NVIDIA Maps Out Europe’s AI Infrastructure in Paris

NVIDIA chose Paris to argue that Europe should not be limited to consuming artificial intelligence developed elsewhere. Speaking at GTC Paris, held alongside VivaTech, Jensen Huang laid out a roadmap built around homegrown computing capacity, country-specific models and industrial applications.

The message carries a clear economic implication: the AI race is no longer fought solely in the labs training language models. It also depends on who builds the data centers, supplies the power, manufactures the servers and brings those systems into factories, transport networks and public services.

From data centers to “AI factories”

Huang placed Blackwell systems—the generation of NVIDIA chips designed to train and run AI models—at the center of that strategy. He described the GB200 NVL72 as a system that operates like a single giant GPU, combining 72 Blackwell GPUs and Grace processors to function as one large-scale computing unit.

NVIDIA says its partners are already manufacturing 1,000 GB200 systems a week. Beyond that figure, Huang linked the hardware to the growth of inference: the work of answering questions, generating content or carrying out reasoning when a user interacts with a service.

According to Huang, the number of users of inference services has grown from 8 million to 800 million in just a few years. The comparison illustrates why chip companies are describing data centers as “AI factories”: facilities that produce tokens, the units of text and other data generated by models.

Europe starts at a disadvantage relative to the United States and China in technology platforms and major public clouds. But it retains assets that are difficult to replicate: manufacturing industry, supercomputing centers, telecommunications operators and regulation that pushes companies and public authorities to control where their data is processed.

Local models and agents for businesses

The company also focused on so-called sovereign models: AI systems that can be tailored to a local language, regulatory framework or dataset without relying entirely on a foreign provider. NVIDIA plans to build on Nemotron, its family of open models, so European developers can create specialized assistants and applications.

The announced integration with Perplexity is intended to offer reasoning-enabled search and multilingual deployments for European customers. A model speaking Spanish, French or German is not enough on its own to make it sovereign. The underlying question is who controls its training data, the infrastructure it runs on and its terms of use.

Huang also introduced tools for building AI agents: programs able to chain together tasks, retrieve information and use digital tools with a degree of autonomy. NVIDIA has launched NeMo Agent and templates for creating so-called data flywheels, cycles in which a company uses the data generated by its operations to improve its systems.

This approach could prove useful for customer service, document analysis or industrial maintenance. It also raises the bar for oversight: an agent that accesses corporate data or carries out actions requires limited permissions, traceability and security controls—not just a strong language model.

Germany, the first industrial cloud

The announcement most directly aimed at the European economy was the creation of an industrial AI cloud in Germany. NVIDIA describes it as the first of its kind and links it to Omniverse, its platform for building digital twins: simulations of factories, vehicles or facilities that make it possible to test changes before applying them in the physical world.

The proposal is particularly well suited to European industry. A manufacturer can simulate an assembly line, train computer vision systems or plan warehouse operations without halting production. Its usefulness will depend less on having a powerful GPU than on integrating reliable data from machinery, design, logistics and operations.

NVIDIA also announced new technology centers in Finland, Germany, Spain, Italy and the United Kingdom, and highlighted CUDA-Q, its platform for combining quantum and classical computing, already operating on Denmark’s Gefion supercomputer. Quantum computing does not yet replace conventional systems for general business uses, but combining the two fields is a significant research bet for Europe.

Infrastructure alone will not close the gap

NVIDIA’s pitch addresses a real need: without computing capacity, Europe’s ambitions for technological autonomy remain dependent on external providers. But building AI infrastructure requires sustained investment, access to energy, adequate power grids and professionals capable of operating it.

NVIDIA also has a direct interest in the effort. The more countries, operators and manufacturers build their own AI platforms, the larger the market for its chips, networks and software. Europe’s opportunity will be to turn that investment into homegrown companies and services, rather than simply installing hardware to run products built elsewhere.