

• Rare earths: motors, drives, cooling systems.

• Yttrium: specialized electronics and defense.

• Mine development timeline is 10–15 years; environmental approval alone can take a decade.

The Feedback Loop: Minerals → AI → Mining → Minerals

• Mining increasingly depends on AI (autonomous vehicles, geological analysis, optimization).

• AI expansion depends on minerals.

• Automation requires hardware built from those same minerals.

• This creates a circular dependency that cannot be “software-optimized” away.

Recycling as a Counterforce

• Critical minerals don’t degrade; they can be recycled indefinitely.

• “Urban mining” (electronics, batteries, data center waste) could become a major future source.

• Western recycling infrastructure is underdeveloped and requires the same capital, talent, and permitting challenges as mining.

Talent and Perception Constraints

• Modern mining is a high-tech domain (LLMs, robotics, real-time telemetry).

• Industry perception is decades out of date, creating a severe talent shortage.

• Universities producing mining engineers are strategic bottlenecks — and unfashionable compared to CS programs.

The Non-Optional Choice

• Electrification and AI require vast mineral extraction.

• The West cannot “clean tech” its way out of mining — it can only choose where and how mining happens.

• Outsourcing extraction has turned into a vulnerability as geopolitics weaponize supply chains.

Strategic Implications for AI Companies

• Vertical integration: hyperscalers may need direct mineral investments or long-term supply deals.

• Geographic diversification: extremely hard because you can’t “spin up” a new mine quickly.

• Recycling investment: a credible hedge for companies with massive hardware footprints.

• Mining-tech acceleration: AI firms may need to invest in mining automation to unblock their own long-term infrastructure expansion.

The Deeper Pattern: Industrial Constraints Return

• You can’t move fast in a domain where permitting takes 10 years.

• You can’t pivot when your hardware requires minerals concentrated in specific regions.

• AI growth is hitting industrial-era constraints that software-era companies are not designed to manage.

• The real bottlenecks sit in unglamorous midstream infrastructure few tech leaders understand.

With massive ♥️ Gennaro Cuofano, The Business Engineer

Then, a few years ago, in early 2023, three frameworks began taking shape in parallel to make sense of what the AI cycle actually was. The frameworks were developed separately because each was answering a different question, and the questions did not initially appear related.

The AI Supercycle asked when: when does a foundational technology like AI become commercially generative, through what phases, and over what time horizon. The answer was a 30-50 year paradigm shift unfolding through three nested cycles, with linear integration into existing industries first, the unlocking of stalled technologies second, and the creation of industries we cannot yet name third.

The Map of AI asked where: where in the architectural stack does value accumulate, where do binding constraints live, where does rent get captured. The answer evolved from a three-layer reading into the current nine-layer structure, with each layer constrained by the layer below it and commercialized by the layer above it.

The Three-Tier Architecture of Stratified Interdependence asked who: who is allowed to operate at any given layer, under what political and commercial terms, with what restraint embedded in the access regime. The answer integrated geopolitical players (states, alliances, multilateral control regimes), macroeconomic players (capital markets pricing the tier system into valuations and procurement), and technology players (companies designing their architectures around tier boundaries).

Three frameworks. Three questions. Three years of refinement. The unification is now overdue.

The unified object is best named for what it actually is: a topology. A structure on a set of coordinates that defines which positions are close to which others, how the structure transforms under continuous deformation, and which properties persist through transformation. The diagonal of dominance is a topological feature. The nine-layer architecture is a topological invariant. The tier system is not — it deforms in quarters, not decades. The defects are coordinates where the structure has localized anomalies that produce novel commercial properties. Cascades are deformation events propagating through the structure.

The Topology is rugged. It is reversible by design. And the most strategically interesting action lives where the three axes pull against each other.

Share The AI Supercycle

On the Spine — The Dual Analogy

Before the three axes, the spine.

