France Is Building AI Sovereignty in Angers and Dismantling It in Parliament
Two pieces of news landed five days apart, and nobody read them together.
On November 3, Schwarz Digits, the digital arm of the group that owns Lidl, signed a memorandum of understanding with Bull to build a European AI infrastructure offering: the STACKIT cloud on the German side, liquid-cooled servers from the Angers factory on the French side, as Reuters reports.
On November 8, Arthur Mensch published a piece titled “Where will we train the next generation of our models?”, in which the head of Mistral asks French lawmakers to reject the Darcos bill, scheduled for debate on November 19, on the grounds that it will weigh on the decision to build his next data centers in the Essonne department or elsewhere.
The first story is about how you build European computing capacity. The second is about how you drive it away. They are the same story.
A discount grocer and a supercomputer maker
The Schwarz-Bull partnership is heartening precisely because there is nothing glamorous about it. A supermarket group that raised 5.6 billion euros for a data center in Germany and, like Amazon before it, turned its internal infrastructure into a commercial offering. A French manufacturer that has just doubled the capacity of its Angers plant after 80 million euros of construction work, the only supercomputer factory in Europe, and one that knows how to build liquid-cooled servers, something almost nobody else on the continent can do. Together they plan to design servers on the Open Compute Project’s open standards and to bring the BullSequana AI software platform into the STACKIT marketplace.
It helps to remember where Bull is coming from. Swallowed by Atos in 2014 and rebranded Eviden in 2022, the business lived through its parent company’s collapse, with crushing debt, one restructuring after another, and a cratering share price, and the Angers plant never stopped. The French state eventually bought Bull back from Atos for an enterprise value of up to 404 million euros, a deal closed on March 31, 2026. Since April, Bull has been wholly owned by the state’s shareholding agency, and the historic brand has been resurrected. In other words, the industrial capacity now signing with Schwarz outlived its shareholder, and the French Republic now holds the keys. Keep that in mind; it will matter later.
No product, no date, and the Clubic article says so plainly: this will take years. So I’m not going to get worked up over a memorandum of understanding. But the nature of the alliance matters more than its timeline. For once, this is not a “trusted cloud” label slapped onto American technology. It’s machines, a factory, and an operator that already runs its own data centers. I wrote in August that sovereignty is an industrial bottleneck, not a line in a portfolio. Angers is a bottleneck. STACKIT is another. That is what sovereignty is made of, when it exists at all.
What the Darcos bill actually says
The bill fits in a single article. It inserts a new article L. 331-4-1 into the French intellectual property code: unless proven otherwise, a protected work is presumed to have been used by an AI system whenever some indication makes that use plausible. The bill’s authors cite as possible indications a technical expert assessment, the regurgitation of excerpts, or output “in the style of” an identifiable author. The Conseil d’État, consulted by the president of the Senate, clarified that this is a presumption of use, not a presumption of infringement: a judge would still have to establish the infringement itself. The Senate passed it unanimously on April 8. In June, the bill bogged down in the National Assembly under a mountain of amendments from the Ensemble pour la République group. It returns on November 19, sixth on the agenda of the Socialist group’s reserved day, while Mistral and the digital industry step up their lobbying.
I’m not going to pretend I’m discovering this subject. I wrote a year ago that large language models are the ultimate copyright laundering machine, and I haven’t changed my mind. A model ingests works it never paid for and returns a “clean” version whose origin nobody can trace. The presumption of use answers exactly that problem: since it’s impossible for an author to prove ingestion, you ask the party holding the corpus to prove the opposite. On principle, the senators are right.
Where Mensch is right
He’s right about the mechanism. Proving that one specific work was not used to train a trillion-parameter model fed on billions of words means proving a negative. The only way to do it is to open up the corpus, the filters, the deduplication procedures, which is to say most of what makes a lab a lab. The Conseil d’État did note that trade secrets should be respected, but an order to produce evidence that only secrecy protects is called a dilemma, not a guarantee.
He’s also right about the geographic asymmetry. A presumption that applies only in France hits those who train in France first. OpenAI and Anthropic have no data center in the Essonne. Mistral does: Large 4 was trained in Bruyères-le-Châtel, on its own hardware, and that’s the first time anyone has been able to say that about a model of this caliber. Mensch has announced a gigawatt of capacity to be built over the coming years, “mostly in Europe and largely in France.” “Largely” is the word that carries the weight. It means the rest is negotiable.
And this is where the two stories converge. Schwarz is German. STACKIT is German. The group’s 5.6 billion went into a German data center. If training a model carries legal risk in France and not across the Rhine, Mistral’s gigawatts will find a home without crossing the Atlantic. European capacity will exist. It just won’t be French. We’ll have built the servers in Angers to run them somewhere else.
