Summary

  • The 18,000 NVIDIA GB200 figure repeated as "already operational" traces to a superseded 2025 procurement plan, not to the fleet Mistral financed in March 2026.
  • The March 2026 debt financing covers 13,800 NVIDIA GB300 GPUs and 44 MW of powered capacity at Bruyères-le-Châtel, south of Paris, with operations intended by end June 2026.
  • No public document inspected records a dated energisation event, metered load, or grid-acceptance notice for that 44 MW fleet.
  • Mistral's commercial layer — Regional Endpoints, Priority Tier and European Compute Units — now sells forward commitments against a physical base documented only indirectly.

A single sentence has done more to obscure Mistral Compute's actual infrastructure status than any megawatt figure. In a blog post dated 17 June 2026, NVIDIA wrote that Mistral "is building a new 44-megawatt data center in Bruyères-le-Châtel" and, in the same paragraph, that its "first deployment is already operational with 18,000 NVIDIA GB200 systems" (NVIDIA blog). Both statements cannot carry the weight the sentence implies. The building and the fleet are described in contradictory states, the GPU count belongs to a plan that was replaced, and no energisation evidence accompanies the claim.

Two fleets, one site

The resolution of the fleet conflict sits in the financing record. On 30 March 2026, Mistral raised $830 million in new debt — its first debt raising — to buy 13,800 Nvidia chips for a data centre near Paris (Reuters). The financing provides access to 13,800 NVIDIA GB300 GPUs and 44 MW of powered capacity at Bruyères-le-Châtel, at a site owned and run by Eclairion, with Mistral intending to start operations there by the end of June 2026 (Data Center Dynamics). Seven banks supported the transaction: Bpifrance, BNP Paribas, Crédit Agricole CIB, HSBC, La Banque Postale, MUFG and Natixis CIB (CNBC). French reporting put the equivalent at roughly €722 million (ZDNet France); French technology reporting also traced Mistral's pivot toward AI infrastructure and industrial applications (Le Monde Informatique).

The earlier plan this financing replaced is the key to the conflict. In 2025, Mistral had said it would procure 18,000 NVIDIA GB200 GPUs from Scaleway, with the data centre opening in summer 2025 at 40 MW — a plan announced in partnership with Eclairion and Fluidstack, the deployment of a high-density AI supercomputer in France for Mistral AI (HPCwire). Data Center Dynamics records explicitly that the project was delayed and that Mistral "has now opted for the newer generation of Nvidia hardware instead" — identifying the 18,000-GB200 figure as a superseded plan and the 13,800-GB300 figure as its replacement (Data Center Dynamics).

So the two numbers are not two fleets. They are successive planning generations at one site: 18,000 GB200 / 40 MW / summer 2025, then 13,800 GB300 / 44 MW / end-June 2026. NVIDIA's June 2026 blog resurrected the first generation and labelled it operational.

What the NVIDIA post actually records

The NVIDIA paragraph contains three defects worth naming precisely. First, it describes Mistral as "building a new" data center while asserting the deployment is "already operational" — a contradiction internal to one sentence. Second, its GPU count and generation match the superseded 2025 plan rather than the March 2026 financing. Third, it locates Bruyères-le-Châtel in "northern France" when the commune lies in Essonne, south of Paris, as Data Center Dynamics and other reporting place it (Data Center Dynamics).

NVIDIA is a commercial party to the project, not a neutral operator confirmation. The post cites no energisation event, meter reading, or grid-acceptance notice — and none of those documents exists anywhere in the inspected public record.

The registry that contradicts itself

Third-party trackers have absorbed the contradiction rather than resolved it. The DEPLOY registry facility page for the Bruyères-le-Châtel campus records Energized 0.04 GW as of 1 June 2026 at "reported" evidence tier — secondary reporting, not primary confirmation — while also listing Announced 1.40 GW as of 19 May 2025 at "claimed" tier, and Energized 0 GW as of the same 2025 date (DEPLOY registry). The same page's FAQ states the campus "is announced but not yet built," and its index table lists status "announced" with energized capacity of zero — while carrying, as a source-tagged chip claim, the assertion that 18,000 NVIDIA GB200 systems are operational at the node (DEPLOY registry). A facility record on Scrutica shows the same pattern: Status "Operational," 44 MW, 13,800 GPUs, owner/operator Eclairion, announced 30 March 2026, operational date "not published" (Scrutica).

These are aggregator artifacts, not evidence. Their internal inconsistency — operational and announced at once, energized and not built — mirrors the upstream contradiction rather than measuring anything (DEPLOY registry index).

What Mistral itself documents

Mistral's own materials are more modest than the claims propagated around them. The launch post for Mistral Compute, dated 11 June 2025, describes an integrated stack from bare-metal to managed PaaS with "availability of tens of thousands of GPUs" but names no hardware generation, unit count, megawatt figure or operational date for Bruyères-le-Châtel (Mistral AI).

The company's product-page timeline is the closest thing to dated operating evidence: GB200 racks landing July 2025, GB300 arriving December 2025, an alpha stack live on GB200, and GB200 "serving production" by February 2026, with a Sweden site run with EcoDataCenter "in motion" (Mistral AI). That timeline supports mixed-generation operation in principle — but it attaches no unit count or megawatt figure to any milestone, so it cannot verify "18,000" as a count, settle 40 MW against 44 MW, or constitute a dated energisation of the financed fleet.

The commercial layer outrunning the record

On 11 August 2026, Mistral announced general availability of Regional Endpoints letting customers choose EU or US inference, a public-preview Priority Tier with committed service levels and an uptime SLA, and support for third-party open models starting with Z.ai's GLM-5.2 (Mistral AI). The same announcement introduced European Compute Units — multi-year enterprise commitments converting into access to Mistral-built infrastructure — and an anchor coalition targeting 200 MW by end-2027 and up to 1 GW by 2030. Reporting on the announcement describes the ECUs as roughly five-year claims with no early exit, across three near-term sites: 44 MW near Paris, 23 MW in Sweden with EcoDataCenter, and 10 MW in Les Ulis (VentureBeat).

The Les Ulis 10 MW inference facility was scheduled to open in the third quarter of 2026 per Mistral's own AI Now Summit page of 28 May 2026 (Mistral AI); no public confirmation of its opening was found in the research for this report. Data Center Knowledge framed the March debt raise as part of Mistral's broader data-center play (Data Center Knowledge).

The pattern is structural, not incidental. The commercial layer sells capacity commitments — Regional Endpoints, SLAs, five-year compute units — whose physical base is documented only by a product-page timeline and vendor blog prose. A buyer of a five-year European Compute Unit is underwriting, in part, a 44 MW fleet whose energisation has never been independently dated.

What the record supports, and what it does not

The defensible statements, as of 1 October 2026, are these: Mistral financed 13,800 NVIDIA GB300 GPUs and 44 MW of powered capacity at Bruyères-le-Châtel with $830 million of debt on 30 March 2026, backed by seven named banks; the site is owned and run by Eclairion; operations were intended by end June 2026; Mistral's product timeline records GB200 systems serving production from February 2026; and the 18,000-GB200 figure belongs to a superseded 2025 plan that NVIDIA's blog repeated as "already operational" three months after the financing.

What the record does not support is any dated energisation event, metered load figure, or grid-acceptance notice for the 44 MW fleet — and therefore the claim that the financed fleet itself is operational. The distinction matters because every tier of Mistral Compute's commercial stack now prices against it.