Foundation Model

A foundation model is a large-scale model — typically an LLM, but also multimodal or vision models — pre-trained on a broad, diverse dataset (a large fraction of the public internet, books, code repositories) to develop general capabilities, with the explicit intent that it will be adapted afterward to specific downstream tasks rather than used only for one narrow purpose. The term (coined by Stanford's Center for Research on Foundation Models in 2021) captures a shift in how AI is built: instead of training a separate model from scratch for every task (sentiment analysis, translation, summarization each getting their own model), you train one enormous general model once, at huge cost, and then "adapt" it cheaply and repeatedly — via prompting, RAG, or fine-tuning — to hundreds of different applications. Claude, GPT-4/5, Gemini, and Llama are all foundation models; a SaaS company's customer-support chatbot, code-review assistant, and marketing-copy generator might all be built on the exact same underlying foundation model, differentiated only by prompts, retrieved context, and (sometimes) fine-tuning. This matters for builders as a mental model for build-vs-buy decisions: virtually no SaaS company today trains a foundation model from scratch (the compute cost is prohibitive for all but a handful of well-funded AI labs) — instead, the entire AI product layer is built on top of foundation models accessed via API or self-hosted open weights. A concrete example: a legal-tech startup doesn't train its own language model to understand contracts; it takes a foundation model like Claude, adds a RAG pipeline over a corpus of contract templates and case law, writes a detailed system prompt establishing legal-review behavior, and ships a "contract review assistant" — the actual language understanding capability is entirely inherited from the foundation model, while the product's differentiation comes from the data, prompts, and workflow built around it. Foundation models are generally categorized as closed (API-only access, like Claude and GPT) or open-weight (downloadable model files you can self-host and modify, like Llama and Mistral), a distinction that matters for data privacy, customization depth, and cost at scale. The foundation-model paradigm also reshapes how SaaS companies think about their technical moat: since the underlying language capability is largely commoditized and accessible to any competitor via the same APIs, durable differentiation increasingly comes from proprietary data (what you can retrieve and ground answers in that competitors can't access), workflow integration (how deeply the AI feature is woven into an existing product surface), and evaluation/quality-control infrastructure (how reliably you can keep the AI feature accurate at scale) — not from the model itself, which any well-funded competitor can access on comparable terms.