On October 16, 2026, Arthur Mensch, co-founder and CEO of Mistral AI, publishes on X an article titled "Build or surrender." In it, he contests an idea he says is widespread in Europe: leaving frontier AI models to the Americans and the Chinese. Its proponents, whom he describes as serious, argue that a frontier model costs hundreds of millions of euros per generation, soon billions, and that it would be better to follow the internet playbook, where value was created in applications, by betting on industrial data, SMEs and domain expertise, and by relying on open models published by others.
For the author, this is a fundamental mistake that would lock Europe into an unprecedented regime of economic dependence. The cloud precedent illustrates this: Europe developed the applications and left the platforms to others; twenty years later, its companies pay more than 260 billion euros per year to American cloud and software providers.
AI differs from the internet, which relied on open protocols with no owner. A closed model has an owner who decides what the applications and agents built on it can do: it changes prices, restricts uses, withdraws a version. The risk is not theoretical: in June, an American lab suspended access to its most advanced models, before restoring it, without Europe being consulted.
Open models only partly solve the problem. They can be downloaded, self-hosted and audited, but without contributing to the ecosystem, one chooses neither what is released, nor when, nor under what license. Building frontier systems becomes a security matter: models play a central role in cyberdefense and in protecting critical infrastructure, and a state must be able to audit them, adapt them, and be certain of accessing them tomorrow. Fundamental research is also affected: these models are becoming indispensable to science, some are reserved for a handful of companies, and the use of AI for AI research creates threshold effects.
Pitting frontier research against industrial deployment would therefore be a trap: for the most demanding uses, mastering the model is a prerequisite. Europe has the means to do so, according to the author: Black Forest Labs and ElevenLabs dominate their category, and Mistral, with fewer resources, trains reasoning models among the best open models. The company will multiply its compute capacity sixfold and will raise capital.
Finally, he calls for breaking with pessimism, a narrative that suits those who are leading. Otherwise, Europe will only have access to second-rate export versions, in exchange for royalties running into trillions of euros.