Six months ago, I started a spreadsheet. It was supposed to be simple — a quick mapping exercise to understand where AI money actually flows. Who builds, who sells, who profits, who pays. I figured it would take a few weeks. I was wrong. The spreadsheet became a database. The database became an obsession. And the obsession became something I wasn’t expecting: a detailed blueprint of the most sophisticated wealth extraction architecture I’ve ever encountered.
I want to be precise about what I mean by “class architecture.” I’m not using the term loosely. I mean a structured system in which different tiers of economic actors are positioned — by design, not by accident — to either extract value or have value extracted from them. And what I found in the AI economy is not a bug. It’s not an unintended consequence. It’s the product itself.
The starting point: follow the invoices, not the hype
My approach was deliberately boring. I didn’t start with Nvidia’s stock price or OpenAI’s valuation. I started with invoices. Actual spending. I wanted to know: when a mid-sized European company “adopts AI,” where does the money go? When a government funds an “AI initiative,” who cashes the check? When a startup raises a seed round to build an “AI-powered” product, what do they spend it on?
The answers, once you aggregate them, reveal a stunningly clear hierarchy. I ended up organizing it into five layers, and the more I refined them, the more they started to resemble something older and more familiar than Silicon Valley — they looked like feudalism.
Layer one: the infrastructure lords. Nvidia, TSMC, the hyperscale cloud providers — AWS, Azure, Google Cloud. These are the landlords of the AI economy. Every single AI application, from a chatbot to an autonomous vehicle, pays rent to this layer. Nvidia’s data center revenue hit $35 billion in a single quarter in late 2024. That’s not product revenue. That’s toll revenue. You cannot participate in AI without paying Nvidia. The cloud market — projected to push past $800 billion by 2026 — functions the same way. This layer doesn’t need to win any particular AI bet. They profit from all of them.
Layer two: the model aristocracy. OpenAI, Anthropic, Google DeepMind, Meta’s AI research division. These entities control the foundational models — the core intellectual property upon which nearly everything downstream depends. They sit in an extraordinary position: they’ve absorbed billions in compute costs (paying up to layer one) and in return, they’ve built moats so deep that competition is structurally discouraged. Training a frontier model now costs hundreds of millions of dollars. This layer doesn’t just sell access to intelligence. It sells dependency.
Layer three: the integration guild. This is where a huge number of European and American tech companies live — the consultancies, the SaaS platforms, the middleware providers who take foundational models and make them “enterprise-ready.” Accenture, Deloitte’s AI practice, a thousand startups with “AI-powered” in their pitch decks. This layer is profitable, but its margins are fundamentally constrained. They’re buying from layer two and selling to layer four. They’re middlemen. Sophisticated middlemen, but middlemen nonetheless.
Layer four: the adopters. Every company, government, and institution that “uses AI.” This is where the real extraction happens, and I want to linger here, because it’s where the architecture gets elegant. The adopters are simultaneously the customers and the product. They pay for AI tools. They feed those tools proprietary data — which improves the models — and then they pay again for the improved version. The market size projections for enterprise AI adoption by 2026 are staggering — hundreds of billions — and nearly all of that money flows upward through the stack.
Layer five: the displaced. The workers whose labor AI is designed to replace or diminish. Customer service agents, junior copywriters, data analysts, translation professionals. This layer doesn’t appear on any balance sheet. They’re not a market segment. They’re a cost reduction.
The elegance is in the dependency loops
What struck me — what kept me at this spreadsheet long past the point of sanity — is how self-reinforcing the whole system is. Each layer is structurally locked into its position. And the mechanisms that lock them in are not coercive in any obvious way. They feel like choices.
Consider a mid-sized logistics company in the Netherlands. They decide to adopt AI for route optimization. They evaluate vendors — layer three companies that offer AI-powered logistics platforms. They choose one. That platform runs on GPT-4 or Claude (layer two), which runs on Azure or AWS (layer one), which runs on Nvidia chips (layer one, sub-basement). The logistics company gets a genuine efficiency gain — maybe 12% reduction in fuel costs. Real value. But the majority of the economic value they’ve created flows upward. Their subscription fee goes to the vendor. The vendor’s API costs go to OpenAI or Anthropic. Those API costs are largely compute costs that go to Microsoft or Amazon. And those compute costs are largely GPU costs that go to Nvidia.
The logistics company has optimized. It has improved. And it has become a tributary feeding an upward river of capital that terminates in about seven companies, most of them headquartered within a thirty-mile radius of each other in Northern California.
I’m not writing from a position of purity here. Brown Brothers Media uses AI tools. I use Claude for research assistance. I pay the toll. The architecture is compelling precisely because opting out isn’t a viable strategy — it’s a competitive death sentence. That’s the sign of a truly elegant extraction system: participation is both rational and extractive simultaneously.
The data layer nobody talks about
There’s a shadow flow I haven’t mentioned yet, and it might be the most important one. Every interaction with an AI system generates data. Usage patterns, prompt structures, domain-specific corrections, preference signals — this data is extraordinarily valuable because it’s how models improve. And it flows, with almost no friction or compensation, from layer four upward to layers two and one.
