Definition
The term was coined by researchers at Stanford's Center for Research on Foundation Models in 2021 to describe models trained at scale on broad data and then adapted, by prompting or fine-tuning, to a wide range of downstream tasks.
In regulation the idea appears as 'general-purpose AI models', which the EU AI Act covers with specific obligations for their providers. For companies implementing AI, the practical point is that most work builds on someone else's foundation model, so its provider's terms, data handling and update cycle become part of your risk.