For nearly a decade, the AI cycle has been described by analogy to the microprocessor wave that began in the 1940s and matured into personal computing, the internet, mobile, cloud, and smartphones across subsequent decades. The analogy is correct, and it has done enormous work in explaining the cycle’s economic structure. But it is incomplete. There is a second analogy running in parallel, and the failure to name it explicitly is why so much commentary on the AI cycle misreads the governance regime as a regulatory footnote rather than the structural feature it actually is.

The microprocessor analogy works for the physical supply chain of AI. It covers Layers 1 through 5 of the Map: energy, foundries, silicon, networking, compute. The lessons of the microprocessor cycle apply directly to this region of the stack. Cascading bottlenecks migrate downward as upper layers scale. Geographic concentration of production creates allocation power. Incumbents capture early value because integration is linear and capital advantages dominate. Profit pools migrate up the stack as commercialization at higher layers creates demand. Everything we know about the semiconductor cycle from 1947 through 2010 applies, with translation, to the physical substrate of AI.

The Cold War analogy works for the frontier of AI. It covers Layers 6 through 9: foundation models, agentic harness, distribution surfaces, and the governance perimeter that wraps around them. A note on the analogy before going further: the Cold War is being used here exclusively as a reference for the governance and institutional architecture that emerged between rival blocs during a period of foundational dual-use technology development. The analogy says nothing about whether frontier AI poses risks comparable to nuclear weapons. That is a substantive question the framework deliberately does not take a position on. What the framework observes is that states are behaving as if they are governing frontier AI under a Cold-War-style control architecture, and the observed behavior is the structural fact the Topology has to account for.

The Cold War regime had specific institutional features that map directly onto the AI governance regime. Containment doctrine. Multilateral export controls (COCOM). Alliance-based technology bifurcation (the Iron Curtain as a technology boundary, not only a political one). Non-aligned blocs courted by both sides through infrastructure deals. Strategic redundancy doctrines (second-source supply chains, dual-track command). Capability gradients tracked and disclosed at controlled intervals. Détente as the institutional mechanism for managed commerce between rival blocs at non-strategic layers. Proxy theaters where rivalry expressed itself without direct confrontation.

Every one of these has a direct analog in the AI cycle. The Three-Tier Architecture is the containment architecture. Strategic Denial is COCOM. Sovereign procurement is non-alignment becoming a procurement category — the UAE, Saudi Arabia, India, and Brazil are courted by both the US-led and China-led stacks with infrastructure deals at sovereign scale. The dual-substrate Anthropic story is strategic redundancy doctrine. The Apple-Google distribution embedding is alliance hardening at the consumer surface. The frontier benchmark race is the capability gradient that intelligence services and procurement teams track.

The two analogies meet at the boundary between Layer 5 (Compute Capacity) and Layer 6 (Foundation Models). Below that line, microprocessor logic. Above that line, Cold War logic. The Three-Tier Architecture exists because the two regimes collide at this boundary — and the H200 deal, with its 25% government surcharge, is the prototype of how that collision gets commercially resolved. A microprocessor-era artifact subjected to a Cold-War-era control structure. The first time in industrial history that a chip has been sold under terms that resemble managed-trade détente more than the export of consumer electronics.

The unified statement: the AI cycle is running on a microprocessor substrate under a Cold War governance regime. That is the spine of the framework. The duality has a location in the stack. The duality has commercial consequences. And the duality is the structural reason the Topology requires three axes rather than the single axis that either analogy alone would imply.

The Full Framework, Stated Once

Before the axis-by-axis walkthrough, the framework stated as a single integrated object.

The AI Supercycle Topology is a three-axis coordinate system that locates any company, technology, or commercial decision in the AI cycle, with five operating components that interact continuously.

The time axis is the AI Supercycle. Three nested phases (5-10 year linear integration, 10-20 year stalled-technology unlock, 20-50 year industry creation) running against four scaling laws operating in parallel (pre-training, post-training, reasoning, agentic tool use), with the binding research question of the next 12-18 months being whether the four scaling laws can be unified into a single self-improvement and continual-learning pipeline. The axis tells you when value emerges.

