Baseten, an AI infrastructure platform specializing in model fine-tuning and deployment, has spotted a trend. Their director Charlie O'Neill noticed that top AI applications like Cursor, Notion, and Cognition are all turning away from big tech's general-purpose APIs, instead training their own models on top of open weights.
https://twitter.com/oneill_c/status/2054604986269802579
The logic comes down to a reversal in cost and performance. In specific niche use cases, these custom models can now match or even beat frontier models — while costing a tenth as much to run. The bigger advantage? The application layer holds exclusive data feedback loops. Take Cursor: the code users keep or delete is the perfect reinforcement learning reward signal. Iterating models on real workflow data is something big tech can't catch up with using just public data and raw compute.
Large model companies' business model relies on a single general-purpose model serving all customers, which means they can't dive deep into vertical use cases to build proprietary data flywheels. As development barriers keep falling, the only way for AI apps to stay competitive is to tightly bind the model to their core business. The exclusive, real user interaction data that big tech can't access is becoming an irreproducible moat.