NVIDIA details agentic AI deployments across French industry

NVIDIA’s roundup describes French organizations using its computing infrastructure and open models across space, public services, banking, retail, scientific research and quantum computing. Its examples range from live systems to planned hardware, previews and demonstrations. Taken together, they do not establish that every project is an operating agentic deployment.

One of the largest projects is in space: Loft Orbital says the first spacecraft in its planned Orbitworks constellation launched on October 1. The company describes a $1 billion program of 50 satellites and says the first carries two NVIDIA Jetson AGX Orin modules and Mistral models. Loft says onboard models can analyze observations in orbit and send alerts instead of raw imagery; those latency and autonomy claims have not been independently measured.

In computing infrastructure, Bull says it plans to build NVIDIA Vera Rubin Ultra NVL72 systems on the European production line used for its Vera Rubin program, with final rack integration in Angers and initial availability targeted for the first half of 2028. The timetable is a company target, not a completed deployment.

The public-sector example has advanced further in procurement than deployment. Docaposte says the CANUT public-sector purchasing group selected a LINAGORA–Docaposte offer built around Twake.ai, making the sovereign collaborative environment available to eligible public bodies. NVIDIA separately says Twake.ai uses Nemotron and is the sole referenced sovereign digital workplace for more than 2,500 French public organizations, details the opened Docaposte announcement does not confirm.

In banking, SBS says its NVIDIA-powered AI Foundation is in production at Scottish Building Society. SBS describes autonomous agentic workflow automation as a near-future capability, however, so the deployment is not evidence that those autonomous workflows already operate in production.

The roundup also includes Mistral Large 4, although the two companies describe its release status differently. Mistral announced a public preview on October 6 and said model weights would follow later in October, while NVIDIA called the model a private preview. DataPhoenix’s coverage of the Large 4 public preview provides additional context on the release. Mistral says it trained the model on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers.

The research and industrial examples are at different stages. Fondation Inria says it launched an open-source AI adoption working group with NVIDIA as a founding member and main contributor, with findings due in spring 2027. L’Oréal says it is integrating NVIDIA ALCHEMI into its research environment to predict how molecules may perform in formulation work. L’Oréal claims the approach can make discovery 100 times faster than traditional methods, but the cited releases provide no independent methodology or validation for that figure.

In commerce, Mirakl says its Catalog Transformer has processed more than 70 million products using internally hosted open models and proprietary services on NVIDIA-accelerated infrastructure. NVIDIA says the system cut seller catalog onboarding to under two hours, a metric not stated in Mirakl’s opened account. Omi describes its NVIDIA-accelerated ProductDrop system as an upcoming feature combining generative AI with approved 3D digital twins, not an operating production service. Omi reports faster inference and lower pipeline latency after GPU optimization, but those figures have not been independently reproduced.

The quantum-computing examples remain demonstrations rather than production systems. Alice & Bob says it used NVQLink and synthetic data in a proof of concept for GPU decoding of LDPC-cat codes, while noting that the work does not yet amount to real-time decoding. Quandela says its GTC Berlin demonstrator links a DGX Spark workflow to its controller through NVQLink; the exhibition setup uses a simulated quantum processing unit, although Quandela says it previously validated the integration path with a physical photonic processor.