NVIDIA’s Vera CPUs earmarked for new Los Alamos supercomputers to run “agentic” scientific AI

NVIDIA says Los Alamos National Laboratory (LANL) is building a trio of new supercomputers with HPE and NVIDIA hardware, targeting faster scientific discovery and a new class of software it describes as “agentic AI for science.” The announcement centers on LANL’s planned use of NVIDIA Vera CPUs as a foundational compute engine inside the systems the lab calls Mission, Vision and Veritas.

According to NVIDIA, the systems will rely on HPE Cray Supercomputing GX5000 architecture paired with the NVIDIA Vera Rubin platform. That combination is intended to bring together Vera CPUs, Vera Rubin GPUs and NVIDIA Quantum-X800 InfiniBand networking, which is the high-speed interconnect used to link compute nodes for large-scale simulations.

Within the planned configurations, NVIDIA says Mission will include NVIDIA Vera Rubin GPU nodes plus 2,300 standalone NVIDIA Vera CPUs deployed using the HPE Cray Supercomputing GX240 blade. Veritas, which the company describes as arriving alongside Mission and Vision, is expected to include approximately 1,150 standalone NVIDIA Vera CPUs to complement the Vera Rubin-based nodes.

NVIDIA frames Veritas’s role as support for LANL’s Laboratory Directed Research and Development (LDRD) program, a mechanism through which the lab funds internal research initiatives. The company says the planned system will also test technologies intended for use in larger LANL systems later on, focusing specifically on “agentic AI for science” workflows rather than just conventional simulation throughput.

A key part of the narrative is LANL’s move toward AI agents that can take actions in a scientific workflow. NVIDIA says the lab is adding tools where software agents can form hypotheses, choose appropriate tools, launch simulations, analyze outputs and then refine the next steps, in a feedback-driven loop aimed at reducing time from idea to result.

NVIDIA points to LANL’s public work on URSA, described as the Universal Research and Scientific Agent. NVIDIA says URSA runs on Venado and will soon extend onto Mission and Vision. URSA is characterized as a modular, feedback-driven framework intended to help scientists brainstorm hypotheses, plan experiments, run simulations and analyze results, effectively turning parts of the research process into repeatable, agent-managed steps.

On performance, NVIDIA says it demonstrated “7x higher performance” on URSA workloads using a Vera CPU versus the CPUs in LANL’s Crossroads x86 supercomputer. In a separate test, NVIDIA says Vera outperformed the Crossroads x86 CPUs by more than 3x on Branson, an open source Monte Carlo heat transfer simulation tool.

The company attributes those results in part to Vera’s design, including what it calls a custom Olympus core, LPDDR5 memory and a fast on-chip fabric. NVIDIA also offers a broader comparison, saying a single Vera CPU outperforms a single socket x86-based CPU by more than 3x and provides more than 4x the memory per core and 6x the memory per node, which it links to faster paths to scientific answers.

For LANL’s computing program timeline and national-security context, NVIDIA says Mission is expected to be operational in 2027 and will be the fifth Advanced Technology System in the National Nuclear Security Administration’s Advanced Simulation and Computing program, replacing Crossroads for classified national security workloads. Vision is also expected to be operational in 2027 and is described as a resource for fundamental science, including materials and nuclear science, energy modeling, biomedical research and AI, with an emphasis on allowing more scientists to run methods and explore ideas before moving into higher-consequence tasks. The company also says the three systems build on LANL’s Venado platform, installed in 2024 with NVIDIA GH200 Grace Hopper Superchips and NVIDIA Grace CPUs.

Still, some details remain open in NVIDIA’s announcement. While the company quantifies several performance comparisons on named workloads and tools, it does not provide full system capacity figures, power or cost information, or specific software benchmarks beyond URSA and Branson. It also does not describe how agentic AI orchestration will be integrated with the lab’s security, data governance and validated simulation pipelines, which can be as consequential as raw compute.

What to watch next is whether LANL and NVIDIA publish additional workload results as Mission, Vision and Veritas move from planned configurations toward deployment. Particularly important will be evidence that the “agentic” loop meaningfully reduces researcher time to insight, not just time to compute, and that the systems deliver consistent performance across a broader set of simulation and analysis applications.