Summary

  • Official company and routing records identify Mistral Compute Holding SAS and AS209952. The earlier account of a missing network footprint was wrong.
  • French GPU financing and a Swedish infrastructure partnership reveal different roles and delivery stages. Neither a route nor a funding announcement proves that every promised unit of capacity is available.

AS209952 is an unglamorous place to begin a story about European artificial intelligence. It is a network number, not a model launch or a billion-euro promise. Yet it answers a question that this article previously left hanging: Mistral's compute business does have a visible network trail.

The RIPE registration names the network Mistral-Compute and identifies Mistral Compute Holding SAS as its organisation registrant. Routing observations show six address prefixes announced by that network over 11–25 October 2026, a period that includes this article's original publication. This is not simply a new announcement arriving after the fact. The earlier claim that no network footprint could be found needs correcting.

The correction makes the business more interesting, not less. A route can tell other networks where to send packets. It cannot tell a customer who owns the accelerator that will process them, whether a reserved cluster is ready, or who must put things right when it is not.

The trail does not stop at a search page

There is a corporate identity to follow, too. A query to the French government's company service returns an active company, MISTRAL COMPUTE HOLDING, under SIREN 993225341, with MISTRAL AI listed as its corporate president. A registration page that failed to render was never a sound basis for treating that identity as unconfirmed. The record establishes the company and an officer relationship; it is not an inventory of assets or a complete ownership chart.

PeeringDB's specific network entry adds a listed connection at nineIX. It is an operator-maintained record, rather than an independent measurement of traffic, but it reinforces the point: the network is named, registered and visible. The absence of a result in a broad name search was the wrong centre of gravity for the story.

Mistral itself describes a business already serving workloads. Its cloud timeline places GB200 production service in February 2026 and the first external customers in March. That is the company's operating account, not an inspection of its machines. It nevertheless goes well beyond the existence of a product name in a news index. The June 2025 launch set out an offering spanning GPU infrastructure and managed services, developed with NVIDIA and other partners.

Two places, different stages

The physical locations are not nameless, either. Bruyères-le-Châtel, south of Paris, appears in a transaction announcement by Gide, counsel to Mistral AI. On 30 March 2026, the firm described roughly $830 million in financing to support the acquisition of 13,800 NVIDIA GB300 GPUs and the development of compute capacity at that French site. Financing a purchase and accepting delivery are separate events, but the announcement identifies considerably more than a country and an aspiration.

In Sweden, the arrangement is different. EcoDataCenter's February announcement names Borlänge and a €1.2 billion long-term partnership, with operations scheduled for 2027. EcoDataCenter is to design, build and operate the underlying infrastructure; Mistral is to deploy the compute. The division of work is explicit. Calling that entire future facility wholly owned and operated by one Mistral entity would erase the partner doing the physical work.

These numbers describe different transactions. The French debt package is not evidence that the same loan is building the Swedish centre. Nor can either commitment be added to Mistral's €3 billion Series D and presented as cash already spent on installed capacity. The equity announcement covers research, compute, infrastructure and commercial expansion. Borrowing, raising equity and buying a usable hour of GPU time belong to different accounts.

Microsoft is on the other side of the counter

One customer complicates the familiar story of a European company seeking independence from large cloud providers. Microsoft's July announcement describes a multibillion-dollar commitment to use part of Mistral's expanded European GPU capacity. Here the hyperscaler is buying from the model company, not merely renting machines to it.

That can change the bargaining relationship. It does not abolish dependence. Control over a compute platform can coexist with commercial reliance on a large buyer. The public announcement does not disclose enough reservation, payment or priority terms to calculate that reliance, and it does not make all of the expansion available today.

For another customer, the consequential questions are concrete. Which legal entity promises the capacity? What hardware, location and delivery date does that promise cover? If a building operator, electricity supplier or equipment provider misses a milestone, who still owes the customer a service—and what remedy follows? A reassuring group name is not a substitute for those answers.

Network observations help with one part of the investigation. They do not count GPUs, measure their utilisation or establish title to a data centre. A provider can also own equipment while using someone else's network. Treating an ASN as a certificate of infrastructure independence would repeat the earlier mistake in reverse.

Mistral's cloud roadmap reaches toward one gigawatt by 2030. Borlänge is still a 2027 project. Those targets belong beside, not in place of, the services the company says it already operates. The useful distinction now is between available capacity, contracted expansion and future ambition. Waiting for Mistral's first trace on the internet would miss the story: that trace is already there.

Updated 25 October 2026: this article corrects its earlier statements about missing company and network records, identifies the named French and Swedish sites, and replaces the earlier financing account with primary sources. Network visibility is distinguished from asset ownership and delivered compute capacity.