Mistral sold five years of compute it has not built yet, and used the money to ship endpoints

On 11 August Mistral made regional inference endpoints generally available, letting customers choose Europe or the US, and put a Priority Tier into public preview with custom rate limits and an uptime commitment ^[1][2]. The third item in the same announcement is what pays for the first two: enterprises hand over money now in exchange for multi-year access to compute Mistral will build later ^[3].

Mistral calls the instrument a European Compute Unit, spendable across its compute products once capacity comes online ^[4]. Chief technology officer Timothee Lacroix told European Business Magazine that partners commit for around five years with no early exit ^[5]. That is a clean trade with an uneven distribution of risk. Mistral gets a demand signal it can finance against; the buyer holds delivery risk on halls that are mostly still drawings.

Sifted reported the anchor group as five companies: ASML in the Netherlands, Amadeus in Spain, Capgemini and Caisse des Depots in France, and the shipping group CMA CGM ^[6]. Each is represented by its chief executive in the announcement ^[7]. Caisse des Depots is a French state financial institution, so one of the five anchor customers is the French state ^[8]. ASML sits on both sides of the table: it led Mistral's 1.7bn euro round in September 2025 at an 11.7bn euro valuation, and has now committed to buy the compute that round helps build ^[9][10]. ASML chief executive Christophe Fouquet said Mistral "is taking on that challenge with the scale, ambition, and staying power" ^[11]. An anchor order from an investor tells you less about the open market than an order from a stranger.

The scale gap is the whole story. Mistral targets up to 1GW by 2030, with about 200MW by the end of 2027 ^[12]. Arthur Mensch has previously put a gigawatt at roughly $50bn of investment ^[13]. The first facility runs to 44MW, sits south of Paris and carries an $830m loan ^[14]: 4.4 percent of the 2030 target ^[15], against an implied $2.2bn of spend at Mensch's own rate ^[16]. There is other iron already in the ground, including a 40MW GPU cluster in the Paris region at an Eclairion facility hosted by Scaleway, per DatacenterDynamics ^[17]. A 1.4GW campus with Bpifrance, the UAE fund MGX and Nvidia has been discussed for the same region, possibly starting 2028 ^[18]. Microsoft made a separate commitment three weeks earlier funding Nvidia Vera Rubin chips for the European capacity, structured to avoid an equity stake ^[19].

The first third-party open model on the same infrastructure, under the same regional controls, is GLM-5.2 from the Beijing lab Z.ai, with a one-million-token context window at $1.40 per million input tokens ^[20][21]. Sovereignty here means jurisdiction over where the weights run, not where they came from ^[22]. Mistral will run Nvidia silicon in European buildings under European law ^[23], which answers the standing objection that renting GPU capacity reinforces the illusion of sovereignty while the chips stay American ^[24] only in part.

Watch whether the 200MW milestone lands by the end of 2027 ^[12], since the five-year lock-ins signed this August mature around 2030 ^[5]. Watch whether any anchor buyer with no equity in Mistral joins the group. And watch how many more third-party models arrive on the endpoints, because the pitch is that the infrastructure layer outlasts any single model ^[25].