Europe’s AI sovereignty is under threat. Could Mistral be the answer?

Arthur Mensch spent 90 minutes in front of a nearly empty room at the French National Assembly back in May, warning lawmakers that Europe had about two years to build its own AI infrastructure before becoming what he called a permanent vassal state. A month later, Washington handed him a live demonstration. The Commerce Department temporarily blocked foreign access to Mythos, Anthropic's flagship model, over concerns about its ability to autonomously find and exploit software vulnerabilities. European firms that had been racing to get their hands on it suddenly found themselves locked out, and Mensch's warning stopped sounding theoretical.

That's the opening Mistral has been waiting for. The Paris startup, founded in April 2023 by Mensch, Timothée Lacroix and Guillaume Lample, has raised roughly $4 billion and sits at a $23 billion valuation — respectable, until you put it next to Anthropic's $965 billion or OpenAI's $852 billion. Mistral's models haven't kept pace either. It didn't ship a reasoning model until mid-2025, and even then DeepSeek beat it there. Its newer Mistral 3 lineup, including the flagship Large 3, still benchmarks close to where OpenAI was a year and a half ago, though Mistral insists most enterprise work doesn't need bleeding-edge horsepower anyway.

So the company is betting on something other than raw model quality: efficiency, specialization, and deep customer entanglement. Smaller Ministral models can run on a single GPU. The pitch to enterprises is Palantir-style — embed engineers, sign multiyear strategic contracts, avoid the runaway bills that reportedly hit $500 million for one Claude customer. It's working well enough that Mistral's annualized revenue passed $400 million in 2025 and is tracking toward $1 billion this year, even if that's a rounding error next to Anthropic's estimated $47 billion run rate. The company has also pushed into industrial AI, buying Austrian startup Emmi AI to feed deals with Airbus and BMW, and into defense, a move Mensch defends bluntly: AI is dual-use technology and the world, unfortunately, isn't at peace.

The awkward part is that Mistral's sovereignty pitch rests on infrastructure it doesn't fully control. Its new data centers in France and Sweden, and its Mistral Compute cloud platform unveiled with Nvidia and praised by Macron as historic, all still run on Nvidia chips. Lacroix says the company wants to diversify away from a single supplier, but building its own silicon isn't happening anytime soon. Mensch himself pushes back on the sovereignty framing, arguing it's less about isolation and more about owning enough of the value chain to keep reinvesting and improving. UC Berkeley's Stuart Russell puts the real gap plainly: leading labs are compounding algorithmic tricks and iterating fast because they have compute to burn, and that's the hard part to close.

Even the memes reflect the pressure. A fake Mistral model called Le Chaton Fat went viral this summer with invented benchmarks claiming it beat Anthropic — a joke, but one that Mensch reads as a signal of how much people want Mistral to actually deliver. Lample says new models are coming this summer. Lacroix, more cautious, just wants people to stop expecting a cat that doesn't exist yet.