The space axis is the nine-layer Map of AI. From physical floor to perimeter: Energy, Foundries, Silicon, Networking, Compute Capacity, Foundation Models, Agentic Harness, Distribution Surfaces, Governance. Each layer is constrained by the layer below it and commercialized by the layer above it. The axis tells you where value accumulates.

The power axis is the Three-Tier Architecture of Stratified Interdependence. Strategic Denial at the frontier, Managed Commerce in the negotiated middle, Open Competition at the friction-laden base. The tier system integrates geopolitical players (states, alliances, control regimes), macroeconomic players (capital markets pricing the tier system), and technology players (companies designing architectures around tier boundaries). The axis tells you who is allowed to operate, under what conditions.

The framework runs on a dual analogy spine. The microprocessor analogy explains the physical supply chain of AI (Layers 1-5): cascading bottlenecks, allocation power, geographic concentration, incumbent paradox. The Cold War analogy explains the frontier governance regime (Layers 6-9): containment, multilateral export controls, strategic redundancy, non-aligned procurement, capability gradients. The two regimes collide at the Layer 5/6 boundary. The collision is the structural reason the Three-Tier Architecture exists at all.

The framework’s most consequential observation is the capex map. Incumbent capex commitments at unprecedented scale (Meta $125-145B, Microsoft $80B+, Google’s doubled run-rate, Apple’s $1B/year embedding) are backward-induced from Phase 2 survival anxiety, not forward-projected from Phase 1 expansion. The incumbents have internalized the second half of the paradox — that native AI players take dominant positions in Phase 2 — and are spending at containment scales to foreclose Phase 2 territory before it can be occupied. The capex behavior is the visible proof that the cycle is being contested as a Phase 2 survival problem, not a Phase 1 integration opportunity.

The unit of analysis is the coordinate: a triple of (layer, phase, tier) that locates a company or technology in the Topology. Coordinates are dynamic — they move quarter to quarter as cascades deform the structure.

The mechanism of change is the cascade: a deformation event that propagates through adjacent coordinates, reshaping which regions are loaded. Cascades propagate faster through layers (technical, demand-driven, months) than through tiers (political, negotiation-driven, quarters to years). Four are currently running: the downward cascade from the harness layer, the dual-substrate cascade at the compute layer, the Physical AI cascade at the distribution layer, the sovereign procurement cascade at the compute layer.

The strategic strategies are the geometries: four distinct traversal patterns through the Topology, each a direct response to the incumbent containment environment. Vertical (Google) — column ownership across nine layers within one tier; incumbent containment perfected. Horizontal (NVIDIA) — band ownership across Layers 3-5 across two tiers; selling shovels to the containment effort. Flywheel (SpaceX) — three engines compounding through five loops at Layers 1, 5, 8; operating outside contested territory. Niche by Adjacent Integration (Cerebras, Groq) — single-coordinate optimization at defects the containment cannot reach.

The reading discipline that uses all of the above is: refuse to talk about AI in the singular. Specify the coordinate. Identify the geometry. Read the cascades adjacent to the coordinate. Ask whether the company’s geometry compounds faster or slower than the coordinates it occupies.

That is the framework. Everything that follows is the detail.

The Three Axes in Full

Axis 1: The Time Axis — The Supercycle

The Supercycle thesis describes AI not as a new industry but as a paradigm shift comparable to the microprocessor wave in its substrate-creating role for subsequent waves of industry formation. The thesis has three core components.

The three nested cycles. Within the 30-50 year supercycle run three sub-cycles of different durations. A short cycle of 5-10 years where AI integrates into existing industries through linear improvements — search becomes generative, advertising becomes hyper-personalized, software becomes 10x cheaper to produce. A medium cycle of 10-20 years where AI unlocks technologies that had stalled for non-AI reasons — robotics, autonomous vehicles, AR, ambient inference at the edge. A long cycle of 20-50 years where AI enables industries that cannot be specified in advance — the way streaming could not be specified before the internet matured.

