Nvidia and HPE will build three new supercomputers at Los Alamos National Laboratory — Mission, Vision and Veritas — anchored on the new Nvidia Vera CPU paired with Rubin GPUs and Quantum-X800 InfiniBand networking. Mission alone will pack 2,300 standalone Vera CPUs alongside Vera Rubin GPU nodes, while Veritas will add roughly 1,150 more Vera CPUs to complement its Rubin partition. In early benchmarks at the lab, Vera outran the x86 CPUs inside the existing Crossroads supercomputer by 7x on agentic-science workloads.
Mission and Vision are slated to be operational in 2027. Mission becomes the fifth Advanced Technology System in the National Nuclear Security Administration's Advanced Simulation and Computing program and will replace Crossroads for classified national-security workloads. Vision will handle open scientific research — materials science, nuclear science, energy modeling, biomedical work and AI — and Veritas will sit alongside both, serving the Laboratory Directed Research and Development program as a smaller proving ground for the same stack.
All three systems use the HPE Cray Supercomputing GX5000 architecture on the Nvidia Vera Rubin platform, with Mission and Veritas built on the GX240 blade. That gives Los Alamos a single hardware substrate spanning classified simulation, open science and exploratory AI research — a notable consolidation for a lab that has historically run distinct architectures for each mission.
Key facts
- 01Mission will ship with 2,300 standalone Vera CPUs alongside Vera Rubin GPU nodes; Veritas adds roughly 1,150 more.
- 02Vera posted 7x higher performance on URSA agentic-science workloads than the Crossroads x86 CPUs.
- 03On the Branson Monte Carlo heat-transfer simulation, Vera beat Crossroads CPUs by more than 3x.
- 04A single Vera CPU delivers 4x more memory per core and 6x more memory per node than an x86 socket.
- 05Mission and Vision are scheduled to be operational in 2027, replacing Crossroads for classified national-security work.
The headline pitch is agentic AI for science: software systems that propose a hypothesis, pick a tool, launch a simulation, read the results and decide what to run next. Los Alamos has been working in public on URSA, its Universal Research and Scientific Agent, which already runs on the lab's Venado system installed in 2024 and is being moved to Mission and Vision. On URSA workloads, Vera delivered the 7x lift over the Crossroads x86 CPUs that Nvidia is leaning on as its marquee number.
“Researchers are adding a new tool for science with AI agents that can form hypotheses, choose tools, launch simulations, analyze outputs and refine the next step.”— Chris Porter, NVIDIA
Simulation performance tracks the same direction. In early testing on Branson, an open-source Monte Carlo heat-transfer code, Vera ran more than 3x faster than the Crossroads x86 CPUs. Nvidia attributes the gains to Vera's custom Olympus core, LPDDR5 memory and on-chip fabric. A single Vera CPU outperforms a single-socket x86 chip by more than 3x while carrying 4x more memory per core and 6x more memory per node — the kind of memory ratio that matters when an agent is juggling intermediate state across a long simulation run.
Vera is the successor to Nvidia Grace, and the Los Alamos relationship has run through both generations. Venado, the HPE Cray EX system installed at the lab in 2024, runs on Nvidia GH200 Grace Hopper Superchips and standalone Grace CPU Superchips, and it is the system where URSA was first deployed in production. Mission, Vision and Veritas extend that arc by more than a decade of CPU codesign work between the lab and Nvidia, aimed specifically at LANL's simulation codes rather than generic benchmark suites.
The three machines also represent a strategic statement about the CPU side of the Vera Rubin platform. Most coverage of Rubin has focused on the GPUs, but Los Alamos is buying thousands of standalone Vera CPUs as compute citizens in their own right — not just as host processors feeding accelerators. That matters for workloads where the bottleneck is memory bandwidth and tightly coupled simulation, rather than dense matrix math.
This continues a recent run of Nvidia science-infrastructure wins. AI Chat Daily has covered the company's CUDA-X science software posting a 14,900x speedup on telescope data, the NAIRR research pilot crossing 700 projects on DGX hardware, and JUPITER becoming Europe's first exascale supercomputer on Grace Hopper. The Los Alamos deal pushes the same playbook into classified national-security work, which is a harder sell on procurement and certification timelines.
What is not yet public: pricing for the three systems, the exact Rubin GPU count per node, the delivery sequence between Mission and Vision, or the power envelope for the full Mission build. Nvidia and HPE have also not detailed how URSA's agent loop will be sandboxed when it runs against classified data on Mission versus open science on Vision, which is the kind of question the NNSA will eventually need to answer in writing.
For Nvidia, the deal cements Vera as the CPU half of the Rubin generation in one of the most prestigious computing customers in the United States. For Los Alamos, it gives the lab a single agentic-AI-capable platform spanning classified simulation and open research years before most national labs will have anything comparable. And for the broader AI-for-science thesis, it puts a real number on agentic workloads — 7x on URSA, 3x on Branson — at a scale where the comparison is to a working DOE supercomputer rather than a synthetic benchmark.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
