JUPITER, Europe's first exascale supercomputer built on NVIDIA Grace Hopper Superchips, trains brain atlas models in days and hits 1km climate resolution

Europe's first exascale supercomputer, JUPITER, is producing science at a scale previously impossible. Located at Forschungszentrum Jülich in Germany and powered by NVIDIA Grace Hopper Superchips, the system is now public with a first wave of results spanning neuroscience, climate modeling, 5G/6G research, and quantum simulation — all published together on June 22, 2026.

What's new

NVIDIA's official account of JUPITER's opening results documents four distinct scientific milestones:

• Neuroscience: The Jülich Brain Atlas project trained CytoNet — a foundation model — on 6.5 petabytes of cellular-scale brain imaging data in under five days. A workload of that scale would have required months on pre-exascale systems.
• Climate: The ICON climate model ran at 1-kilometer global resolution while simulating a fully coupled Earth system, processing 146 days of climate data in a single 24-hour compute run. Kilometer-scale global resolution had previously been a theoretical target; JUPITER achieved it operationally.
• Connectivity: Ericsson and the Jülich Supercomputing Centre announced a collaboration to develop brain-inspired AI architectures for 5G and 6G networks, using JUPITER to prototype neuromorphic approaches to next-generation wireless infrastructure.
• Quantum: Researchers simulated a 50-qubit universal quantum computer on JUPITER — surpassing the previous record of 48 qubits and extending the frontier of what classical hardware can simulate.

Thomas Lippert, director of the Jülich Supercomputing Centre, described the system's breadth plainly: "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."

Context

Exascale computing — systems capable of at least one exaFLOP (10¹⁸ floating-point operations per second) — has become the benchmark for national-scale AI and scientific infrastructure. In the United States, Frontier at Oak Ridge National Laboratory and Aurora at Argonne crossed the exascale line first. JUPITER is Europe's entry into this tier.

NVIDIA's Grace Hopper Superchip underpins JUPITER's architecture. It combines an ARM-based Grace CPU with a Hopper GPU on a single package, using NVLink-C2C for chip-to-chip interconnect. This integration reduces the energy and latency cost of data movement between CPU and GPU — critical for memory-bound workloads like the brain atlas imaging that JUPITER handled. The same architecture forms the basis of NVIDIA's DGX GH200 systems deployed at major AI labs globally.

Why it matters

JUPITER's early results make a concrete case that exascale is a qualitative threshold, not just a benchmark number:

• Foundation model training for science: Training CytoNet on 6.5 petabytes in under five days changes the economics of building domain-specific foundation models at national labs. Work that would anchor a research program for months becomes a sprint.
• Operational km-scale climate modeling: The ICON run at 1km resolution is not a test workload — it is the resolution required for meaningful regional climate projections. Doing it in 24 hours of compute makes iterative ensemble modeling possible.
• Record quantum simulation: Surpassing the 48-qubit classical simulation record demonstrates the system's cross-domain utility; it is not a single-purpose AI machine.

For NVIDIA, JUPITER is the flagship European reference deployment for the Grace Hopper architecture. European science funding bodies and national HPC programs have historically been cautious about hardware monoculture; a system of this scale producing this breadth of results publicly is a significant data point for NVIDIA's pitch to European governments and research consortia evaluating Rubin-class infrastructure.