Mistral AI announces Mistral Compute with NVIDIA

What was announced

On 11 June 2025, Mistral AI announced Mistral Compute, an offering in AI infrastructure.^[1] In the announcement the company described it as a private, integrated stack covering GPUs, orchestration, APIs, products and services, provided in whatever form factor a customer needs.^[1]

That description matters because it defines the boundaries of the product: not a single piece of software, but a layered stack. At the bottom sit the graphics processors (GPUs) that train and run AI models; above them sits orchestration software that decides which job runs on which machine; above that sit the APIs applications call, and then the services built on top. Mistral Compute is presented as covering all of these layers together, under one supplier.^[1]

How it is delivered

Customers can take the offering in two main forms. It is available as bare metal, meaning dedicated machines that the customer runs with its own software stack, and as a managed platform-as-a-service (PaaS), meaning the orchestration and APIs are run for the customer.^[1] The first form suits organisations that want full control over their environment; the second suits those that want the capacity without operating the layers above the hardware themselves.

Both forms are described as private, in the sense that the stack is dedicated to the customer rather than shared.^[1]

The NVIDIA link

Mistral described itself as a premier NVIDIA partner, offering tens of thousands of GPUs.^[1] The announcement also said the offering would include the latest NVIDIA reference architectures.^[1] In practice a reference architecture is a vendor's published blueprint for how to lay out a GPU system — which components go where, and how they are connected — so that a customer or systems integrator can build a working cluster without designing it from scratch. Including those designs means customers can follow a known layout rather than assembling one.

The combination of GPU supply and reference designs is what lets a customer take the offering either as raw capacity or as a running service.^[1]

Launch partners

The announcement named a set of launch partners: Black Forest Labs, BNP Paribas, Kyutai, Mirakl, Orange, Schneider Electric, SNCF, Thales and Veolia, with SLB Groupe also among the partners named.^[1] The list spans a wide range of sectors — banking, telecommunications, rail transport, energy and industrial equipment, defence electronics, and AI research — which suggests the offering was positioned at large organisations that want AI capacity inside their own environments rather than through a public cloud. The partners were named as launch partners; the announcement did not describe individual deployments.

Where it sits

Mistral Compute puts the company in the position of supplying the computing layer itself, together with the orchestration and API layers above it.^[1] For a customer, the practical effect of an integrated stack sold by one supplier is fewer separate contracts and fewer places where the layers have to be made to fit together: the GPUs, the scheduling software, the APIs and the services arrive as one package, in a form the customer chooses.^[1] Whether organisations buy it as bare metal or as a managed service, the same underlying hardware and reference designs are on offer.^[1]