Foundation Model

Foundation models are the big general-purpose models that underpin AI products. Claude, GPT, Gemini, Llama — these are the foundations on which more specialized applications are built.

The term emphasizes that these models are general (not task-specific), expensive to train (typically tens to hundreds of millions of dollars), and designed to be adapted via prompting, fine-tuning, or distillation rather than trained from scratch by every team that wants to use them.

For operators in 2026, the practical implication: you almost certainly don't need to train a foundation model. You build on top of one. The interesting decisions are about which foundation model (capability per dollar varies meaningfully), how to prompt it, how to ground it with your data, and how to wrap it in product.