JUPITER, Europe's First Exascale Supercomputer, Goes Live on NVIDIA Grace Hopper; Maps Brain at Cellular Scale, Climate at 1km Resolution
JUPITER, Europe's first exascale supercomputer, is now live at the Jülich Supercomputing Centre in Germany, powered by approximately 24,000 NVIDIA GH200 Grace Hopper Superchips and NVIDIA Quantum-2 InfiniBand networking. The system delivers over 90 exaflops of AI performance and 1 exaflop for HPC applications while consuming just 18.2 megawatts of power, achieving 60 gigaflops per watt—the highest energy efficiency among the top five systems on the TOP500 list. JUPITER runs 1 quintillion FP64 operations per second and is owned by the EuroHPC Joint Undertaking, serving as a flagship for European scientific sovereignty in AI and high-performance computing.
Early science applications are moving from theoretical possibility to delivered results. The Jülich Brain Atlas project trained a foundation model for brain microarchitecture analysis (CytoNet) in under five days on 4,096 Grace Hoppers, processing 6.5 petabytes of cellular imaging data from 21 post-mortem brains. For the first time, researchers built an AI agent that 'thinks through' neuroscience experiments rather than simply analyzing static data—a capability that opens the door to accelerating drug discovery and therapies for neurodegenerative diseases. ICON, a coupled Earth system model, won the Gordon Bell Prize by simulating the planet's entire climate at 1-kilometer resolution (atmosphere, ocean, land, carbon cycle) at once. The simulation achieves 146 days of real climate per 24 hours of compute. Researchers also set a world record by simulating a 50-qubit quantum computer—surpassing the previous 48-qubit benchmark.
JUPITER's architecture combines CPU-GPU coherence (Grace CPU spills data into main memory when GPU memory fills, maintaining bandwidth with minimal latency loss), modular design (Booster for AI/HPC workloads, Cluster for general compute), and warm-water cooling that heats the Jülich campus. The system enables Ericsson and Jülich to develop brain-inspired, energy-efficient AI models for 6G networks, and supports quantum research through NVIDIA CUDA-Q and cuQuantum toolkits. For European enterprises and national labs, JUPITER removes the infrastructure bottleneck that previously required outsourcing to US cloud providers.
The strategic implication extends beyond capability. JUPITER signals Europe's commitment to digital sovereignty in AI infrastructure and anchors a planned network of ~13 specialized AI data centers across the continent. For architects evaluating where to run exascale workloads—climate simulation, large-scale model training, quantum research—the existence of a European system with formal EuroHPC governance removes dependency lock-in to US commercial clouds. The system is open to 18 German and 15 European teams as early access users, with broader access expected in coming months.