Foundation Model
A foundation model is a large machine-learning model trained on broad data at scale — text, code, images — that can be adapted to many tasks it was not specifically trained for, by prompting or fine-tuning. Large language models are the best-known kind, and the companies that build them are the most expensive acquisition targets, and acquirers, in technology.
- Also known as
- Base model, Frontier model, Pre-trained model
- Used at
- the base layer of nearly every AI product; a fintech "AI feature" is usually a foundation model with the fintech's data around it
- Built by
- a handful of labs — OpenAI, Anthropic, Google, Meta, Mistral and a few others — at a cost of hundreds of millions to billions per generation
- Consumed by
- everyone else, through an API priced per token or as open weights run on the buyer's own hardware
- In deals
- the labs are bought for their people and models (acqui-hires and licensing deals); the users are bought for their data and distribution
- The risk
- dependency on a supplier whose pricing, terms and availability the customer does not control
How a foundation model works in practice
A foundation model is trained once, at enormous cost, to predict the next token in a vast corpus, and what it learns in doing so — language, reasoning, code, the structure of a contract — can then be pointed at any task a prompt can describe. A fintech does not train one; it calls one. Its underwriting assistant, its customer-service agent, its document-extraction pipeline and its compliance screening each send a prompt and the company's data to a model it rents by the token, or to an open-weights model it runs itself, and build their product in the layer above: the retrieval of the right documents, the guardrails, the evaluation, the audit trail a regulator will ask for.
That layered structure shapes the M&A. The labs are acquired — or, since the regulatory scrutiny of 2024, licensed and acqui-hired — for the model and the people who can build the next one, at prices that bear no relation to revenue. The application companies are acquired for what the model cannot give a buyer: proprietary data, a distribution channel, a regulated licence, a customer base that has already adopted the workflow. A buyer of a fintech with an "AI product" is buying the second kind of company, and prices the model dependency as a supplier risk: what happens to the margin if token prices double, or the terms of service exclude financial advice, or the supplier is acquired by a competitor.
For a regulated firm the model is also a compliance question: where the data goes, whether the outputs can be explained to a supervisor, and who is liable when they are wrong.
Example
A lending platform's credit-decision assistant runs on a rented foundation model, sending 40m tokens a day at a cost of €1,100. A buyer values the platform at €60m, largely for its 14,000 SME customers and its lending permission. Diligence models three scenarios: the supplier doubles prices (−€400,000 of annual margin), the supplier's terms restrict credit decisions (six months and €1.5m to move to an open-weights model on the buyer's own infrastructure), and the supplier is acquired by a bank the platform competes with. The buyer takes a 15% price reduction and a covenant that the sellers migrate the pipeline to a second model before the earn-out is measured.
Common mistakes
- Valuing a company as an "AI company" when its AI is a rented model and its value is its data and customers. - Diligence that reads the AI feature and not the supplier contract behind it. - No second supplier, so a pricing change is a margin change the company cannot answer. - Sending regulated customer data to a model whose terms the compliance team never read. - Confusing a fine-tuned model the company owns with a foundation model it rents.