NVIDIA at ISC 2026: Exascale Supercomputing and AI-for-Science Announcements

On June 22, 2026 — the opening day of ISC High Performance in Hamburg — NVIDIA published four coordinated blog posts covering its AI-for-science agenda across three fronts: production science results from JUPITER, Europe's exascale system built on Grace Hopper Superchips at Jülich [1]; a LANL supercomputing procurement adding a third system (Veritas) alongside the previously announced Mission and Vision, with Vera CPUs and Rubin GPUs targeting 2027 [2]; and expanded NAIRR pilot contributions spanning over 700 U.S. research projects [5]. A fourth post detailed new scientific computing software with large benchmark claims across astrophysics, materials science, and particle physics [6].

NVIDIA is presenting exascale compute as now delivering production science results — climate records, quantum simulation milestones, brain modeling — not future capability. The LANL procurement and federal NAIRR backing represent sustained U.S. national-lab commitment to NVIDIA's hardware stack at a moment when AMD is also winning large supercomputer contracts.

NVIDIA opened ISC 2026 with a coordinated set of announcements on June 22 covering four distinct fronts: the JUPITER exascale system in Europe, the Los Alamos National Laboratory supercomputing procurement, the NAIRR pilot program, and new AI-for-science software. All four announcements originate from NVIDIA's own blog and present partner and customer results without independent benchmarking. ISC runs through June 26, so further announcements from competitors and third parties may follow.

JUPITER, operated by Jülich Supercomputing Centre and built on NVIDIA Grace Hopper Superchips, is the centerpiece of the European narrative. Three headline scientific results are cited: CytoNet, a foundation model for brain microarchitecture trained on 6.5 petabytes of data from 21 post-mortem brains, completed training in under five days using 4,096 Superchips; the ICON climate model won the Gordon Bell Prize for Climate Modelling by simulating a fully coupled Earth system at 1-kilometer global resolution, running 146 simulated days in 24 hours of compute on 20,480 Grace Hopper Superchips; and Jülich set a quantum simulation record by fully simulating a 50-qubit universal quantum computer, exceeding the previous 48-qubit record by exploiting coherent CPU-GPU memory in GH200 chips. Ericsson is also using JUPITER to develop energy-efficient AI for 6G networks [1].

At Los Alamos, NVIDIA announced that a third system — Veritas — joins the previously announced Mission and Vision supercomputers, all using HPE Cray GX5000 hardware combining Vera CPUs, Rubin GPUs, and Quantum-X800 InfiniBand, with 2027 operational targets [2][3]. NVIDIA claims the Vera CPU delivers 7x better performance on LANL's URSA agentic AI workloads compared to the x86-based Crossroads supercomputer, and 3x better on Monte Carlo simulations with 4x more memory per core [2]. URSA is a modular, feedback-driven agent system designed to autonomously run scientific workflows — forming hypotheses, launching simulations, and analyzing results [2]. Tom's Hardware noted the announcement follows recent AMD wins in the supercomputer market [4].

The NAIRR pilot has supported over 700 U.S. research projects over two years, with NVIDIA providing DGX node access and technical support. Highlighted outcomes include Polymathic AI's publicly released Walrus foundation model for fluid-like physical simulations, the University of Michigan's MIST molecular models matching or exceeding state-of-the-art across 400+ structure-property benchmarks, and Boston University's BEACON pipeline reducing infectious disease outbreak report generation from several hours to roughly two minutes [5]. On the software side, new tools include cuPhoton (claiming 14,900x faster astronomical image loading for the Rubin Observatory LSST on GB200 NVL72), CERN's A-GHOST project using DAQIRI for real-time AI on previously-discarded collision data, and ALCHEMI-optimized kernels for materials screening with claimed 50x acceleration [6].