France's economy minister Roland Lescure told reporters in San Francisco on September 3, 2026: "If European AI comes down to just Mistral, then we're screwed. You can't put all your eggs in one basket — it's a whole ecosystem that needs building." He was in the Valley to meet investors and AI companies, OpenAI and Stripe among them named on his itinerary. Asked directly whether Mistral should keep developing frontier models rather than pivot toward being an infrastructure and services provider, he declined to say: "Whether they end up more focused on foundational model development, or become a very efficient, profitable services company — we'll see. Companies do what they can and what they want."
The EU already made this call in June
Lescure's "don't put all your eggs in one basket" isn't hypothetical — the European Commission has already acted on it. In June 2026, the Commission's Frontier AI Grand Challenge selected the EUROPA consortium — led by the Italian startup Domyn (based in Milan, formerly iGenius) with Germany's Fraunhofer-Gesellschaft — to build the EU's own sovereign open-source frontier model: over 400 billion parameters, trained across all 24 official EU languages, targeted for delivery within a year. The consortium gets up to 2.5% of EuroHPC's total compute capacity on EU public supercomputers, backed by a roughly 6,000-chip Nvidia Blackwell cluster already committed. Mistral wasn't part of it. The EU's flagship bet on sovereign frontier AI is already running through Italy and Germany, not through the company most people would name first if asked for Europe's AI champion.
The same week, the state doubled down on Mistral specifically
Set against that, the domestic picture reads almost like a different argument. On August 18, budget minister David Amiel said the French state would rely on "sovereign" AI providers like Mistral for government use — explicitly excluding OpenAI — following a cyberattack on the French tax administration that leaked tax data on 678,000 individuals and professionals. On September 3, the finance ministry confirmed a reinforced, non-exclusive state partnership with Mistral, concluded through a public-procurement process: €6 million over 2026–2027, with Mistral engineers seconded directly into government ministries to build out three priority use cases — cybersecurity, processing mass litigation for an overloaded justice system, and fraud detection at the tax authority. The tools will run on whichever infrastructure fits the data's sensitivity: Mistral's own, the "sovereign cloud" provider Outscale, or the ministries' own data centers. This builds on a June rollout that already put a Mistral-built AI assistant in front of two million public-sector employees. So the same government arguing publicly that Europe can't just be Mistral is, at the same time, narrowing its own sovereign-AI procurement specifically around Mistral for the use case where sovereignty matters most: the state's own data. Those two positions aren't strictly contradictory — "Europe needs a broader ecosystem" and "France trusts Mistral with government workloads specifically" can both be true — but they pull in different directions loudly enough that reporters asked about it directly.
Mistral itself, valued near €12 billion with the French state an indirect shareholder via Bpifrance, is reportedly in talks for a new funding round of roughly €3 billion that would value it near €20 billion — a figure that would still be a rounding error against OpenAI's reported $852 billion valuation. The company has spent this year building out infrastructure — a first French data center, a commercial deal with Microsoft — while continuing to publish its own frontier models. Its three co-founders, Timothée Lacroix, Guillaume Lample, and Arthur Mensch, now rank 20th on Challenges' list of France's wealthiest people, their combined fortune having doubled over the past year to €6 billion — the fourth-largest gain of anyone on that list.
Palantir's Alex Karp says the sovereignty strategy could backfire
Not everyone treats "buy French" as the safer bet. Palantir CEO Alex Karp, speaking on the sidelines of a G20 ministerial meeting in North Carolina the same day, warned that France risks hurting itself if it believes routing government work through French companies delivers real sovereignty when those companies' own AI still runs on American or Chinese models underneath. He was responding to Prime Minister Sébastien Lecornu's June announcement that France would replace Palantir's data-analysis software — used by the DGSI, France's domestic intelligence agency, since 2016 — with a system from French company ChapsVision, in the name of "strategic autonomy." Without naming ChapsVision, Karp criticized the idea of handing deployment of American and Chinese "closed models" exclusively to French companies, and argued that if Paris lets its decisions be "guided by fear and by people who don't really understand what's at stake," "that will hurt France." He added that "the regulatory environment clearly isn't helping Europe or France, and you need regulation that actually helps" — a jab at the same EU rulebook this blog covered funding compliance support for, a day earlier. Karp said he wasn't speaking in American companies' interest, though there's a fair chance his remarks get read that way regardless.
"Unpredictability," in one word
Asked whether a Trump administration could pressure Europe by cutting off access to the best US models, Lescure answered with one word: "unpredictability." That's not an abstract fear. Washington forced Anthropic to suspend all foreign nationals' access to Fable 5 and Mythos 5 for 19 days in June, an episode that's since become the reference point European officials reach for whenever the sovereignty argument comes up. Lescure reaching for the same example, unprompted, while standing in the city where that decision was made, is the clearest evidence yet of how much that one incident reshaped the conversation.
What to expect next
- Watch whether Mistral responds directly. Lescure's refusal to recommend a path — frontier lab or profitable services company — puts the decision back on Mistral's own leadership in public, in a way the company hasn't had to answer to directly from its own government before.
- Watch how "non-exclusive" the reinforced state partnership actually is in practice. The ministry's own wording leaves room for other sovereign providers; whether any others actually get government cybersecurity work will say more than the announcement itself.
- Watch for the EUROPA consortium's first checkpoints. A 400-billion-parameter open-source model delivered within a year is an aggressive timeline — the Domyn/Fraunhofer team's first public benchmarks will be the first real test of whether "not Mistral" was the right bet.
- Watch for a response to Alex Karp's warning, from Sébastien Lecornu, Bercy, or ChapsVision. A public swipe from a major US vendor's CEO directly at the "strategic autonomy" framing is the kind of comment this fight has been missing so far.