Foundation Models
The labs that train the models the rest of the stack runs on.
What this layer does
This is the layer of capital-burning research labs. A frontier training run costs hundreds of millions to billions of dollars in compute alone, before headcount or data. The output is a model that other layers either resell (cloud APIs), wrap (applications), or compete with (open-weight challengers).
The economics here are extreme: revenue scales fast, but compute cost scales with usage. Gross margins depend on how cheaply the lab can serve inference — which depends on the model architecture, the GPUs it’s tuned for, and any custom silicon (Anthropic on Trainium, Google on TPU).
There are no pure-play public model labs. Exposure comes through the partners and infrastructure each lab depends on.
Sub-categories
The four (or five) labs training at the absolute frontier of compute scale.
Hyperscaler and platform-owned models built for internal product distribution.
Models released with downloadable weights — the “Linux of AI” thesis. Critical floor on closed-model API pricing.
Smaller labs aiming at specific markets (enterprise, regulated, sovereign).
Modality-specific generative model labs. Often the engines behind the creative apps in Layer 01.
Foundation models for the physical world — the next frontier of training scale.
Bio, materials, weather, math — non-text models, often partnered with big-tech compute providers.