The four scaling laws operating in parallel — and the unification question. As of mid-2026, four scaling laws are running simultaneously. Pre-training scaling (more parameters, more data, more compute). Post-training scaling (RLHF, constitutional approaches, agentic fine-tuning). Reasoning at inference time (test-time compute, chain-of-thought, deliberative search). Agentic tool use (the harness layer as capability multiplier). Each has different hardware implications, different cost structures, and different commercial profiles. The shape of the time axis is no longer a single curve. It is a four-curve composite.

But the binding research question of the next 12-18 months is not which curve dominates. It is whether the four can be unified into a single pipeline that closes into a self-improvement and continual-learning loop. If unification succeeds, the four scaling laws stop being parallel curves and become one self-reinforcing engine. Pre-training feeds capability into the base model. Post-training feeds task-specific competence into the deployed model. Reasoning feeds deliberate problem-solving into agent execution. Agentic tool use feeds environmental feedback back into the training loop. The model executes agentic tasks in production, generates feedback signal, the feedback gets distilled into post-training updates, the post-training updates feed pre-training continuation, the better pre-trained model reasons better at inference, and the better reasoning generates better agentic execution.

The strategic significance of unification is enormous. If unification succeeds at any single frontier lab before others, that lab compounds at a rate no competitor can match — because every other lab is still running disconnected pipelines. The model race that supposedly ended would resume in a more brutal form, with the winning lab pulling away through compounded self-improvement. The Phase 1/Phase 2 boundary would also become more porous: self-improving models running agentic tasks generate training signal from physical-world deployment, collapsing the distinction between “models in Phase 1” and “Physical AI deployment in Phase 2” into the same engine. And the capex map would change materially, because continuous inference at production scale becomes a training-signal generator, making inference a capex category rather than only an opex one.

The unification question is currently the binding strategic variable at the model layer. Every other variable in the framework — frontier-lab valuations, compute procurement at scale, the dual-substrate doctrine, the Apple-Google distribution embedding — is currently being structured around bets on whether unification works and which lab gets there first.

The incumbent paradox. In Phase 1, incumbents capture most of the value because integration is linear and their distribution and capital advantages dominate. Google captures generative search. Meta captures generative advertising. Microsoft captures enterprise integration. In Phase 2 and Phase 3, native AI players take dominant positions because the underlying business logic changes — and the change is non-linear enough that incumbent advantages do not translate. The paradox sits at the time axis but produces the most consequential behavior at the capital axis, which Part II of this piece treats as a standalone analytical section.

The time axis tells you which phase a layer is in, with the discipline that not all layers sit at the same phase simultaneously. Physical AI deployment surfaces are firing in Phase 2 right now. Reasoning models are deep in Phase 1. Sovereign procurement is creating Phase 1 demand against a Phase 2 deployment timeline. The clock runs differently in different regions of the Topology.

Axis 2: The Space Axis — The Map of AI

The Map of AI describes the architectural geography of the cycle. Four structural shifts forced the current nine-layer structure.

Networking separated from compute. Rack-scale interconnects now exceed 20% of system cost at scale. The unit of AI infrastructure is no longer the GPU. It is the rack. And the rack is bound together by protocols (NVLink, Spectrum-X, InfiniBand) that capture rent at every interface they touch.

Foundries separated from silicon. TSMC’s CoWoS-L allocation became the most contested resource in semiconductors. The HBM oligopoly has been forward-purchased through the late 2020s. Allocation power moved up the stack to become more consequential than chip design itself. A frontier chip without foundry allocation is a slide deck.

Silicon bifurcated internally into three categories. Generalist substrate (NVIDIA at 80%+ share). Hyperscaler custom (TPU, Trainium, MTIA, Maia, Titan). Merchant custom challengers (Cerebras, Groq, SambaNova, Tenstorrent) selling to anyone on physics dimensions NVIDIA does not optimize for. This is a within-layer split but it changes the strategic logic at every layer above silicon.