When your company’s employees use an AI coding assistant and correct its outputs, they’re performing unpaid labor that improves the model. When a hospital adopts an AI diagnostic tool and its doctors flag errors, those corrections become training signal. The language we use — “feedback,” “improvement,” “optimization” — obscures what’s actually happening: a massive, distributed, uncompensated transfer of expertise from domain professionals to model owners.
This is not new, of course. Scholars like Astra Taylor have been writing about “fauxtomation” — the hiding of human labor behind the mask of automation — for years. But the AI economy has industrialized this process at a scale that makes previous versions look artisanal.
What makes the AI data extraction different from, say, Facebook’s advertising model is the directness of the value loop. With social media, your data was used to target ads — a step removed from the core product. With AI, your data literally becomes the product. Every enterprise customer is simultaneously training its own future competition, because the improvements they fund through usage get rolled into models that serve everyone — including their competitors.
The European position — and why geography is class
I’m based in Singapore, but I’ve spent significant time covering the European tech space, and one of the most sobering realizations from this mapping exercise is how clearly it reveals Europe’s structural position in the AI class hierarchy. Europe is, overwhelmingly, a layer three and layer four participant. It integrates and adopts. It does not — with very few exceptions — own the infrastructure or the foundational models.
This has massive implications that go beyond tech industry navel-gazing. When European governments talk about “AI sovereignty,” they’re implicitly acknowledging their position in this architecture. But most policy responses — regulatory frameworks, innovation funds, digital market acts — are operating at layers three and four, trying to improve the terms of participation rather than changing the structural position itself. It’s like negotiating better rent terms while someone else owns every building in the city.
The market projections tell the story clearly. European enterprise AI spending is projected to grow aggressively through 2026, but the vast majority of that spending will exit the European economy entirely, flowing to American cloud providers and model companies. Europe is funding American AI dominance with its adoption budgets. This is not a conspiracy — it’s the natural behavior of the architecture.
And this geographic dimension maps onto the class architecture in ways that should alarm European policymakers far more than it currently does. The infrastructure layer is American (and partially Taiwanese, for semiconductors). The model layer is American (and partially Chinese, for domestic models). The integration layer is globally distributed but margin-constrained. And the adoption layer — the layer that pays — is global, with Europe as one of the largest and most eager customers.
What I got wrong at first
Early in this project, I assumed the architecture was primarily about corporate profit. Big companies extracting from small companies. But the longer I spent with the data, the more I realized the extraction operates across multiple dimensions simultaneously — corporate, geographic, and labor — and that these dimensions reinforce each other.
The labor dimension is particularly important and under-discussed. When AI replaces a junior analyst’s work, the productivity gain accrues to the employer (layer four), who uses it to justify their AI subscription (flowing to layer three), which funds API access (flowing to layer two), which funds compute (flowing to layer one). The junior analyst, meanwhile, faces a choice between reskilling — often into a role that serves the AI architecture in a different capacity — or accepting diminished economic power. The system doesn’t create unemployment as cleanly as previous automation waves. It creates underemployment and wage compression. It’s harder to see, harder to organize against, and more effective as an extraction mechanism.
I also got the timeline wrong. I initially thought this architecture would take a decade to mature. It’s already mature. The market size figures I’ve been tracking across infrastructure, cloud services, and enterprise AI adoption — they’re not projections of something that might happen. They’re measurements of something that’s already happened. The class architecture isn’t forming. It’s formed. We’re living inside it.
So what do you do with this map?
I don’t have a neat prescription. I’m suspicious of anyone who does. But I think the mapping exercise itself has value, because the architecture’s greatest strength is its invisibility. It presents itself as a technology story — breakthroughs, innovations, exciting new capabilities — rather than what it actually is: a political economy story about who captures value and who doesn’t.
A few things I’ve concluded, tentatively:
First, the most important competition in AI is not between companies. It’s between layers. The infrastructure lords are trying to expand into models (Google, Microsoft). The model aristocrats are trying to build their own chips and infrastructure (OpenAI, reportedly). The integration guild is trying to become indispensable enough to capture more margin. These inter-layer competitions may, over time, create openings that disrupt the current architecture. Or they may simply consolidate it further.
Second, the data extraction problem is solvable — in theory. Data cooperatives, usage-based compensation models, regulatory frameworks that treat training signal as a form of labor — these ideas exist. They’re not technically impossible. They’re politically difficult because they would redistribute value away from layers one and two, and those layers have extraordinary political power.
Third — and this is the one I keep coming back to — awareness of the architecture changes your relationship to it. Not dramatically. I’m still going to use AI tools. You’re still going to adopt AI if you run a business. But understanding that your adoption is simultaneously value creation and value extraction — understanding that you’re both the customer and the raw material — gives you at least the possibility of making different choices. Demanding different terms. Building, where possible, alternatives that distribute value differently.
I started a spreadsheet six months ago. What I found isn’t a technology. It’s a class system — elegant, self-reinforcing, and almost invisible. The least we can do is make it visible.
Feature image by cottonbro studio on Pexels
