Model layer vs. app layer
Also known as: foundation layer vs application layer, infrastructure vs app, AI stack layers
The AI value chain has rough layers: at the bottom is compute infrastructure (chips, data centers); above that sits the model layer (the frontier labs training the actual LLMs); above that is the application layer (products built on those models). Each layer has different economics, different barriers to entry, and different ways of building a moat.
The key strategic question for builders is: where does durable value accrue? The model layer is capital-intensive and dominated by well-funded labs. The application layer requires less capital but faces competition from any team that can call the same APIs. Over time, as models commoditize, application-layer products that have built proprietary data, deep workflow integrations, or distribution advantages look more durable than those that are essentially a thin interface over a foundation model.
This framing surfaces a real tension for early-stage AI startups: should you build on top of existing models (faster, cheaper, more accessible) or invest in fine-tuning or specialized models to create differentiation? Most successful early-stage products start at the app layer and add model-level differentiation only when there is strong evidence that generic models can't solve the problem well enough.