35 new NVIDIA AI supercomputers for Europe

From a technical point of view, the new systems are based on NVIDIA Blackwell and Hopper full-stack hardware platforms, supported by Quantum InfiniBand interconnection networks and CUDA-X software libraries. Since last year’s budget, the overall computational capacity distributed or planned on European territory has reached 800 AI exaflops, a quota intended to support predictive models in the fields of climatology, biomedicine, energy decarbonization and quantum simulation. Jensen Huang, founder and CEO of NVIDIA, described the strategic direction of this expansion: “Artificial intelligence is the new tool of science and Europe is building the infrastructure to put it in the hands of millions of researchers. With NVIDIA Accelerated Computing, scientists can simulate more complex systems, train scientific AI models, and structure agentic AI workflows that transform European data and expertise into world-changing discoveries”.

Jensen Huang, founder and CEO of NVIDIA

The configuration of the new European installations, called AI factories, touches on several key points of continental research. In Spain, the Barcelona Supercomputing Center will expand the MareNostrum5 system through the integration of GB300 NVL72 and GB200 NVL4 modules connected via the InfiniBand Quantum-X800 platform, a credited architecture of 20 exaflops for AI training and 33 exaflops for inference. Mateo Valero Cortés, director of the Barcelona centre, framed the initiative: “The BSC is committed to building an artificial intelligence infrastructure that advances science, industry and society. With the upgrade to MareNostrum5 and NVIDIA accelerated computing, the consortium made up of Spain, Portugal and Turkey will provide European researchers with the tools to tackle some of the world’s most complex challenges, from climate modeling to biomedical discovery”.

In Germany, the BavariaAI program involves the activation of the Blue Swan platform at the Friedrich-Alexander University of Erlangen and the LRZ centera 1,000 GB200 NVL4 GPU cluster credited with 11 exaflops in training and 22 in inference. Markus Blume, Minister of Science of Bavaria, commented on the project: “With the ‘Blue Swan Platform’ project, Bavaria is working on an innovative and independent multimodal foundational artificial intelligence model, aimed at important application areas such as healthcare and robotics. This will allow us to provide a powerful AI tool for science and industry, fully compliant with European standards. For this ambitious goal, we are currently building a special computing infrastructure at the Friedrich-Alexander University of Erlangen, the largest GPU cluster ever installed at a German university”. Also on German territory, the HLRS center in Stuttgart will host HammerHAIthe first national AI factory equipped with over 850 GB200 NVL4 GPUs for engineering simulations and large language models, capable of expressing 8 exaflops in training and 15 in inference. Michael Resch, director of the Stuttgart centre, underlined the industrial value of the structure: “Germany has long been a leader in engineering, science and industrial innovation. With HammerHAI, Germany’s first AI factory, we are building on this foundation through a secure, national AI infrastructure that will help researchers and industrial users accelerate scientific simulation, inference and discovery, strengthening Europe’s ability to turn advanced computing into concrete breakthroughs for the real world”.

On the Italian front, the inter-university consortium CINECA, in coordination with the Ministry of University and Research and the National Cybersecurity Agencywill lead the development of the IT4LIA AI factory, an infrastructure based on more than 8,000 GB200 NVL4 GPUs designed to deliver 82 exaflops of training and 164 exaflops of AI inference. Gabriella Scipione, director of CINECA’s HPC sector, highlighted the strategic objectives of the Italian node: “IT4LIA marks a strategic step in strengthening the European AI and HPC ecosystem, providing a high-performance infrastructure to the research and innovation ecosystem. Through advanced accelerated computing, EuroHPC with CINECA, the Italian Ministry of University and Research and the National Cybersecurity Agency are creating a safe and reliable environment for the development of open AI models and applications in the agritech, cybersecurity, meteorology, climate and manufacturing sectors, enhancing Europe’s technological autonomy and strengthening Italy’s role in the global artificial intelligence scenario“. Complete the initial picture Sweden with the Mimer AI Factory at Linköping University, a facility owned by the EuroHPC Joint Undertaking that will have 100 GB200 NVL4 systems for the development of trusted artificial intelligence and life sciences.

High-performance computing flows find immediate application in energy transition and industrial climatology. An example of this integration is represented by the engineering activities of Siemens Energy, employing the Siemens Xcelerator suite accelerated by Omniverse libraries and CUDA-X frameworks to model the computational fluid dynamics and combustion processes of 100 percent hydrogen-fueled gas turbines, a complex physical transition due to extreme temperatures and fluid behaviors. The adoption of digital simulation-driven workflows has allowed calculation times for burner design iterations to be reduced by up to 77 percentspeeding up the subsequent technological validation through components made in additive manufacturing.

In parallel, European infrastructure sees progress in hybrid quantum-HPC architecture through the open CUDA-Q platform. At CINECA, EuroHPC and Pasqal are integrating an atom-neutral quantum processing unit (QPU) operated via the Slurm orchestrator, providing an environment for developing algorithms for materials science. In Germany, Fraunhofer FOKUS is working on interoperability between the CUDA-Q platform and the Eclipse Qrisp quantum programming language, while the Barcelona Supercomputing Center has integrated the same protocols into the control software of the analog quantum computer developed by Qilimanjaro Quantum Tech. In terms of simulation records, researchers at the Jülich Supercomputing Center performed the complete simulation of a 50-qubit universal quantum computer called JUQCS-50 on the JUPITER supercomputer, based on the GH200 Grace Hopper superchip, establishing a numerical reference for the verification of complex quantum problems on accelerated computing architectures.