JUPITER, Europe's first exascale supercomputer, used its debut at ISC in Hamburg this week to publish results across four scientific domains, including a 1-kilometer global climate run on 20,480 NVIDIA Grace Hopper Superchips and a fully simulated 50-qubit quantum computer that breaks the prior 48-qubit record. The machine, hosted at Germany's Forschungszentrum Jülich, runs on Grace Hopper chips paired with NVIDIA Quantum-X800 InfiniBand networking. Each of the four projects targets a problem that was effectively out of reach on prior hardware.
The throughline is scale. JUPITER's CPU-GPU memory architecture lets workloads spill from GPU memory into CPU memory without the usual performance cliff, which is what allowed the 50-qubit simulation to clear the previous record set on smaller systems. Coupled with the system's 20,480-chip headline configuration for climate work, the machine is being positioned by Jülich as a general-purpose exascale platform rather than a single-domain showcase.
“With JUPITER, Europe doesn't just join the exascale era — it leads it, across the widest range of science and AI of any system worldwide.”— Thomas Lippert, Director of the Jülich Supercomputing Centre
The brain project is the most ambitious AI-native workload. The Jülich Brain Atlas team, anchored at the Institute of Neuroscience and Medicine (INM-1) with Helmholtz AI, trained CytoNet, a foundation model for brain microarchitecture, on JUPITER in under five days. The training run consumed 6.5 petabytes of imaging data drawn from 21 post-mortem brains, using 4,096 Grace Hopper Superchips. The target is mapping a brain with 86 billion neurons and roughly 100 trillion connections at single-neuron resolution.
Key facts
- 01JUPITER, Europe's first exascale system, runs on NVIDIA Grace Hopper Superchips and Quantum-X800 InfiniBand at Forschungszentrum Jülich.
- 02Jülich researchers trained CytoNet — a brain-mapping foundation model — in under 5 days using 6.5 petabytes of data on 4,096 Grace Hopper Superchips.
- 03An ICON climate run on 20,480 Grace Hopper Superchips simulated 146 days of Earth climate at 1-km resolution in 24 hours of compute.
- 04JSC and NVIDIA simulated a universal 50-qubit quantum computer, surpassing the prior 48-qubit record.
- 05Ericsson and Forschungszentrum Jülich are using JUPITER to train AI models for 5G and 6G networks.
CytoNet is meant to feed an agentic system. The team's next step is building an AI agent for brain researchers that can reason across modalities and answer questions directly against the underlying brain data, using open models including NVIDIA Nemotron 3 120B. The framing — using a foundation model not just to analyze but to drive experimentation — is a shift in how computational neuroscience is being structured.
Katrin Amunts, who leads INM-1, said the agent approach changes what neuroscience can attempt, and that JUPITER is the system that makes the work tractable today. The paper describing CytoNet is on arXiv.
The climate workload is the headline performance result. A novel configuration of the ICON model, developed by ETH Zurich, the German Climate Computing Centre, Jülich Supercomputing Centre, the Max Planck Institute for Meteorology, NVIDIA, the Swiss National Supercomputing Centre, and the University of Hamburg, won the Gordon Bell Prize for Climate Modelling at SC25 last November. ICON is the first coupled Earth-system model to run at 1-kilometer global resolution with ocean, atmosphere, land, biogeochemistry and the full carbon cycle exchanged across all components.
On 20,480 Grace Hopper Superchips, the model simulated roughly 146 days of real climate in 24 hours of compute — a world record in global climate simulation. At kilometer resolution, ocean eddies, upper-ocean mixing and fine-scale winds emerge from physics rather than parameterization, which is the qualitative jump that justifies the hardware spend.
“At a global resolution of just 1 kilometer, many of these interactions emerge directly from the laws of physics rather than being approximated.”— Daniel Klocke, Group leader at the Max Planck Institute for Meteorology
The wireless project is more forward-looking. In March, Ericsson and Forschungszentrum Jülich announced a collaboration to develop AI for ongoing 5G evolution and for 6G networks, with JUPITER as the training and testing platform. Research priorities include AI models for Ericsson's radio and core networks, energy-efficient inference at the radio edge using neuromorphic approaches, and modular supercomputing concepts drawn from Jülich's exascale architecture.
The quantum result is narrower but symbolically large. Researchers at JSC, working with the jointly operated NVIDIA Application Lab, fully simulated a universal 50-qubit quantum computer, exceeding the previous 48-qubit ceiling. The simulator, JUQCS-50, will be accessible inside JUNIQ, the quantum user facility at JSC. The practical point: today's physical quantum hardware can't yet beat classical machines on useful problems, so high-fidelity simulation at 50 qubits is where the algorithms that future hardware will run get designed and stress-tested.
The caveats are real. Exascale results at this scale are still rare events tied to specific machine allocations and hand-tuned configurations, not workflows that ordinary research groups can spin up on demand. CytoNet's agentic next step is unbuilt; ICON's kilometer-resolution runs remain expensive enough that they will be reserved for headline experiments rather than routine forecasting. And classical simulation of 50 qubits, while a record, is also a reminder that scaling beyond ~50 qubits gets exponentially harder on any classical system, exascale or not.
JUPITER matters less as a benchmark trophy than as evidence that the Grace Hopper platform now anchors frontier science workloads outside the US hyperscaler stack. Four results across neuroscience, climate, telecom AI and quantum simulation, all on the same machine inside a year of operation, give NVIDIA a durable European reference customer at a moment when sovereign-compute arguments are reshaping how governments buy AI infrastructure. For Jülich, the bet is that exascale becomes the default substrate for foundation-model science — and the early results suggest that bet is paying off faster than the usual supercomputer ramp.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
