Six months ago the conventional view of the generative-AI sector held that a long tail of specialist model providers would coexist with three or four general-purpose frontier labs. That view has, very quietly, stopped being conventional.

What is forming instead looks more like the consolidation phase of an extractive industry: a small number of vertically integrated operators who own the chips, the data-center capacity, the leading frontier model, and — increasingly — the application layer that sits on top of it.

The four stack layers

Silicon. Compute. Frontier model. Application. In every previous platform cycle — mainframes, PCs, the web, mobile — different companies won different layers, and the interfaces between them created durable rents. In this cycle, the same three or four names are appearing in every layer at once.

Nvidia sits at the base of nearly every training cluster. Microsoft, Google, Amazon, and Oracle sit atop most of the compute. The frontier-model layer is a four-name list. And the application layer — the office suites, the developer tools, the enterprise agents — is, increasingly, being built by the same four names, sometimes through acquisition and more often through native product rollout.

Who gets squeezed

The specialist model providers are the first casualty of this dynamic. Their pitch — better performance in a narrow domain at a lower cost — is being undercut on both sides simultaneously. On the top end, the frontier models keep absorbing narrow tasks. On the bottom end, the same models are being offered at prices that make the unit economics of a smaller vertical model difficult to defend.

The second casualty is the independent application-layer startup. The most successful vertical apps — legal research, medical scribing, developer tooling — are the ones being copied fastest by the platform players.

What breaks the pattern

Two things could. One is a regulatory intervention that separates the layers, in the way antitrust separated Microsoft's OS from its application business a generation ago. That is, at this moment, extremely unlikely. The other is a foundational architectural change — the arrival of a training approach that no longer requires the compute footprints that make the current oligopoly rational. That, too, is unlikely, but it is not impossible.

Absent one of those, the map you are looking at now is, roughly, the map for the next four years.

About the author

Elena Vance

Tech Editor based in San Francisco. Covers AI infrastructure and the people building it.