The three founders of Mistral came up through Polytechnique and the École Normale Supérieure, spent a few years at DeepMind and Meta, and then returned to Paris in the spring of 2023 to build their own thing. Coming home is the part that took twenty years to become possible. For most of the stretch I want to talk about, France was superb at manufacturing people and close to hopeless at keeping companies. Yann LeCun did his early work on convolutional nets partly in Paris and then spent three decades building AI for American labs, most of it at Meta’s. The Hugging Face founders trained at Polytechnique, incubated at a Paris startup campus, and moved the headquarters to New York the moment it got serious. That was the pattern. The brains were French, and the cap tables weren’t.
So when someone asks me what actually changed in French tech between roughly 2006 and now, I don’t point at any single program. I point at the moment the export machine started running in reverse.
The scaffolding got built in layers, and the first one predates the smartphone. In September 2004, the government set up the pôles de compétitivité, regional clusters meant to bring companies, public labs, and universities together around a shared theme. The first call for projects closed in early 2005; by 2007, there were 71 of them, and from 2009 to 2012, the state invested about €1.5 billion in cluster research partnerships. The clusters are the least glamorous part of this story and the part foreign investors understand the least. They mostly reinforced existing innovation hubs rather than creating new ones. But they made “go find the lab and the supplier and the co-funder in your region” a normal, cheap thing to do instead of a heroic act.
The temperature changed in 2013, when Fleur Pellerin launched La French Tech, and the government folded its investment arms into a single public bank, Bpifrance. French Tech was mostly branding at first, a red rooster and a label handed to nine cities in late 2014, but branding was exactly what a country with a complex about its own entrepreneurs needed. Bpifrance was the real machine. It has spent the decade since doing volume: in 2023 alone, it put roughly €677 million of equity into startups, around a tenth of all French venture that year, on top of billions more in loans and subsidies. When American founders ask me who the anchor co-investor is on a French round, the answer is almost always the state, and that is not the insult it would be in Silicon Valley.
Physical proof showed up in June 2017, when Xavier Niel finished pouring something like €250 million of his own money into a derelict rail freight hall in the 13th arrondissement and reopened it as Station F. It is the biggest startup campus in the world by a wide margin, 34,000 square meters under one 310-meter roof, a thousand startups at a time, thirty-odd programs run by Google, Microsoft, Meta and LVMH. Macron inaugurated it in the same summer he started calling France a “startup nation,” which was no coincidence. Hugging Face was the campus’s first unicorn, which tells you both what the place could produce and where those companies tended to end up.
The money got serious in October 2021, when Macron announced France 2030 and then let it grow from €30 billion to €54 billion by 2023. Most of that has nothing to do with software. It goes to reindustrialization: small modular nuclear reactors, hydrogen, batteries, and a €5.7 billion expansion of STMicroelectronics’ fab at Crolles near Grenoble, run with GlobalFoundries, to make chips on a node where France actually holds a lead. By last year, the plan had committed north of €30 billion across several thousand projects, with something like €3.4 billion pointed at AI and quantum. This is the layer my own employer spends the most breath explaining to companies, because it is the one with real money and real factories attached.
And then the reason all of this suddenly looked like a strategy instead of a subsidy: Mistral. Arthur Mensch, Guillaume Lample and Timothée Lacroix raised $113 million in June 2023, one month after founding the company and before they had a product. It was the largest seed round in European history, valuing a slide deck at $260 million. What happened next is the part I still find slightly hard to believe. A Series A that December put the company past €2 billion. By the following June the Series B had it at €5.8 billion, and in September 2025 a €1.7 billion Series C arrived led, of all people, by ASML, the Dutch lithography monopoly, which took an eleven percent stake and became the largest shareholder. By March of this year Mistral was raising $830 million in debt just to build data centers, and as I write this it is reportedly in talks for roughly €3 billion more at a €20 billion valuation.
I want to keep that number in perspective, because the French AI press mostly won’t. Twenty billion euros is enormous for a three-year-old European company and rounding error next to the American frontier labs, which sit in the high hundreds of billions and have raised on a scale Mistral will never match. Its whole pitch is that it doesn’t need to. The models are open-weight, so a German bank can run them on its own servers and keep the data in-house, which it cannot do with a closed American model. Mistral signed France’s army, Luxembourg’s government and a wall of European industrials on exactly that sovereignty argument, and it is spending its debt raises building its own compute in France and Sweden rather than renting from AWS. That reframes the company from an AI lab into a bet that Europe will pay a premium to not depend on American infrastructure.
None of this comes from nowhere, and here the boring French institutions earn their keep. Polytechnique alone accounts for something like 15% of the country’s AI founders, and INRIA, the public computer science institute, sits at the very top of the list of organizations that produce them. That depth is why Meta put its AI research lab in Paris in 2015 and DeepMind followed in 2018. For years those labs were the problem, because they were where French talent went to vanish into American org charts. What flipped is that the same people started leaving the American labs to found things at home. Mistral’s founders walked straight out of DeepMind and Meta. Even LeCun, the patron saint of the brain drain, left Meta at the end of 2025 to start his own outfit, AMI Labs, and headquartered it in Paris. When even LeCun comes home, you know the current has reversed.
The victory lap was February 2025, when France hosted the AI Action Summit at the Grand Palais and Macron walked in with €109 billion of private investment pledges, which he cheerfully billed as France’s answer to America’s Stargate. Some of that is a UAE fund promising 30 to 50 billion for a data-center campus; another chunk is Brookfield; and the rest is French telcos and Thales. How much actually lands is an open question I won’t pretend to answer here. But the summit did something the clusters never could: for two days it made France look like a place where the frontier gets built rather than a place that trains the people who go build it elsewhere.
I’m skipping the energy-grid pilots, which the tidy version of this story always includes and which never amount to what the slides promise. The mobility version deserves one hard look, though, because it is the counterexample that keeps me honest. Navya was the French autonomous-shuttle darling, founded in Lyon in 2014, running an early prototype around a nuclear plant, floated on the Paris exchange in 2018 at roughly a €190 million valuation. Its stock ended at three cents. It went into receivership in early 2023 and got sold for parts. Every layer I just described was available to Navya, and the layers did not save it, because the technology wasn’t ready and the market didn’t exist. The stack is not magic. It lowers the cost of trying without doing a thing to guarantee the trying works.
So here is what I actually tell an American company looking at France, which is more or less my day job. The layered structure is real, and its honest advantage is that you don’t have to build the whole stack yourself: the cluster, the campus, the public co-investor and the training pipeline already exist, and you can plug into them instead of reproducing them. That beats an ecosystem where one strong VC scene sits on top of nothing. But I would be lying if I sold it as bulletproof. The capital is still thin next to the US, and the best schools still feed the American labs as hard as they feed the French ones. And the whole “loop finally closed” thesis is currently resting on the shoulders of exactly one company. If Mistral stumbles, the story I just told becomes much more fragile. Ask me again after the next funding round.