Agentic artificial intelligence in France is moving beyond the demonstration phase into operational workflows across public administration, banking, industrial research, telecommunications networks, and even Earth observation from space. This is the picture emerging from the initiatives presented by NVIDIA, where on-soil computing capacity and the adoption of open models have become central pillars of the French strategy.
The distinguishing feature is not the adoption of a single model or platform, but the interplay between the hardware supply chain, local software, and sovereignty requirements. In Europe, the AI debate has often centered on regulation and the gap with the United States; in France's case, however, the applications described point to a more concrete effort to build environments where public and private organizations can deploy generative models and agents while retaining control over data, infrastructure, and deployment.
The push has also been laid out on political and industrial fronts. The AI Action Summit hosted in Paris in February 2025 and subsequent editions of the Choose France summit secured commitments for new computing investments in the country. While those announcements do not automatically equate to immediately available capacity, they have set a clear direction: turning France into a European base for data centers, advanced computing systems, and AI services aimed at organizations unwilling to entrust sensitive information to platforms operated outside the Union.
From digital sovereignty to production capacity
One piece of this strategy concerns the manufacturing of the systems that will host AI workloads. Bull, an Eviden group company, inaugurated its rack integration facility in Angers on October 2 and plans to expand the production of its AI servers across Europe. The roadmap includes the upcoming NVIDIA Vera Rubin Ultra NVL72 systems, expected in 2028: the trays are slated to be manufactured in Europe, while rack integration will take place at the French facility.
This is still a forward-looking project, yet relevant to industrial policy: localizing part of the production chain does not equate to making the entire technology stack independent, given that core components continue to be designed by NVIDIA. However, it can reduce certain logistical constraints, provide integration capabilities close to European users, and build a technical supply chain that until now has been concentrated primarily in the United States and Asia.
On the software side, NVIDIA points to the CANUT framework agreement as a significant case, through which LINAGORA and Docaposte secured the reference for Twake.ai, a sovereign digital workplace based on NVIDIA Nemotron models. The offering is aimed at more than 2,500 French public organizations. The main point is less the chatbot itself than the delivery model: collaborative tools and generative assistants can be deployed in administrative environments with stringent requirements for security, traceability, and data processing.
The same logic drives the initiative by SBS, which is developing a sovereign platform for banking and financial services with NVIDIA. The stated goal is to enable institutions to run generative and agentic services within their own governed environments. For the financial sector, where an agent can consult documents, extract data, and trigger procedures, the ability to define permissions, audit trails, and operational boundaries is an essential condition, not an add-on.
Open models, but not without constraints
France can also rely on a local presence in the foundation model market. Mistral has made Mistral Large 4 available in private preview, an open-weight model geared, among other things, toward use cases in cybersecurity, manufacturing, finance, and multimodal tasks. The value of open-weight models for enterprises and public administrations lies in the ability to fine-tune them to proprietary vocabularies, documents, and processes, offering greater latitude in choosing the infrastructure that runs them.
Open does not automatically mean transparent or secure. Organizations must still assess data quality, hallucinations, the protection of confidential information, and accountability for decisions made or suggested by agents. For this reason, Fondation Inria, with support from NVIDIA, has launched a working group dedicated to advanced open-source modeling and the tools needed to adopt it; results are expected in the spring of next year. It is a sign of the remaining gap between the availability of a model and its reliable deployment in a critical process.
Industrial applications show where this transition is already taking place. L’Oréal is integrating NVIDIA ALCHEMI into its research workflows to evaluate potential formulations computationally before heading to the lab. The company estimates that combining computational predictions with experimentation can make the discovery process up to a hundred times faster compared to traditional methods. While this is a corporate estimate rather than a result generalizable to every project, it clearly illustrates the role these tools aim to play: narrowing down the hypothesis space that needs to be physically tested.
In digital commerce, Mirakl has uploaded over 70 million products using optimized and open models on NVIDIA GPUs and inference software. The company claims to have reduced seller catalog onboarding to under two hours. Photoroom, Vestiaire Collective, and Omi, meanwhile, focus on the visual side: Omi combines brand-approved 3D digital twins with open models optimized on NVIDIA and reports four times faster inference for producing generated marketing imagery.
Agents operating in the physical world
The most critical aspect of agentic AI emerges when systems do not merely generate text or images, but contribute to managing physical assets. In the Côte-d’Or department, ThinkDeep and Akila have developed a digital-twin-based agent to monitor nearly 5,800 kilometers across roads and buildings. The system uses Cosmos and Nemotron models alongside NVIDIA Omniverse libraries. According to those involved, the project has already reduced overall operating costs by 10% and aims for 30% as deployment expands.
Port Sud de France is also deploying ThinkDeep's DeepBrain platform, based on Nemotron and Cosmos, across cargo processing and port security operations. In these scenarios, the potential benefit lies in unifying data from disparate sources and prioritizing tasks. However, the need for human oversight is growing: a misclassification, an opaque recommendation, or unauthorized system access can have tangible consequences, ranging from traffic flow management to the security of strategic infrastructure.
Bouygues Telecom and Prisme.ai have also launched AI Studio V2, a platform that enables the operator's employees to build agents within a controlled environment. The model aims to prevent internal automation from developing via consumer tools outside corporate governance. Enchanted Tools, a Parisian robotics startup, is bringing agents into healthcare facilities: its social robots are used to support patients, including in pediatric oncology care. Here, AI does not replace clinical work, but enters a context where reliability, interaction design, and the protection of vulnerable individuals require particularly rigorous verification.
The most extreme example comes from orbit. Altair-1, the first satellite in a planned 50-unit constellation backed by one billion dollars in funding, was launched this month and reached space. Integrated by Loft Orbital and built in Abu Dhabi by Orbitworks, the satellite uses NVIDIA Jetson AGX Orin modules to run Mistral models on board. The goal is to perform real-time Earth observation and autonomous operations, limiting dependence on ground station links.
The Altair-1 case does not represent an entirely French production, but it is indicative of the international network underpinning the ecosystem: models developed in France, US hardware, and integration and manufacturing distributed across multiple countries. In this perspective, technological sovereignty is not absolute self-sufficiency. It is the ability to choose where models run, who can access data, how they are adapted, and which skills remain available within the country.
The toughest test begins now. Projects will need to demonstrate measurable results beyond pilot stages, maintain control over compute costs, and comply with European obligations regarding data and artificial intelligence. France boasts a rare combination of public institutions, tech companies, research, and industrial demand. It remains to be seen whether this critical mass will successfully translate into durable platforms capable of competing in the market without relying entirely on the vendors currently enabling their growth.
