Microsoft has agreed to spend billions of dollars on Mistral's computing infrastructure in Europe. Microsoft will use Mistral's data centers to serve its own cloud and AI customers, while Mistral's models get embedded into Microsoft's tools for building business software. The exact dollar figure has not been disclosed, but both companies confirmed it runs into the billions.
This is not a new relationship. Microsoft made a small equity stake in Mistral back in 2024. What changed now is the scale: this deal plugs Mistral's European servers directly into Microsoft's commercial cloud capacity, and it puts Mistral's models in front of every company using Microsoft's software-building tools.
The deal is tied to Microsoft's broader pledge, announced in April 2025, to expand its European data center footprint by 40% over two years, adding operations across 16 countries and reaching more than 200 data centers on the continent by 2027. The Mistral deal counts as part of that commitment. Microsoft is essentially renting capacity from Mistral to help meet European demand faster than it could build alone.
For Mistral, the arrangement is distribution at scale. The company's CEO, Arthur Mensch, has been vocal about a specific concern: companies that send their data to closed AI systems run by foreign companies risk exposing trade secrets and proprietary information. By running inside Microsoft's infrastructure tools, Mistral's models can now be deployed entirely within a company's own servers, with no data leaving their environment at all.
This matters most to a specific type of buyer. European banks, hospitals, defense contractors, and government agencies face strict rules about where their data can sit and who can access it. A 2025 survey found that roughly 72% of European enterprise IT decision-makers cited data control as a primary or secondary factor when choosing a cloud vendor. Mistral already won a framework agreement with France's armed forces and has partnered with SAP and the French and German governments to build AI tools for public administrations. The Microsoft deal extends that reach globally.
Mistral's numbers back up the momentum. Founded in 2023, the company grew revenue roughly twentyfold to over $400 million annually. It raised 1.7 billion euros in a Series C round led by ASML in September 2025, and in March 2026 raised $830 million in debt specifically to build a data center near Paris housing nearly 14,000 Nvidia chips. The company has also invested 1.2 billion euros in a separate data center in Sweden. It is reportedly now in talks to raise another 3 billion euros at a valuation of around 20 billion euros.
There is a real tension worth watching here. Mistral's sovereignty pitch rests on being European-controlled infrastructure. But the company trains largely on US chips, distributes through Azure, AWS, and Google Cloud, and counts US venture firms among its largest backers. As one analysis put it, the sovereignty story holds where Mistral can demonstrate clean jurisdictional control end to end. A French legal address does not automatically satisfy every regulatory audit.
Still, the momentum is real and the commercial logic is clear. European companies with serious data compliance requirements need AI tools that fit within their legal constraints. American providers have struggled to offer that cleanly without complicated workarounds. Mistral's open models, which companies can download and run on their own servers, combined with Microsoft's distribution network, give procurement teams a straightforward answer to a question they are increasingly asked: who controls the infrastructure?
For operators using Microsoft tools today, the immediate practical change is simple: Mistral's models, including its latest Medium 3.5 and document-reading model OCR 4, are now accessible inside Copilot Studio, the tool many organizations use to build internal AI assistants. They can be used in the cloud, in a connected private environment, or fully disconnected. Regulated industries looking for AI they can actually deploy within their compliance constraints now have a more direct path to doing so.