The exascale offensive: America's race to rule AI HPC (www.theregister.com)

The Department of Energy announced a sweeping build-out of nine AI-optimized supercomputers across Argonne, Oak Ridge and Los Alamos national labs, delivered through unprecedented public–private partnerships with Nvidia, AMD and HPE. Argonne will host five machines including Solstice (reportedly 100,000 Nvidia Blackwell GPUs) and Equinox plus Minerva, Tara and Janus for specialized modeling and workforce development. Oak Ridge will add Lux (AMD Instinct MI355X GPUs + EPYC CPUs, early 2026) and Discovery (HPE Cray GX5000 with next‑gen EPYC Venice CPUs and Instinct MI430X GPUs, ~2028) — the latter expected to eclipse Frontier’s performance and cross exascale thresholds. Los Alamos gets Mission and Vision (HPE/Nvidia) for nuclear‑security simulation and open science. New architectures — Nvidia’s Vera Rubin (Vera CPU + Rubin GPU), next‑gen AMD silicon, and Quantum‑2/X800 InfiniBand networks — emphasize heterogeneous hardware and mixed‑precision AI paths that claim orders‑of‑magnitude gains (vendors cite multi‑exaflop AI throughput). The push is explicitly AI‑centric: these machines are meant to fuse traditional simulation with large‑scale ML for climate, materials discovery, fusion, biomedical research and stockpile stewardship, while shoring up U.S. strategic advantage against China and Europe. Technically, the program accelerates adoption of specialized accelerators, tighter CPU–GPU integration, and low‑precision AI kernels at unprecedented scale — but it also highlights system‑level challenges (storage, software stacks, secure open tooling and workforce readiness). In short, the DOE drive aims not just to chase raw FLOPS but to make exascale the backbone of “AI‑enabled science,” with major geopolitical and scientific ramifications.