One base, many uses
A foundation model is a large model, pre-trained on an enormous dataset, that serves as the base for a wide range of uses. The term was proposed by a Stanford research group in 2021 and covers LLMs such as GPT, Gemini, and Claude, as well as large models for images and audio.
What was new about it
Conventional AI meant building a separate model for each job — one for translation, one for image classification. Foundation models invert that: grow one enormous general base, then adapt it to each use through fine-tuning or prompting.
It is the difference between building custom houses one at a time and building one solid apartment block, then fitting out each unit. Development economics changed fundamentally.
Why it matters to users
Most AI services in use today sit on top of a small number of foundation models. ChatGPT, Copilot, and the assistants various companies ship all trace back to a handful of lineages — GPT, Claude, Gemini, Llama. Countless applications, very few foundations, is one of the most useful structural facts about the industry.
What to take away
When coverage mentions competition to build foundation models, or a national foundation model program, read it as being about building the base models themselves. Very few organizations can do it, because it requires enormous compute and data — which is why it turns up in economic as well as technology news. It overlaps heavily with LLM, but foundation model is broader, covering models beyond language.
A foundation model is not a finished product. What you use has finishing stages such as fine-tuning and RLHF, plus safety work, layered on top. Keeping the base and the product distinct makes launch announcements much easier to read.