Mistral is selling multi-year reservations on compute it has not built yet

Mistral said on August 11 that ASML, Amadeus, Capgemini, Caisse des Depots and CMA CGM have made multi-year commitments to buy compute capacity from its planned European infrastructure ^[4]. The operative word is planned: the contracts reserve future capacity through a mechanism Mistral calls European Compute Units, and reporting on the announcement stresses that much of what is being reserved has not been built ^[1][7].

That makes this a demand-underwriting exercise rather than a procurement round. Mistral frames the agreements as a way to secure long-term demand while it expands the physical capacity behind its services ^[2], and describes the group as a coalition assembled because compute is strategic yet scarce and fragmented ^[3]. Five large European organisations agreeing to pay for capacity before it exists is what construction finance normally requires, and it is the part of the announcement with real consequence: the commitments are contractual promises to use infrastructure as it becomes available, not evidence that the buildout is running ^[8].

The shape of the curve is worth reading closely. Mistral targets roughly 200 MW by the end of 2027 and up to 1 GW by 2030 ^[5], which means about 800 MW, five times the 2027 base, has to arrive in the following three years or so ^[21]. Delivery depends on data-centre shells, grid power and hardware landing on schedule ^[13], and none of those are things a customer commitment accelerates.

Neither Mistral's own post nor the trade coverage discloses contract values, term lengths, price per European Compute Unit, or whether the commitments carry take-or-pay teeth ^[14]. Until those terms surface, the announcement is a signal about willingness, not a measurable liability on either side.

What is shippable today is narrower and more useful. Mistral Regional Endpoints are generally available, letting customers pin inference to Europe or the US at api.eu.mistral.ai or api.us.mistral.ai ^[9], at a 10 percent surcharge over standard pricing, according to The Decoder ^[15]. The same report notes that among add-on tools only function calling works at the regional endpoints, with agents, batch processing and file management absent ^[16], that model availability varies by region with no published list, so customers must query each endpoint ^[17], and that account settings, API keys, billing and usage statistics may still be processed outside the chosen region ^[10]. The Priority Tier is in public preview: a fast lane during congestion, priced at 1.75x standard and requiring a signed contract rather than self-service ^[11][18]. It carries a 99.5 percent uptime SLA, roughly three and a half hours of permitted downtime a month, where the standard tier has no guarantee at all ^[11]. Customers switch it on with a service_tier parameter, overflow above negotiated rate limits falls back to standard instead of failing, and the API response reports which tier actually served the request ^[19]. Prompt caching discounts, which can reach 90 percent, are calculated before the priority surcharge ^[20].

Mistral also says its platform will host third-party open models under the same regional controls and service commitments, beginning with Z.ai's GLM-5.2 ^[6], and claims to be the only European lab offering both regional processing choice and an SLA-backed service level ^[22]. The New Stack reads the combination as a move from model developer toward infrastructure and distribution ^[12]. Note that Mistral says most of its customers currently run its models in their own data centres and clouds ^[23], so the reservations ask those buyers to shift part of that dependency onto capacity Mistral operates.

Watch three things. Whether the stateful features reach the regional endpoints, since without agents, batch and file handling the sovereignty claim covers the compute step only ^[16][10]. Whether the Priority Tier exits preview with SLA credits attached to the 99.5 percent number ^[11]. And whether the end-2027 200 MW milestone is met on time, because the 2030 figure is arithmetically dependent on it ^[5][21].