Vast Data and Mistral Compute announce a partnership focused on the deployment of the Vast AI Operating System as the data layer of Mistral AI’s inference computing center in Les Ulis, Essonne. The infrastructure, powered by Nvidia GB300 NVL72 accelerators, is presented as the largest concentration of its type in Europe. It covers the entire life cycle of AI models, from training to production.
The announcement comes during the AI Now Summit organized by Mistral AI on May 28, 2026 in Paris, where the French company detailed its positioning as a complete stack supplier for large industrial accounts. The partnership with Vast Data is part of the infrastructural dimension of this strategy. The 10 megawatt data center planned in Les Ulis for the third quarter of 2026 will be based on a unified data architecture provided by the American publisher, alongside Nvidia GB300 NVL72 accelerators grouped in racks of 72 units. It is on this technical basis that Mistral Compute, the group’s infrastructure division, intends to offer inference services to its industrial and public clients.
The combination of the three players draws a distinct layered architecture where Vast Data provides the data management foundation, Nvidia provides the computing power with the GB300 GPUs, and Mistral Compute orchestrates the whole thing as the platform operator. This scheme aims to reduce friction between the storage, acceleration, and inference layers that are the most common bottlenecks in large-scale deployments.
The Vast AI Operating System unifies data flows
The Vast AI Operating System is the central building block provided by Vast Data. It is a platform that simultaneously manages the flows necessary for model training, augmented recovery operations and inference pipelines in production. Its particularity lies in the elimination of the silos usually separated between high performance storage, object storage and vector databases, which reduces latencies and simplifies the architecture of MLOps teams.
“AI is poised to surpass existing data architectures, and Vast AI Operating System is the foundation for building the future,” summarizes Alon Horev, CTO of Vast Data. This formulation covers a precise technical reality for large-scale deployments. The consistent management of data between the training phase and the production phase remains one of the most costly challenges in terms of engineering time in organizations that have passed the experimentation milestone. The Vast platform aims to reduce this friction without requiring a rewrite of existing pipelines.
Yann Léger, vice-president of cloud computing at Mistral AI, confirms that the choice of Vast Data is due to its ability to align storage and processing needs across the entire life cycle of the models. The decision to build a proprietary infrastructure rather than relying on the shared offers of hyperscalers responds to a declared objective of direct control of GPU resources and visibility into the security of processing.
Nvidia GB300 NVL72 racks size the capacity
The choice of the Nvidia GB300 NVL72 accelerator structures the entire architecture. Each NVL72 rack brings together 72 GB300 graphics processors interconnected by NVLink, which makes it possible to distribute the inference of large models on a compact physical unit while maintaining high memory bandwidth. For Mistral Compute, this sizing directly determines the inference throughput per task and the cost per token charged to clients.
Rod Evans, Vice President EMEA Supercomputing at Nvidia, places the deployment in the trajectory of large-scale AI infrastructures in Europe: “Mistral Compute is creating a world-class AI infrastructure. “The wording anchors the announcement in the competition for sovereign computing capacity in Europe, a subject that has mobilized both industrial players and public institutions since the publication of the European Commission’s AI strategy in 2024. According to the data communicated, the Les Ulis center will be the largest deployment of racks GB300 NVL72 in Europe when it opens.
Sovereignty claimed
Mistral AI is a company incorporated under French law with no direct exposure to the American Cloud Act. This legal status, combined with the proprietary operation of the Les Ulis center, constitutes a justifiable regulatory advantage for customers who process sensitive data within the meaning of the GDPR or who fall under the obligations defined by NIS2, transposed in France in October 2024.
The presence of Vast Data and Nvidia in the architecture introduces a nuance that public organizations’ procurement and compliance teams will need to examine. Vast Data is an American company and Nvidia is an American manufacturer subject to United States export controls. The data hosted in Les Ulis passes through a software layer published in the United States and through accelerators designed by an American entity. The assertions of sovereignty relate to the operator of the platform and the location of the processing, not to the neutrality of the entire chain thus constituted.
The Les Ulis center has not yet been the subject of a qualification published by ANSSI. Public buyers and critical operators who plan to rely on Mistral Compute will need to verify, at the time of calls for tender, the level of qualification obtained and the contractual conditions of access of Vast Data and Nvidia support teams to production environments. The commercial availability of the service was not specified during the Summit.
All that’s missing is the commissioning date.
For industrial organizations leading large-scale AI model deployment projects, the announcement raises three practical questions. The first concerns the service model. Mistral Compute positions itself as a shared infrastructure operator. Guarantees of data isolation and partitioning of GPU resources between clients must be contractually integrated before any deployment of models processing sensitive data.
The second concerns compatibility with third-party models. The infrastructure is presented as designed for the Mistral models, but nothing in the Vast and Nvidia architecture technically prohibits the deployment of other open source models. Mistral Compute’s commercial policy on this point was not detailed during the Summit. The third touch is on the calendar. The Les Ulis center is announced for the third quarter of 2026 without a firm date of commercial availability. For projects scheduled to start before the end of the year, validation of the commissioning schedule will be a prerequisite for the inclusion of Mistral Compute in the reference architectures.
