Microsoft is building an AI stack it doesn’t fully own — on purpose

Microsoft doesn't own the data center, and that's the whole point. The company's expanded, multibillion-dollar partnership with Mistral, announced Tuesday, hands Mistral the job of running European compute infrastructure that Microsoft will tap into — outside its usual first-party facilities and leased space. Mistral plans to load that infrastructure with thousands of NVIDIA's next-generation Vera Rubin GPUs, hardware NVIDIA says can push agent throughput up to 10x higher than the prior Grace Blackwell generation, which matters a lot if you're running the kind of multi-step agentic workflows that sovereign, regulated environments increasingly demand.

The pitch here is really about control, not raw model quality. Brad Smith, Microsoft's Vice Chair and President, put it plainly: Europe should get the world's most capable AI without giving up control over its data, operations, or digital future. That line is aimed squarely at organizations spooked by the US CLOUD Act, even though whether a US-headquartered company can ever fully wall off data from American jurisdiction remains a genuinely unsettled legal question. The deal builds on Microsoft's European Digital Commitments from 2025, and it turns Mistral into something more than a model vendor — it becomes an independent European compute supplier woven into Microsoft's enterprise stack.

On the model side, the news is more concrete. Mistral Medium 3.5 and Mistral OCR 4 are now live in Microsoft Foundry, with Medium 3.5 also folded into Copilot Studio. Medium 3.5 is an open-weight model with 128 billion parameters and a 256,000-token context window, which suggests it's built for chewing through long documents or extended conversations rather than quick one-off queries. OCR 4, meanwhile, is aimed at document-heavy workflows — it handles 170 languages and holds onto layout details like bounding boxes, document structure, and confidence scores, the kind of metadata that matters when you're digitizing contracts or regulatory filings at scale.

What ties the compute and model layers together is portability. Engineers can build and fine-tune inside Microsoft Foundry, then push that same workload to public Azure, Azure Local, or Mistral's sovereign infrastructure without rewriting anything underneath. Some of those environments can even go fully air-gapped, cut off entirely from the public internet — a option that used to be a niche request from defense or finance clients and is now getting treated as a standard deployment path.

Microsoft hasn't disclosed the financial terms, capacity numbers, or a rollout schedule, and that vagueness is worth sitting with. But the strategic logic is clear enough: Microsoft is betting that flexibility in where and how AI runs will matter to enterprise buyers as much as which model tops a leaderboard. If regulated industries buy into that framing, Microsoft ends up owning the control plane for their AI even while the compute underneath belongs to someone else entirely.