One gigawatt. That is what Mistral now sells.
On August 11, the company announced a three-part plan: regional endpoints to choose where inference runs (Europe or the United States), a "Priority Tier" with an uptime guarantee for mission-critical deployments, and above all "European Compute Units." Five-year purchase commitments, converted into compute rights, meant to underwrite 200 megawatts of infrastructure by the end of 2027, then a full gigawatt by the end of 2030. Four European groups have already signed those multi-year commitments: Amadeus, ASML, CMA CGM and Capgemini.
The announcement in detail: "Mistral AI wants to build 1 gigawatt of European compute by 2030", Michael Nuñez, VentureBeat, August 11, 2026. Illustration: VentureBeat (Midjourney).
Look closely at how this gets financed
A one-gigawatt data center costs around $38 billion according to Epoch AI, and no European funding round covers that. So Mistral makes the people who will use the infrastructure pay for it up front: the customer subscribes for five years to the capacity it will then buy, a mechanism borrowed from energy contracts, not from cloud.
The system feeds the system, right down to the cap table: ASML, which took a €1.3 billion stake in Mistral in 2025, now signs as an anchor customer of the infrastructure it helped capitalize.
That leaves demand. Who manufactures it?
GDPR, the AI Act, data residency requirements, public buyers checking the "European" box. Fear works as marketing fuel: fear of the Americans, fear of the Chinese, fear of falling behind. Mistral would exist neither without the European companies funding it, nor without the European laws that make its offer necessary in practice.
Brussels creates the constraint, the constraint creates the market, the market funds the infrastructure, and the infrastructure becomes the argument for the next constraint.
Servers rented from Redmond, a model trained in Beijing
Still, on that same "sovereign" platform: Microsoft, a multi-billion partnership, rents capacity in Mistral's European data centers, which run on Nvidia chips. And on Tuesday, Mistral announced it would host GLM-5.2, the model from Chinese lab Z.ai, because it performs well and it is open source. So protection against foreign influence ships with servers rented from Redmond and a model trained in Beijing.
There is coherence here. It is commercial, not geopolitical.
The question is not settled by the data center's zip code
A Chinese model hosted in Essonne is still a Chinese model: the weights carry their training, the texts selected, the answers rewarded, the silences learned. Your teams will think with it, your customers will consume what comes out of it. The influence question is not settled by the data center's zip code, it is settled by what shapes the reasoning of the people using it every day.
They sell you data residency; cognitive residency appears in no contract.
I hold nothing against Mistral
It is well played, in fact: going after value where the market makes sense, on compute supply, a scarce commodity, and on European conservatism, an inexhaustible one. A seller sells, and this one sells very well.
What I take issue with is the protection narrative wrapped around the deal, because on real security as much as on the end customer's interest, nothing is less certain.
Who funds, who hosts, who trained
At Catalia, we hand the executives we work with a different grid. Before signing a "sovereign" compute contract, ask what you are actually buying. Capacity, an SLA, a jurisdiction? Fine, that is real, that is useful, negotiate the price. Independence? It does not ship in megawatts.
Three questions are enough to see clearly: who funds, who hosts, who trained. The answers tell you whether you are buying sovereignty or subsidizing it.
At Catalia, we don't teach you to produce with AI. We teach you to think with it.
Want to go further on this topic?
Let's talk about your needs.
