Mistral set out the plan on 11 August. Its own announcement covers three things at once.

Regional endpoints are now generally available, letting customers pick Europe or the US for inference. A Priority Tier in public preview adds custom rate limits and an uptime commitment.

The third part is the one that funds the other two.

What a European Compute Unit is

The instrument is a forward sale. Enterprises commit money now, and it converts into multi-year access to compute Mistral will build later.

Mistral calls these European Compute Units. Customers can spend them across its compute products once capacity comes online.

The terms are long. Partners commit for around five years with no early exit, chief technology officer Timothée Lacroix told European Business Magazine.

That is the trade. Mistral gets a demand signal it can finance against, and the buyer accepts delivery risk on infrastructure that is mostly still drawings.

Who signed

Five companies form what Mistral calls the anchor group, Sifted reported. ASML in the Netherlands, Amadeus in Spain, Capgemini and Caisse des Dépôts in France, and the shipping group CMA CGM.

Each is represented by its chief executive in the announcement. Christophe Fouquet for ASML, Luis Maroto for Amadeus, Aiman Ezzat for Capgemini, Olivier Sichel for Caisse des Dépôts and Rodolphe Saadé for CMA CGM.

Fouquet framed it as backing scale. Mistral “is taking on that challenge with the scale, ambition, and staying power”, he said.

Caisse des Dépôts is worth naming twice. It is a French state financial institution, so one of the five anchor customers is the state itself.

ASML is on both sides of the table

The Dutch company is not only a customer. ASML led Mistral’s €1.7bn round in September 2025, which valued the company at €11.7bn.

Now it has committed to buy the compute that round helps build. Both facts come from the companies themselves, and neither is hidden.

It does change how to read the demand signal. An anchor order from an investor tells you less about the open market than an order from a stranger.

Microsoft made a separate commitment three weeks earlier. That deal funds Nvidia Vera Rubin chips for Mistral’s European capacity and deliberately avoids an equity stake.

The buildout behind the promise

The target is up to 1GW by 2030, with about 200MW by the end of 2027. Arthur Mensch has previously put a gigawatt of compute at roughly $50bn of investment.

The first facility is smaller than the headline. It runs to 44MW, sits south of Paris, and carries an $830m loan.

The gap between 44MW and 1,000MW is the project. Everything the anchor group has bought sits on the far side of it.

Mistral has other iron in the ground already, per DatacenterDynamics. A 40MW GPU cluster in the Paris region runs at an Eclairion facility hosted by Scaleway.

A larger campus is a joint venture. Mistral, Bpifrance, the UAE fund MGX and Nvidia have discussed a 1.4GW site in the same region, with a possible 2028 start.

The first outside model it will run is Chinese

Mistral will host third-party open models on the same infrastructure, under the same regional controls. The first is GLM-5.2, from the Beijing lab Z.ai, also known as Zhipu.

The model has a one-million-token context window and costs $1.40 per million input tokens. Z.ai released GLM-5.2 as open weights with no usage restrictions.

Mistral’s pitch is that the infrastructure layer matters more than any single model, as The New Stack put it. Customers get one place to run open models without starting over each time they switch.

Matan Grinberg, chief executive of Factory, supplied the customer line. Mistral “allows us to run open models under strict regional controls and service commitments”, he said.

Sovereignty here means jurisdiction over where the weights run, not where they came from. Those are different claims, and the announcement makes the first one.

The argument this walks into

Europe has been having a version of this debate for two years. Mistral itself spent them warning that American providers could switch European customers off, a case the desk covered as its sovereignty moment.

The counterargument is about the layer below. Renting GPU capacity reinforces the illusion of sovereignty while the chips stay American, one contributor argued here in May.

Owning the buildings answers part of that and not all of it. Mistral will run Nvidia silicon in European halls under European law.

The physical constraints are real too. Research this summer put power availability, planning delays, build costs and skills shortages in the way of Europe’s data centre plans.

Those constraints apply to Mistral exactly as they apply to everyone else. A gigawatt by 2030 needs grid connections that are already scarce.

The company is also raising again. Reports put it in talks for about €3bn at a valuation near €20bn.

That round and this coalition do the same job from different directions. One brings equity, the other brings contracted demand, and both are needed before a spade goes in the ground.

What would settle it

Three things, and the first is disclosure. Mistral has not published the capacity, pricing or delivery dates attached to any European Compute Unit.

Without those, the size of the commitments cannot be read. Five chief executives on a page is a strong signal and not a number.

The second is the 2027 milestone. About 200MW by the end of next year is checkable, and it is the first point at which the forward sale either has product behind it or does not.

The third is who signs next. Five anchor customers, one of them a lead investor and one of them the French state, is a start rather than a market.

Mistral has sold the idea of European compute before it has the compute. The interesting part is that five of Europe’s largest companies were willing to pay for it anyway.

Get the TNW newsletter

Get the most important tech news in your inbox each week.