NVIDIA Vera CPUs land at Los Alamos to power three new future supercomputers - Yahoo Tech

Vera's deployment at Los Alamos is not a product milestone—it is a structural encirclement of the HPC processor market. For two decades, AMD EPYC and Intel Xeon duopolized supercomputing CPUs while NVIDIA plugged in solely as a GPU accelerator. Vera shatters that division: NVIDIA now owns the full CPU+GPU+NVLink+InfiniBand stack, and CUDA's software moat makes switching costs far exceed the silicon itself. The technical ripple is immediate. Three supercomputers running Vera will force OpenMPI, HPC compilers, and memory schedulers to be re-optimized for NVIDIA's custom microarchitecture. That engineering burden becomes a competitive moat—exactly the ecosystem lock-in playbook NVIDIA has perfected since 2007. Upstream, Vera almost certainly still depends on TSMC (Taiwan, China) for leading-edge fabrication, a single-point supply-chain vulnerability under tightening export controls. Downstream, DOE's concentrated procurement effectively crowns NVIDIA the default supplier. AMD's most probable counter: accelerate EPYC Turin plus MI300X bundling to match the full-stack pitch. Intel may pull Falcon Shores forward. But the 2019 DOE shift from Intel to AMD took three years; reversing a software-stack migration is equally glacial. Twelve-to-twenty-four-month outlook: by 2026, at least two non-US national labs will follow suit. The HPC CPU landscape shifts from a duopoly to a structural NVIDIA-full-stack-versus-AMD/Intel-combo confrontation.