NVIDIA's Grip on Scientific Computing Tightens as Europe and U.S. Build Exascale Supercomputers Around Its Hardware

This week at the ISC conference in Hamburg, NVIDIA's dominance in scientific computing reached a new inflection point: JUPITER, Germany's Forschungszentrum Jülich supercomputer, went live as Europe's first exascale machine running almost entirely on NVIDIA silicon—specifically Grace Hopper Superchips paired with Quantum-X800 InfiniBand networking. Simultaneously, results from the U.S. National Science Foundation's National Artificial Intelligence Research Resource (NAIRR) pilot program revealed that over 700 research projects across protein prediction, materials simulation, and disease modeling have been powered by NVIDIA infrastructure over the past two years. These aren't isolated wins; they represent a structural shift in how governments and institutions fund and build scientific computing capacity. The convergence matters because it signals that NVIDIA's ecosystem—hardware, networking, and increasingly proprietary software libraries like DAQIRI and ALCHEMI—has become the assumed baseline for exascale research globally.

The economics driving this consolidation are compelling. NVIDIA's new 45-degree-Celsius liquid cooling systems promise substantial energy cost reductions at scale, a critical advantage when operating exascale facilities that consume megawatts continuously. JUPYTER itself demonstrates the appeal: using Grace Hopper and Quantum networking meant scientists could consolidate compute density without custom engineering. Yet this efficiency comes with a cost rarely discussed publicly: institutional dependence. Los Alamos National Laboratory's three new HPE Cray supercomputers—Mission, Vision, and Veritas—will run on NVIDIA Vera CPUs, further embedding the company's architecture into classified research workflows. When a lab's entire software stack, from cuBLAS to custom CUDA-optimized kernels developed over years, runs on NVIDIA, switching vendors becomes economically and technically prohibitive, even if AMD's EPYC or Intel's latest Xeon processors offer comparable performance. The switching costs are immense.

Competitive alternatives exist in name only. AMD's MI300 accelerators target some HPC workloads, and Intel maintains R&D programs, but neither vendor has invested in the software ecosystem—the CUDA libraries, optimized compilers, proven workflows—that makes NVIDIA frictionless for scientists. The NAIRR program, while nominally vendor-agnostic, essentially standardized on NVIDIA because that's where the mature tooling lives. This isn't malice; it's path dependency. But as exascale computing becomes critical infrastructure for everything from climate modeling to drug discovery, the concentration of compute, networking, and software control in one vendor's hands deserves scrutiny. NVIDIA's engineering is genuinely excellent, but excellence without competition is not a sustainable arrangement for scientific research.