Table of contents
Open Table of contents
Los Alamos Bets on Vera CPUs for Agentic Science
Category: Infrastructure and Hardware Source: NVIDIA Blog
Los Alamos National Laboratory said its upcoming Mission, Vision, and Veritas supercomputers will use NVIDIA’s Vera CPU alongside Rubin GPUs and Quantum-X800 networking, with Mission and Vision expected in 2027. The headline claim is not just more accelerator capacity, but a system design tuned for simulation-heavy scientific AI workflows, including LANL’s URSA research agent framework.
The notable detail is the emphasis on CPUs, not only GPUs. LANL says Vera delivered 7x higher performance than the CPUs in its Crossroads x86 system on URSA workloads, and more than 3x on an early Monte Carlo heat transfer test. If those gains hold in production, this matters because many “agentic AI for science” loops still depend on orchestration, simulation, memory bandwidth, and tool execution that don’t map cleanly to GPU-only scaling.
What changed today is that NVIDIA’s scientific AI pitch got a concrete national-lab anchor with named systems, deployment timelines, and benchmark claims tied to actual research software. That is stronger signal than generic “AI for science” positioning. It also reflects a broader shift in HPC procurement: labs are no longer treating models as separate from simulation stacks; they’re buying for hybrid workloads where planning agents, numerical codes, and analysis pipelines run as one system.
What to watch next: independent evidence beyond vendor benchmarks, especially how URSA-like workflows perform end to end once these systems are live. The bigger question for practitioners is whether this architecture meaningfully reduces time-to-discovery, not just component-level runtime.
NAIRR Shows Shared AI Compute Is Turning Into Research Output
Category: Infrastructure and Hardware Source: NVIDIA Blog
The U.S. National Artificial Intelligence Research Resource pilot has now supported more than 700 projects over two years, according to NVIDIA’s update on its contribution to the program. The most useful signal here is that NAIRR is moving past access rhetoric and producing visible work across domains, including molecular modeling, fluid dynamics, and infectious disease monitoring.
Examples cited include Polymathic AI’s Walrus model for fluid-like behavior, the University of Michigan’s MIST molecular models for energy-storage materials, and Boston University’s BEACON outbreak-monitoring pipeline. None of this proves breakthrough science on its own, but it does show that giving researchers time-bound access to serious compute can produce diverse, domain-specific model development rather than just demos.
Why it matters now: policy conversations around national AI infrastructure often get stuck between frontier-model geopolitics and academic scarcity. NAIRR is one of the clearer signs that mid-scale, shared access can still matter operationally, especially for researchers who do not control hyperscaler budgets. For engineering leaders, it is also a reminder that research value often comes from well-supported access and onboarding, not only from the absolute largest cluster.
The next thing to watch is durability. Pilots are useful; sustained public-interest AI infrastructure requires predictable funding, transparent allocation, and evidence that outputs justify the spend.
Quick Updates
-
Wave energy gets pitched as future AI infrastructure power: Eco Wave Power outlined pilot work using NVIDIA-backed digital twins and AI optimization to connect wave energy systems to coastal compute infrastructure, including a Los Angeles port project testing whether wave power can run a data center without the grid. Interesting as an energy-adjacent experiment, but still far from a proven answer to data center-scale demand. Source: NVIDIA Blog
-
A pragmatic case for AI-assisted code review on large diffs: A developer essay argues reviewers should stop line-checking huge AI-generated changes and instead use models to scan for implementation issues while humans focus on “out-of-distribution” knowledge: architecture decisions, local conventions, deprecations, and high-level design risk. Sensible for fast-moving software teams; less applicable where every line is safety-critical. Source: Simian Words