# What is a foundation model? Definition and business implications

> A foundation model is a generalist AI model trained on massive, diversified data, which serves as a reusable base for dozens or even hundreds of different business applications, by adaptation (fine-tuning, RAG, prompt) rather than by specific training.

## Definition
The term *foundation model* was introduced by the Stanford CRFM in 2021 (Bommasani et al., *On the Opportunities and Risks of Foundation Models*) to designate a new class of models characterised by two properties. First, large-scale training on generalist corpora: these models are not designed for a specific task, but to capture the statistical regularities of a vast domain (language, image, code). Second, their downstream adaptability: the same model can be specialised afterwards for translation, summarisation, classification, code, image analysis, through adaptation techniques (fine-tuning, RAG, prompt engineering) far less costly than the initial training. GPT, Claude, Gemini, Llama, and Mistral are textual foundation models. Stable Diffusion and GPT Image (OpenAI) are visual foundation models. Nearly all current enterprise AI applications rely on a foundation model, without necessarily knowing it.

## Example
The 2021 Stanford CRFM report identified foundation models as a paradigm shift: before 2018, each AI task (translation, summarisation, classification, vision) required its own model specifically trained, on dedicated data. After 2018 and the emergence of large transformers, a single generalist model can address dozens of tasks through light adaptation. Illustrative case for a mid-cap: instead of funding five specialised AI projects, each with its own data and model, it adapts the same foundation model five times, through prompting, RAG or light fine-tuning. The entry cost drops sharply, because the provider has already paid for the initial training.

## For a leader
The emergence of foundation models is the most important structural change in enterprise AI since 2018. Three implications for the executive. First, you will not build a model, you will choose one. Training a frontier foundation model costs tens to hundreds of millions of dollars, and that cost has grown about 2.4-fold each year since 2016, according to Epoch AI. It is out of reach. The strategic competence shifts to the selection, adaptation, and orchestration of existing models. Second, the choice commits: changing foundation model after going to production requires rewriting the prompts, redoing the fine-tuning, sometimes re-indexing the RAG bases. It is a multi-year decision, to be arbitrated according to the provider's trajectory, not on short-term price alone. Third, model homogenisation creates a systemic risk: a flaw or bias in a foundation model propagates to all adapted applications downstream. Upstream auditing becomes a governance issue.

## FAQ

**What is a foundation model?**
It is a generalist AI model trained on massive, diversified data, serving as a reusable base for dozens, even hundreds, of different business applications. You adapt it (fine-tuning, RAG, prompt) rather than train it specifically for each task. The term was introduced by the Stanford CRFM in 2021.

**How does a foundation model change the economic equation?**
Before 2018, each AI task (translation, summarisation, classification, vision) required its own specifically trained model. Since then, a single generalist model addresses dozens of tasks through light adaptation. For a mid-cap, that means adapting an existing model five times at 10,000 to 30,000 euros each, instead of funding five specialised developments at 200,000 euros apiece: the entry barrier collapses by an order of magnitude.

**Should you build your own foundation model?**
No. Its cost (tens to hundreds of millions of dollars) places it out of companies' reach. You will not build a model, you will choose one. The strategic competence shifts to selecting, adapting, and orchestrating existing models.

**Is the choice of a foundation model reversible?**
Only with difficulty. Changing it after going to production means rewriting the prompts, redoing the fine-tuning, sometimes re-indexing the RAG bases. It is a multi-year decision, to be weighed against the provider's trajectory, not short-term price alone.

**What governance risk should you anticipate?**
Homogenisation. Because nearly all applications rely on a handful of foundation models, a flaw or bias in one of them propagates to every adapted application downstream. Upstream auditing therefore becomes a governance issue, not a technical detail.

## See also

- [LLM (Large Language Model)](https://wyp.agency/en/glossary/llm/)
- [Transformer](https://wyp.agency/en/glossary/transformer/)
- [Generative AI](https://wyp.agency/en/glossary/generative-ai/)
- [Open-source model](https://wyp.agency/en/glossary/open-source-model/)
- [Multimodal](https://wyp.agency/en/glossary/multimodal/)
- [AGI (artificial general intelligence)](https://wyp.agency/en/glossary/agi/)
- [Deep learning](https://wyp.agency/en/glossary/deep-learning/)

## Sources

- On the Opportunities and Risks of Foundation Models, Bommasani et al., Stanford CRFM, arXiv:2108.07258, 2021 (https://arxiv.org/abs/2108.07258)
- Center for Research on Foundation Models, Stanford HAI (https://crfm.stanford.edu/)
- How much does it cost to train frontier AI models?, Epoch AI, June 2024 (https://epoch.ai/blog/how-much-does-it-cost-to-train-frontier-ai-models)
- Deprecations, OpenAI (DALL·E removed from the API on 12 May 2026) (https://developers.openai.com/docs/deprecations)

---
Entry from the WYP.agency AI glossary · https://wyp.agency/en/glossary/foundation-model/ · Last updated : 2026-10-14