Skip to content
Predictive Systems
PSI Daily

Lescure: Europe Is Doomed If AI Sovereignty Is Just Mistral

Sovereignty

Roland Lescure went to San Francisco to meet OpenAI and came home with a warning for Paris. A national champion is not an ecosystem. One kill switch in Washington already made the point.

mistral ai

If Mistral is the plan, there is no plan

France’s finance minister flew to Silicon Valley to see OpenAI and Stripe. The line that travelled was about a French company.

The quote

“If European AI amounts to Mistral, then we are doomed,” Roland Lescure told reporters in San Francisco on Wednesday, in AFP’s English rendering. The French, carried by Le Monde, is sharper: “Si l’IA européenne se résume à Mistral, alors nous sommes fichus.” He added that Europe cannot put all its eggs in one basket. “It is a whole ecosystem that we need to develop.”

The French state is an indirect shareholder in Mistral through Bpifrance. The startup is valued at nearly 12 billion euros. It is still described as Europe’s only house regularly publishing first-tier models, and still behind US and Chinese peers. Lescure called it “a company we can be proud of.” That was the compliment. The policy was the warning.

The same day, more Mistral

The timing is the story. On the same morning Le Monde, citing Les Echos and Bercy, confirmed that Paris is deepening a non-exclusive public contract with Mistral: 6 million euros in 2026-2027, engineers seconded into ministries, trials in cybersecurity, tax-fraud targeting, and mass litigation at Justice. Tools can sit on Mistral’s kit, Outscale’s “sovereign” cloud, or government data centres, depending on sensitivity. This follows a Mistral assistant rolled out in June to two million public servants, and a summer hack at the tax authority that leaked data on 678,000 people and firms. On 18 August, budget minister David Amiel had already said the state would lean on “sovereign” suppliers such as Mistral and keep OpenAI out after that breach, Maddyness notes, again from AFP.

So the minister is not dumping the champion. He is refusing to let the champion be the strategy.

One basket, one kill switch

Asked whether the Trump administration might squeeze Europe by cutting off the best models, Lescure answered with one word: “imprévisibilité.” Unpredictability. AFP and Le Monde both tie that to June, when Washington forced Anthropic to cut foreign access to its most powerful Claude models for two weeks. CSIS, writing on 16 June, named the models (Fable 5 and Mythos 5), dated the Commerce letter to 12 June, and recorded Anthropic’s own line: it had to “abruptly disable” them for all customers to stay legal. That is the single-supplier risk in hardware form. A US agency, not a market, turned the lights off.

Maddyness and the AFP English wire add a second diversification already under way: in June the Commission gave an Italian-German consortium, not Mistral, a contract to build a European frontier model on public supercomputers. Lescure’s own gloss: US models are ahead, Chinese models are not far behind, “and there is stuff going on in Europe which might lead to development as well.”

What he would not pick

AFP asked the obvious follow-up. Should Mistral stay a frontier-model lab, or become a compute and services shop, the path implied by its first French data centre and its Microsoft commercial deal? Lescure would not choose. “Companies do as they can and as they want,” he said, in Le Monde’s French. In English, per AFP: “How they’re going to do it, we’ll see.”

That dodge is the European AI problem in miniature. Sovereignty is sold as a national model. The minister just said a national model is not enough. He still will not say what the company should be. The ecosystem he wants does not yet exist. The contract with Mistral does.

Sources

Source: Le Monde with AFP