French companies across space, public services and finance are deploying agentic computing systems built on Nvidia hardware, Nvidia said on 13 October, marking a transition from two years of summit pledges to operational infrastructure.
The wave of deployments—spanning European-manufactured server racks, an orbital edge-AI satellite and sovereign digital workplaces—shows how France and its partners are building a compute supply chain that reduces reliance on non-European cloud providers while creating demand for Nvidia’s next-generation data-centre and edge products.
France’s AI infrastructure push began with the February 2025 AI Action Summit and the Choose France summits of 2025 and 2026, which anchored commitments for large-scale compute on French soil, Nvidia said. Since then, the ecosystem has moved from blueprint to delivery.
Servers and Satellites: The Infrastructure Layer
Bull is extending its European AI server production to Nvidia’s Vera Rubin Ultra NVL72, a liquid-cooled rack-scale platform, according to Bull. The system packs 72 Rubin Ultra GPUs into a single rack via seventh-generation NVLink and scales to 576 GPUs by connecting eight racks through direct optical links. Bull plans to manufacture trays in Europe and complete rack integration at its Angers factory—inaugurated on 2 October—in the first half of 2028. The platform keeps the same rack footprint and maximum power draw as the current Vera Rubin NVL72, so data centres designed for the older generation can upgrade without reworking power distribution. It supports AC and 800 VDC power and integrates Caliptra, an open-source root of trust, with post-quantum secure boot. Bull and Nvidia did not disclose pricing or initial production volumes.
Meanwhile, the first satellite of a $1 billion, 50-satellite AI constellation reached orbit on 1 October aboard a SpaceX rideshare mission, Loft Orbital announced. The satellite, named Altair-1, carries two Nvidia Jetson AGX Orin modules running Mistral’s Ministral 3 model alongside nearly 20 other computer-vision and language models from NASA’s Jet Propulsion Laboratory, Google, IBM, Meta, Microsoft and Hugging Face. The edge hardware corrects images against ground reference points and detects events on orbit, so a user receives an alert within seconds rather than a raw image file hours later. Applications include maritime domain awareness, wildfire and disaster response, and port security. Onboard data is encrypted with post-quantum protection using ML-KEM-1024, the highest security level in the new NIST standard, Loft Orbital said. The $1 billion programme, announced by President Emmanuel Macron on 9 September, is held by a dedicated French company and built by Orbitworks in Abu Dhabi, combining Emirati capital with European supply‑chain expertise. For future satellites, Loft Orbital is evaluating Nvidia’s Space-1 Vera Rubin compute module and IGX Thor system.
Agents, Models and the Sovereign Stack
On the software side, LINAGORA and Docaposte secured the CANUT framework agreement that makes Twake.ai—a digital workplace built on Nvidia’s Nemotron open models—the sole referenced sovereign platform for more than 2,500 French public organisations, Nvidia reported. In financial services, SBS is launching a sovereign banking platform that lets institutions run governed generative and agentic services on-premises, also in collaboration with Nvidia. Separately, Mistral introduced a private preview of its Mistral Large 4 open model, with capabilities in cyber defence, manufacturing, finance and multimodal tasks, though no public release date was given. Fondation Inria, with Nvidia’s support, is launching a working group on open-source advanced modelling and tool adoption, with findings expected in spring 2027.
Enterprise deployments are producing the first performance claims. L’Oréal is integrating Nvidia’s ALCHEMI platform to let scientists explore formulation possibilities in silico; the company told Nvidia that combining computational predictions with hands‑on experimentation could make the discovery process up to 100 times faster than traditional methods. Mirakl said that fine‑tuned open models running on Nvidia GPUs cut seller‑catalog onboarding to under two hours, after onboarding more than 70 million products. Omi, a start‑up that uses 3D digital twins as a “product source of truth,” stated that its pipeline—built on Nvidia OptiX, CUDA and RTX PRO 6000 Blackwell Server Edition GPUs, running on the European Verda cloud—delivers a 10× speedup on Microsoft TRELLIS inference, reducing overall generation‑pipeline latency by 4×. In Côte‑d’Or, a digital‑twin agent powered by Nvidia Cosmos, Nemotron and Omniverse manages nearly 5,800 km of roads and buildings; it has cut total operating costs by 10% and is targeting 30%, Nvidia said. Enchanted Tools, a Paris‑based robotics company, is deploying social robots for patient support in paediatric cancer care. Bouygues Telecom and Prisme.ai launched AI Studio V2 on Nvidia’s full‑stack infrastructure to let employees build their own agents securely. In quantum computing, Alice & Bob integrated Nvidia NVQLink for superconducting cat‑qubit error correction, and Quandela connected a DGX Spark system to photonic qubits for hybrid quantum‑classical exploration.
None of the performance figures has been independently verified, and neither Bull nor the satellite programme offered per‑unit pricing. Nvidia’s GTC Berlin conference, where CEO Jensen Huang will give a keynote on 21 October, is likely to surface more details.