Governance separated from generic policy and trust and became a priced variable. Single government directives can switch off frontier models globally in days. Governance is no longer a footnote. It is a control plane perpendicular to every layer beneath it.

The current nine layers, from physical floor to perimeter:

LayerNameBinding constraintAnalogy regime9GovernancePolitical stability of the tier systemCold War8Distribution SurfacesEndpoint surface areaCold War7Agentic HarnessRouting economics, cost-per-tokenCold War6Foundation ModelsCapability per parameterCold War5Compute CapacityGW of operational power, rack countBoundary4Networking and ProtocolsInterconnect bandwidth, rack-scale bindingMicroprocessor3SiliconArchitecture, transistor densityMicroprocessor2Foundries and PackagingLithography, advanced packaging allocationMicroprocessor1Energy and PhysicalGW of generation, rare earths, waterMicroprocessor

Two reading rules govern the space axis. Each layer is constrained by the layer below it — you cannot ship a model without compute, you cannot ship compute without silicon, you cannot ship silicon without foundries, you cannot ship foundries without energy. Binding constraints migrate downward when capacity at a lower layer becomes scarce. And each layer is commercialized by the layer above it — silicon needs models to monetize, models need a harness, harnesses need distribution. Profit pools migrate upward when commercialization at a higher layer creates demand.

The space axis tells you where value accumulates, where rent gets captured, and which analogy regime governs the strategic logic at any given layer.

Axis 3: The Power Axis — The Three-Tier Architecture

The Three-Tier Architecture is the power axis. It is not a geopolitical framework alone, a macroeconomic framework alone, or a technology framework alone. It is the integrated regime under which all three operate, because the tier system simultaneously specifies geopolitical access conditions, prices into macroeconomic variables, and forces technological architecture decisions.

Tier 1 — Strategic Denial. Hard wall. The answer is structurally no, and the answer holds across administrations because the underlying logic is alliance capitalism: capability denial as strategic imperative.

• Geopolitical layer: Frontier GPUs (Blackwell, Rubin). EUV lithography. Sub-3nm fabrication. Frontier model weights. Quantum computing primitives. The walled perimeter of the allied bloc.

• Macroeconomic layer: The wall is priced into NVIDIA’s forward guidance as a permanent base case (zero China data center compute revenue). It is priced into secondary-market valuations of frontier labs (Anthropic’s dual-substrate compute strategy carries a premium that single-substrate execution does not). It is priced into procurement contracts that include compliance clauses with extraterritorial reach.

• Technological layer: Forces dual-substrate architectures at frontier labs (Anthropic on Google TPU plus SpaceX COLOSSUS). Forces silicon co-design on the non-allied side of the line (DeepSeek V4 on Huawei Ascend). Forces the inference layer to route on cost-per-token rather than peak capability when the peak is walled off.

Tier 2 — Managed Commerce. Permissioned flows with leverage embedded. Washington controls the tap, monetizes the flow, makes the arrangement politically sustainable.

• Geopolitical layer: Previous-generation semiconductors (the H200 deal). Agricultural commodities. Aircraft. Energy. Green technology. The negotiated middle ground.

• Macroeconomic layer: The 25% government surcharge on H200 sales is a revenue-sharing mechanism, not a tax. It transforms a layer into an export industry with embedded political conditions. The surcharge model is the prototype that will extend to other previous-generation strategic technologies as Tier 1 walls move forward.

• Technological layer: Creates a category of products designed for Tier 2 — chips deliberately specified at one generation behind the frontier, models deliberately released as open-weight to flow through the managed perimeter, infrastructure deliberately scoped to operate under the surcharge regime.

Tier 3 — Open Competition. Non-strategic goods at permanently elevated but negotiated friction. The new structural cost of doing business, not a temporary disruption.

• Geopolitical layer: Consumer goods, industrial inputs, general manufacturing. Section 122 (15%) plus Section 301 (7.5-100%) plus Section 232 (sector-specific).

• Macroeconomic layer: The friction is priced into supply chain decisions, into reshoring incentives, into the equity valuations of companies positioned for Tier 3 commerce.

