Nvidia used ISC 2026 to flog Vera Rubin as Europe’s new science engine, which sounds wonderfully sovereign until you read the vendor list.
At ISC High Performance 2026 in Hamburg on 22 June, Nvidia pitched Vera Rubin as a scientific computing platform, rather than just another cloud AI accelerator.
The GPU maker said 35 new AI supercomputers are being built across 23 European countries, making this its biggest scientific computing land grab in years.
This sounds a bit strange because EuroHPC was supposed to give Europe sovereign supercomputing away from US tech bros who can switch the whole lot off if their supreme dalek gets his knickers in a twist about them not supporting his latest war.
Nvidia says it now powers more than 90 per cent of Europe’s AI factory buildout. The Vera Rubin NVL4 scientific variant supports up to 144 GPUs per rack and is designed to run high-precision simulation and AI inference without data conversion faff.
A fully loaded rack delivers more than seven exaflops of AI performance and five petaflops of native FP64 computing, although those numbers measure different things.
Nvidia senior director of HPC and AI Factory Solutions Dion Harris said:“Native FP64 precision remains absolutely vital for accurate fluid dynamics, climate modeling, and geoscience. We are committed to maintaining that support moving forward.”
The big changes are NVLink 6, HBM4 memory and tighter CPU-GPU coupling through Nvidia’s 88-core Vera CPU and Rubin GPU.
NVLink 6 doubles GPU-to-GPU bandwidth to 3.6TB/s per GPU, while HBM4 lifts per-GPU memory bandwidth from 8TB/s to 22TB/s.
That should help memory-bound workloads such as computational fluid dynamics, molecular dynamics and lattice quantum chromodynamics, where Nvidia claims up to four times the performance of earlier systems.
System builders include Bull, the Grey Box Shifter, Dell., GIGABYTE, HPE and Supermicro announced custom Vera Rubin NVL4 racks, with OEM availability aimed at the fourth quarter of 2026.
Germany’s Leibniz Supercomputing Centre will put Vera Rubin into Blue Lion, an HPE Cray exascale-class machine due in 2027.
Barcelona Supercomputing Center will upgrade MareNostrum 5 with Nvidia GB300 NVL72 and GB200 NVL4 systems, while BavariaAI’s Blue Swan and Sweden’s Mimer AI Factory are joining the same gold rush.
Barcelona Supercomputing Centre director Mateo Valero Cortés said: “With the upgrade to MareNostrum5 and NVIDIA accelerated computing, the consortium composed of Spain, Portugal and Türkiye will make available to European researchers the tools to tackle some of the world’s most complex challenges, from climate modelling to biomedical discovery.”
The US national labs are piling in too, with Lawrence Berkeley’s NERSC, Los Alamos and other federally funded boffins committing Vera Rubin systems for simulations, AI training and scientific discovery.
Nvidia is pitching this as agentic AI for science, where systems chew through papers, run tasks and make researchers wonder how much of the lab is still theirs.
Harris said: “Agentic AI is already emerging as a powerful tool to do science at a scale that isn’t possible when human scientists alone drive the process. Agents don’t need to sleep, or eat, or take breaks. They can consume thousands or millions of technical papers and remember the details.”
The open-source HPC crowd now has to update Spack, Kokkos, MPI and OpenMP runtimes for NVLink 6, HBM4 and the new coherent CPU-GPU fabric. That is the awkward bit, because NVLink is proprietary and non-Nvidia silicon cannot use it, leaving Europe’s “sovereign” science workflows cuddled up to a single supplier.
Nvidia chief executive Jensen Huang said: ““AI is the new instrument of science, and Europe is building the infrastructure to put it in the hands of millions of researchers.”
