Mistral is now letting customers choose the region for AI inference, hosting a third-party model for the first time, and courting long-term offtake agreements to fund new European compute capacity. With this, the French AI developer is deliberately positioning itself as a pan-European AI operator and a driving force behind Europe's compute infrastructure buildout.
Regional endpoints, with limits
Mistral AI is now offering so-called regional endpoints for production use. Customers can specify whether inference, meaning the processing of requests by a trained model, happens in Europe or the US. The company cites data residency, regulatory, and latency requirements as the reasons for the move. According to its documentation, Mistral charges a ten percent surcharge for using regional endpoints.
The pledge of regional processing isn't absolute, though. Mistral reserves the right to make limited, secured data transfers to subcontractors outside the chosen region. The option also only covers inference, not account data, billing, API keys, or usage statistics. In regional mode, the AI models also have access to only a limited selection of external tools, agents, for instance, aren't supported yet. At the same time, a so-called priority tier launches as a public preview, meant to offer individually agreed capacity and a 99.5 percent availability guarantee for business-critical applications, at a 75 percent surcharge over standard processing, according to the documentation.
GLM-5.2 as the first third-party model
Alongside its own models, Mistral is adding a third-party model to the platform for the first time. First up is GLM-5.2 from the Chinese company Z.ai, which Mistral positions for long contexts, coding tasks, and agentic workflows. According to Mistral, the model will get the same infrastructure, the same regional controls, and the same contractual commitments as its own models, meaning that even for a Chinese model, data processing for inference workloads should be selectable to stay exclusively within Europe.
Long-term contracts meant to fund the buildout
For its infrastructure expansion, Mistral is courting long-term commitments from companies and institutions. In exchange, they receive so-called European compute units, meaning access to Mistral products based on the compute capacity to be built in the future. By the end of 2027, Mistral wants to build up 200 megawatts of European AI compute capacity this way, and up to one gigawatt by 2030. Buyers of compute units are also meant to gain influence over which capacity gets built where.
The model resembles long-term infrastructure financing more than a flexibly usable cloud plan, and it shows many parallels with the business models of US companies like Nvidia and OpenAI. Mistral gets predictable commitments for products, hardware, and data centers, while buyers, in turn, tie themselves for years to a single provider and to capacity whose scope and availability aren't fully settled at the time of contracting. Separately from the new compute units, Microsoft expanded its existing partnership with Mistral in July 2026, securing access to future European GPU capacity through an undisclosed multi-billion-dollar contract.