We should be precise about what the problem is. It isn’t regulation as such: an enforceable copyright regime across the entire single market would scare nobody off, because there would be nowhere in Europe to run to. The problem is a national, uncoordinated rule on a subject where compute is the most mobile asset there is. A data center is decided in a quarter and amortized over ten years. A French law that runs for three years before being superseded by the text Brussels is preparing under Article 53 of the AI Act means three years during which siting decisions get made elsewhere. That isn’t a bluff. It’s arithmetic.
One detail you couldn’t make up: Laure Darcos is the senator for the Essonne. The department where Bruyères-le-Châtel sits.
And here is the point I asked you to keep in mind. The French state has just spent up to 404 million euros to own the factory that builds AI servers. That same state’s parliament is about to vote on the rule that will make those servers riskier to run in France than in Germany. The state shareholding agency buys the hardware; the culture committee drives away the workload. Two arms of the same body, and neither knows what the other is doing.
Where you have to read between the lines
Mensch isn’t writing as a worried citizen. He’s writing as the head of a company whose biggest customer is the French state, and which turned sovereignty into a sales pitch before reducing it to a point of access. The threat of relocation is a classic negotiating lever, and it should be read as one. Laure Darcos, for her part, fired back the same day with an open letter accusing Mistral of retreating from generative AI to refocus on cloud and data centers, and of never having signed a single license with the cultural sector. The attack is clumsy, since a data center is hardly a retreat, and the post earned a community note on X pointing that out, but it says something about the mood.
His counterproposal deserves more attention than it’s getting: a levy on the revenue of every AI provider operating commercially in Europe, Americans and Chinese included, feeding a fund for creative work. What’s piquant is that the Senate rapporteurs themselves had floated a revenue tax as their plan B if the bill failed in parliament, and that on July 1, Renaissance deputy Céline Calvez delivered a parliamentary report that likewise proposes a mandatory contribution from AI providers above a revenue threshold, regardless of their country of origin. Everyone has the same plan B. The disagreement is over plan A.
But not all contributions are created equal. The Calvez report is careful to state that its contribution amounts to neither authorization nor immunity, that it leaves authors’ rights fully intact, and that a provider who signs licenses is exempted pro rata: it’s an incentive to negotiate. Mensch’s version feeds a fund for creative work, full stop. That isn’t copyright; it’s a compulsory license in disguise. You pay a flat fee, you ingest everything, and you never have to ask anyone for anything again. For an author, it amounts to making the laundering official in exchange for a royalty stream distributed by a collecting society. Mensch presents this as the path where “regulation protects.” What it mostly protects is Mistral from the burden of proof. It pays, but it doesn’t authorize, and that is precisely the distinction rights holders refuse to give up.
What this affair says about us
There’s a French constant in this file: we know how to vote for principles, and we don’t know how to build the conditions for applying them. The Senate unanimously passed a single article on a subject that deserved thirty pages, and unanimity on a short text is rarely a sign that the hard questions have been settled. The Assembly responded with a hundred-odd amendments rather than a counterproposal, which is no more glorious. In between sits an executive that wants both its AI champion and its cultural industries, and has refused to arbitrate between them. Mistral threatens to leave. And meanwhile, a German grocer and a manufacturer from Angers sign, without a press conference, the kind of agreement that will actually decide who owns compute in Europe ten years from now.
Mensch has one line I’ll grant him without reservation: if Europe stops producing this technology, American and Chinese players will dictate their terms entirely. I made the same argument when Jensen Huang made the case for open models. A continent that no longer trains anything has nothing left to regulate; all it can do is accept terms of service written in San Francisco or Hangzhou.
But the reverse is just as true, and Mensch doesn’t say it: a continent that trains its models on the unpaid work of its own creators hasn’t won its sovereignty. It has taken it from someone else.
What I’ll be watching on November 19
The Darcos bill will probably be buried under amendments once again, and everyone will see in that the proof of what they already believed. I’ll be watching something else. I’ll be watching whether anyone in the Assembly tries to hold both ends of the rope: a presumption that forces providers to document their corpora, paired with an evidentiary regime that doesn’t force them to publish their trade secrets, and a timeline aligned with what Brussels is preparing under Article 53 of the AI Act, so that France isn’t carrying the risk alone for three years.
And in the months ahead, I’ll be watching where Mistral puts its next gigawatt. If it’s in the Essonne, the bill will have been aligned with Europe or buried for good. If it’s somewhere else, Mensch will have been telling the truth, and we’ll have built our servers in Angers to plug them in abroad. Either way, we’ll find out what sovereignty is worth when you ask it to move from the speech to the power outlet.