The Stanford AI Lab published a paper in August 2021 introducing the term "foundation models" to describe the family of large pre-trained models — GPT-3, BERT, CLIP, and their successors — that form the basis for a wide range of downstream AI applications. The paper makes an important observation: these models function as a platform, not just as a tool. The capabilities they provide are general enough that entire categories of applications can be built on top of them, in the same way that entire categories of mobile applications are built on top of iOS and Android.
I want to explore what the platform analogy implies for infrastructure investing — because I think it is both useful and importantly limited.
What the Platform Analogy Gets Right
The mobile platform analogy captures something real. When iOS launched, it created demand for an entirely new layer of development tools, analytics infrastructure, monetization platforms, and distribution services. The companies that built those layers — Flurry, Appsflyer, RevenueCat, Adjust — became large businesses by serving the ecosystem of applications that grew on top of Apple's platform. A similar dynamic is plausible for foundation models: as applications proliferate on top of GPT-3 and its successors, demand for the infrastructure layer that supports those applications will grow.
The parallel also captures something about the power dynamics. Platform owners capture significant value; the infrastructure layer above them is less defensible unless it provides enough depth that the platform can't easily replicate it. This is the key risk in the foundation model ecosystem: OpenAI building features into its API that commoditize what independent infrastructure companies do.
What the Analogy Misses
The mobile platform analogy breaks down in one important way: foundation models are not locked platforms. GPT-3 was not a singular capability that required building on OpenAI specifically. It demonstrated a pattern — large-scale pre-training on diverse data — that others could replicate. Anthropic, Cohere, AI21, and dozens of open-source projects followed. The infrastructure layer for foundation model applications does not need to be coupled to a single platform provider in the way that iOS apps are coupled to Apple's APIs.
This actually makes the infrastructure thesis stronger, not weaker. A developer ecosystem that needs to switch between multiple foundation models — or wants the optionality to do so — needs a routing and abstraction layer that the model providers themselves will not provide. The infrastructure bet is on the plurality of foundation models, not on any single platform winning.