NVIDIA's latest rundown of its France ecosystem turns sovereign AI from a slogan into a checklist: on October 12, 2026, NVIDIA's official blog surveyed agentic deployments by French companies and institutions, and the most striking item is no longer on the ground — Altair-1, the first satellite of a $1 billion, 50-satellite constellation, launched this month and reached orbit carrying NVIDIA Jetson AGX Orin edge modules and running Mistral's models for real-time Earth observation and autonomous operations in orbit.
The ground half: servers to be built in Europe
Beyond the constellation, ground infrastructure is also being pulled onto European soil: server maker Bull is extending its European AI server production to NVIDIA Vera Rubin Ultra NVL72 systems, planned for 2028, with trays produced in Europe and rack integration at its Angers factory, inaugurated on October 2. The significance is not one more assembly plant; it is that the sovereignty story is reaching manufacturing. Where compute hardware is built and integrated is becoming a first-order procurement question for European customers, alongside where models run.
The model half: open weights as the entry ticket
A recurring combination in the rundown is NVIDIA's open Nemotron models running inside local institutions' own environments. LINAGORA and Docaposte secured the CANUT framework agreement, making Twake.ai, built on Nemotron, the sole referenced sovereign digital workplace for more than 2,500 French public organizations. In finance, SBS is launching a sovereign platform that lets banks deploy governed generative and agentic services within their own environments. Last week Mistral introduced its open model Mistral Large 4 under private preview, aimed at cyber defense, manufacturing, finance and multimodal tasks. Read together, European buyers are not asking for the strongest closed model; they want models that stay inside their own boundaries and can be audited and customized. Here, openness is a compliance condition, not a cost-saving option.
The deployment list hides checkable output numbers
What keeps this rundown from being only a brochure is a set of verifiable business results: L'Oréal is integrating NVIDIA ALCHEMI into research workflows, with discovery claimed to run up to 100 times faster than traditional methods; marketplace platform Mirakl has onboarded more than 70 million products with fine-tuned and open models, cutting seller catalog onboarding to under two hours; a digital-twin agent in Côte-d'Or manages nearly 5,800 kilometers of roads and buildings, with operating costs already down 10% and a 30% target; and Bouygues Telecom with Prisme.ai has launched AI Studio V2 so all employees can build their own agents in a controlled environment. The right way to read the list is in layers: the satellite and server lines are 2028 positioning, the workplace platform and product onboarding are production systems already in use, and figures such as 100x remain vendor claims awaiting independent customer confirmation. The sovereign AI contest is shifting from conference statements to who gets systems running inside the daily workflows of governments, banks and factories first.
