The Seven-Layer Concentration

The competition you see at the top is real. It is balanced on a stack where the layers below are not competitive — a cascading natural monopoly.

The mental picture most people have of the AI industry is a competition. OpenAI versus Anthropic versus Google versus Meta versus a dozen well-funded startups, all racing to build the smartest model. Customers pick a winner. The market sorts it out. Standard tech industry stuff.

This picture is wrong in a way that takes about ten minutes to see, and once you see it, you cannot unsee it.

What’s actually happening is concentration. Not at the level of “which company makes the best chatbot,” but at the level of the entire stack of physical and economic infrastructure that makes any chatbot possible. The stack has roughly seven layers, and at almost every layer, between one and three companies hold an effective monopoly.

Layer one: rare earths and raw silicon. The materials that go into modern chips come from a small number of mines, mostly in countries that don’t always like the United States. China currently controls a majority of the global rare earth refining capacity. This is not because the elements are rare in the earth’s crust — they aren’t. It is because the refining is dirty, expensive, and politically inconvenient, and one country decided thirty years ago to do it.

Layer two: chip fabrication. There are exactly three companies on Earth that can make the most advanced chips. TSMC in Taiwan, Samsung in South Korea, and Intel in the United States. Of these, TSMC has roughly 90% of the leading-edge market. A single fab, in a single country, ninety miles from a hostile neighbor, makes the chips that the entire AI industry runs on. There is no Plan B. Building a fab takes about five years and twenty billion dollars.

Layer three: chip design. NVIDIA, currently, has somewhere around 80% of the market for AI training chips, and an even higher share of the inference market. Their nearest competitor is AMD, which has roughly 10%. Everyone else, combined, has the remainder. This is not a market. This is a near-monopoly with a tail.

Layer four: networking and storage. The chips have to talk to each other. The data has to live somewhere. The companies that make the high-speed switches and the high-density storage that AI requires — Broadcom, Marvell, a handful of others — are themselves a small club, and several of them depend on the same fabs as NVIDIA.

Layer five: cloud and data center. Three companies — Amazon, Microsoft, and Google — operate the vast majority of the cloud capacity in the world. They are also, increasingly, the ones building the new mega-clusters required for frontier AI training. Each of them has a different “preferred” AI lab. Microsoft’s is OpenAI. Amazon’s is Anthropic. Google has its own.

Layer six: foundation models. Maybe four to six labs in the world, today, are doing serious frontier-scale training. OpenAI, Anthropic, Google DeepMind, Meta, and a handful of others. Below this layer, there are hundreds of smaller players, but they are running on models trained, distilled from, or descended from the work of those few labs.

Layer seven: applications. This is the only layer that looks like a competitive market. Thousands of startups, tens of thousands of products, all built on top of the six layers below — most of which they don’t control, can’t replace, and don’t fully understand.

The competition you see at the top is real. But it is balanced on a stack where the layers below are not competitive. They are not even close to competitive. They are exactly the kind of stack that economists, in any other industry, would describe as a cascading natural monopoly.

This has consequences.

When OpenAI signs a deal with Microsoft for compute, it is not negotiating in a free market. There are three options for that scale of compute, all of them are taken, and the pricing reflects this. When Anthropic signs a deal with Google for TPUs, same situation. When NVIDIA decides what to charge for its chips, it is not constrained by competition — it is constrained only by what its customers can afford to pay before they go bankrupt.

The concentration also means that policy decisions in a small number of places can change the entire industry. A new export rule from the US Commerce Department, an earthquake near Hsinchu, a single CEO decision at TSMC, a regulatory shift in Beijing — any of these can move the global AI industry by 20% in a week. This has happened multiple times in the past three years. It will keep happening.

The application layer — the only layer where it looks like normal competition — is the cheapest part of the stack to enter, and therefore the most crowded. Most of the breathless reporting on “AI startups” is about this layer. Most of the actual money and power is in the layers below, where there are no startups, no breathless reporting, and very little public visibility.

If you want to understand where the AI industry is going, do not watch the application layer. Watch the fabs. Watch the substations. Watch the rare earth mines. Watch the three or four executive teams whose decisions ripple all the way up.

The future is being made in places that don’t have demo days.

Part of the Aperture series at Nous. The structural reading here pairs with The Physical Ceiling (which constraints close in first) and The 30-Year Lag (the timeline within which this stack will keep mattering).