Mistral's CEO urges Europe to build its own frontier AI models

“Produce or endure: the choice is ours for now.” That is the last sentence of the op-ed that Arthur Mensch, co-founder and chief executive of Mistral AI, published on October 15 in Le Monde, then in English on LinkedIn. The head of Europe's leading generative AI company attacks an idea that, he says, is gaining ground in Europe: leaving the most advanced models, known as “frontier” models, to the United States and China.

The reasoning he challenges has two parts. A frontier model costs hundreds of millions of euros per generation, soon billions. So Europe would be better off repeating the Internet playbook: creating value in applications, industrial data and the know-how of SMEs, relying on open models published by others. For Arthur Mensch, this is a “fundamental mistake.”

Why the Internet is the wrong model, in his view

His first argument: the Internet was built on open protocols that belonged to no one. A closed AI model, by contrast, has an owner. And whoever controls the model sets the terms: the price, the permitted uses, the withdrawal of a version. According to the op-ed, European companies already pay more than €260 billion a year to US cloud and software providers. He does not say how this figure is calculated.

Open models are not enough either, he continues. They can be downloaded, hosted and audited, but Europe would choose neither the release schedule, nor the features, nor the license of later versions if it does not take part in their development. For cyberdefense and critical infrastructure, a state must be able to audit, adapt and count on lasting access.

The June example

To illustrate this risk of dependence, Arthur Mensch points to a decision made in Washington in June. Silicon.fr reads this as the temporary, worldwide suspension of access to Anthropic's most advanced models, requested by the US government. “The decision was not made in Europe, nor for European reasons,” he writes.

The op-ed also contains an announcement: Mistral plans to multiply its computing power by six over the next twelve months, and calls for more capital to speed up. It cites Black Forest Labs and ElevenLabs as other European examples. The company's latest flagship model, Mistral Large 4, was unveiled in early October. Computing capacity is also being built by other French players, such as Scaleway and its Swedish supercomputer.

A plea that also serves its author

The text comes from an interested party, one that raises funds and sells its models. Silicon.fr notes that NGOs and European officials have accused Mistral of helping, along with Aleph Alpha and the French and German governments, to water down the first version of the AI Act, the EU regulation on AI. Mistral says it supports transparency on training data, without an obligation to disclose trade secrets.

Finally, Arthur Mensch warns against the narrative of a Europe that has already fallen behind: “all it takes is for us to believe it to give our competitors the advantage.” Mistral's next fundraising rounds, and how the European Union allocates its AI budgets, will show whether the appeal has landed.