• Technological layer: Most consumer-facing AI products operate in this tier — the harness layer routing is largely Tier 3, the merchant custom silicon market is largely Tier 3, open-weight model deployment outside the frontier is largely Tier 3.

Two properties of the architecture matter for the Topology.

The tiers are not universal. They are specific to the layer and the asset class. The same two countries in strategic confrontation over semiconductors are simultaneously cooperating on agricultural commodities and co-managing energy stability. The rules of access differ by what is being accessed.

The architecture is reversible by design. It is built on rolling truces, not treaties. A Supreme Court ruling reconfigured the US tariff architecture in 72 hours in February 2026. A China-side breakthrough at CXMT or SMIC could collapse Strategic Denial in semiconductors and reorganize the power axis below the foundation model layer. A single sovereign procurement crossing $50 billion formalizes a new sub-tier inside Compute Capacity that does not currently exist as a stable category.

The power axis tells you who is allowed to operate at a given layer, under what political conditions, with what restraint embedded — and how the access regime simultaneously prices into capital markets and forces architectural decisions on the companies operating across it.

Part II — The Incumbent Paradox and the Capex Map

The most analytically important behavior in the current cycle is not the model race, the silicon race, or even the geopolitical contest. It is the scale of incumbent capex spending. And the spending makes no sense under a forward-looking Phase 1 logic. It only makes sense under a backward-induced Phase 2 survival logic. Reading the capex map correctly is the key to reading the entire cycle.

The Numbers That Demand Explanation

The capex commitments of the largest incumbents are at scales that have no precedent in software industry history.

Meta is guiding to $125-145 billion in capex for 2026, with free cash flow projected down nearly 90% on the year, and has executed approximately 8,000 layoffs to redirect headcount into compute build-out. Microsoft committed $80 billion+ across the Fairwater facility and its successors, then absorbed displaced Stargate capacity (the 700 MW that Norway and the UK declined to host) into its own infrastructure plan. Google’s capex run-rate has doubled in 24 months, and the Apple distribution deal at roughly $1 billion per year is itself a capex substitute — paying to rent distribution capability rather than build it. Amazon’s AWS infrastructure spend is at multi-year peaks. NVIDIA is selling into all of this and recycling the cash into networking acquisitions, foundry pre-payments, and the sovereign procurement pipeline.

These are not the spending patterns of companies that believe they will simply integrate AI into their existing dominance and continue compounding linearly. A linear-integration Phase 1 thesis would justify capex at perhaps half these levels. The remainder is something else.

The Second Half of the Incumbent Paradox

The incumbent paradox has two halves. The Phase 1 half says that in early integration, incumbents capture most of the value because integration is linear and their distribution, capital, and operational advantages dominate. That half is well understood and uncontroversial.

The Phase 2 half says that as the underlying business logic changes — as native AI players reach scale, as the substrate becomes the product, as distribution surfaces themselves get disintermediated by AI-native interfaces — incumbents lose their structural advantages, and the dominant players of the new phase are often not the same companies as the dominant players of the previous phase. The Phase 2 half is less discussed but more consequential.

The capex commitments tell us that the incumbents have internalized the second half of the paradox. They are not spending at these scales because they believe Phase 1 integration will be valuable — though it will be. They are spending at these scales because they believe Phase 2 will be existentially dangerous to their current positions if they do not lock in territory before native AI players reach it first.

The aggressive lock-in is rational under this reading. Meta’s capex bet is that its 3-billion-user distribution moat plus its own MTIA silicon plus its own Muse Spark model can outrun the disintermediation. Microsoft’s bet is that Copilot embedding plus Azure neutrality on model choice locks in enterprise procurement before native AI players reach enterprise scale. Google’s bet is the only one that fully bridges the two phases — the vertical column from silicon through distribution gives it the chance to remain dominant across the Phase 1/Phase 2 transition rather than only on one side of it.

