Before, you built an AI per task: one system for translation, another for summarising, a third for classification. Each required its own data, its own training, its own team. A foundation model upends this logic: you train a single general-purpose system, at enormous cost, then adapt it to a thousand uses. It's the difference between building a house room by room and pouring foundations that everyone else builds on.
The term is everywhere, often without a definition. Yet it names the idea that structures the entire current AI economy.
The definition
A foundation model is a model trained at very large scale on highly diverse data, with no specific task in mind, which then serves as the basis for numerous applications.
Three characteristics define it. It is general-purpose: it wasn't trained to translate, it learned language. It is adaptable: it can be specialised through prompting, through RAG or fine-tuning. And it exhibits emergent capabilities: it can do things it was never explicitly taught, as we explained in our article on LLMs.
Why it changed everything economically
Here's the genuinely interesting part, because it explains the shape of the market.
Training a foundation model is extraordinarily expensive: months of compute across tens of thousands of chips, as we described in our article on the two costs of AI. This cost is fixed and concentrated.
By contrast, adapting that model to a task costs very little. Writing a good prompt costs nothing, hooking up a RAG pipeline is modestly priced, and light fine-tuning remains accessible.
This asymmetry produces a very particular market structure: a handful of players able to pay for the foundations, and thousands of companies building on top. That's why the sector is at once extremely concentrated at the model level and extremely open at the application level.
The practical consequence is significant: building your own foundation model makes sense for almost no one. The maths is simple, and it's crushing. What makes sense is choosing carefully which foundation to build on, knowing how to switch, and focusing your efforts on what's your own: your data, your interface, your understanding of the business. This is exactly the recommendation we made in our selection guide: don't pick a model, pick an architecture that lets you switch models.
The fragility of this structure
This business model rests on one assumption: that foundations remain expensive and scarce. Yet several trends are undermining it.
Open-weight models make quality foundations available for free. API prices are collapsing. And smaller models, like the one Meta released to run on a single graphics card, make accessible what yesterday required an entire infrastructure.
If foundations become abundant, value shifts entirely to what's built on top. That's already happening, and it explains why the most highly valued players are no longer necessarily those with the best model, but those who own the relationship with the user.
What to remember
The concept of a foundation model explains just about everything that seems strange about this sector: why so few companies matter, why they burn so much money, why everyone builds on the same bricks, and why any single decision by a supplier affects thousands of downstream products.
It's also what makes the question of dependency so central. When an entire ecosystem rests on a few foundations owned by a few players, the resilience of the whole depends on choices no one else controls.