Capex as Containment

In Cold War terms, the incumbent capex bet is containment. The incumbents are not racing to win Phase 1. They are racing to make Phase 2 too capital-intensive for any native AI player to challenge their position. The strategy is to build moats deep enough to survive the disintermediation phase by foreclosing the territory before the rival reaches it.

This explains several otherwise puzzling features of the current cycle. It explains the willingness to crater near-term free cash flow — Meta’s free cash flow projected down nearly 90% on the year is incomprehensible under a Phase 1 thesis but perfectly rational under a Phase 2 containment thesis. It explains the urgency of the embedding deals — the Apple-Google arrangement is paying $1B/year for distribution embedding because the alternative is letting a native AI player reach a billion devices first. It explains the willingness to absorb displaced infrastructure — Microsoft picking up the 700 MW that Stargate could not place is rational because compute capacity is itself the contested territory.

It also explains the asymmetric incumbent behavior. Apple’s containment strategy is to rent capability (the Google deal) while keeping option value on building in-house, because Apple’s distribution moat is durable enough to outlast a two-year window of rented capability. Meta’s containment strategy is to spend more aggressively than anyone because Meta’s distribution moat (social and advertising surfaces) is more disintermediation-vulnerable than Apple’s (hardware and OS). Google’s containment strategy is the most coherent because Google is the only incumbent with a full vertical column that can be defended at every layer simultaneously.

The Mechanism: Containment Forecloses Optionality

The structural mechanism is foreclosure of optionality. By spending at unprecedented scale at every layer of the substrate (compute, models, silicon, distribution embedding), the incumbents are foreclosing the optionality that a native AI player would otherwise have to enter at any single layer. The native AI player cannot challenge Google by building only a model, because Google’s column denies the silicon access, the compute access, the distribution access. The native AI player cannot challenge Meta by building only a distribution surface, because Meta’s capex denies the compute substrate and the silicon underneath.

The containment is not absolute. Two categories of native AI players have already broken through despite the containment: the frontier labs (OpenAI, Anthropic, xAI) by going directly to the model layer with sovereign-scale capital and dual-substrate compute arrangements, and the merchant custom silicon players (Cerebras, Groq) by occupying defect coordinates where the incumbent containment did not reach. Both categories survive because they did not try to enter on the incumbents’ chosen terrain.

But containment is succeeding broadly. The total number of native AI players currently positioned to threaten incumbent dominance in Phase 2 is small — perhaps 6-10 companies globally that have credible Phase 2 positions, and most of them are either embedded inside incumbent partnerships (OpenAI/Microsoft, Anthropic/Google + SpaceX) or operating in cells of escape (DeepSeek, Cerebras) rather than directly confronting incumbent territory.

What the Capex Map Predicts

If the containment thesis is correct, three predictions follow.

First, incumbent capex will not normalize as Phase 1 matures. It will continue at elevated levels through Phase 2 because the strategic logic is foreclosure, not expansion. Investors expecting capex normalization in 2027-2028 are reading Phase 1 logic into a Phase 2 contest.

Second, the dominant native AI players of Phase 2 will be the ones that occupied territory the incumbents could not reach in time. The frontier labs, the merchant custom silicon players, and the open-weight + silicon co-design coordinate are the working examples. New native players entering after 2026 will need to find new defects, because the incumbents have closed the obvious entry points.

Third, the geometries from Part IV of this piece — vertical, horizontal, flywheel, niche — are not just strategic descriptions. They are direct responses to the containment environment. Vertical geometry is incumbent containment perfected. Horizontal geometry is selling shovels to the containment effort. Flywheel geometry is the only geometry that does not play the incumbent containment game (because it operates outside the incumbents’ chosen terrain). Niche geometry is occupying defects the containment cannot reach.

The capex map is not a side variable. It is the most visible direct evidence that the incumbent paradox is fully operational, that the second half is what’s actually driving behavior right now, and that the entire cycle is being structured around the question of which incumbents survive the Phase 2 transition and which native AI players break through.

Coordinates, Cascades, and Geometries