Nvidia is pitching agentic AI as the engine for the next era of scientific supercomputing, backed by a new hardware and software stack it unveiled at ISC High Performance 2026 in Hamburg. The GPU giant claims its upcoming Vera Rubin platform, combined with new domain-specific tools, will power autonomous AI co-scientists capable of running experiments, simulations, and data analysis around the clock without human intervention.
"We are currently witnessing a massive inflection point with agentic AI. AI is shifting from a tool that simply answers questions to an autonomous system that executes complex tasks," said Dion Harris, Nvidia's senior director of HPC and AI Factory Solutions, in a media briefing ahead of the event.
Nvidia's argument is straightforward: scientific discovery at a scale humans cannot manage alone requires a new computing stack. That stack connects AI agents directly to simulators, surrogate models, and instruments. Harris described workflows where agents plan experiments, write code, run simulations, and fold data analytics into a single continuous process. The compute, memory, and networking demands of this, Nvidia says, are met by its Vera Rubin and Grace Blackwell architectures, plus Quantum InfiniBand interconnects.
The hardware: Vera Rubin and early deployments
The Vera Rubin NVL rack, slated for availability in Q4 this year, packs up to 144 GPUs per rack and delivers 5 petaFLOPS of FP64 floating-point performance. Memory bandwidth gets a 2.8x boost over Blackwell, using 41 TB of HBM4 memory per rack to hit three petabytes per second of bandwidth. Harris said many HPC workloads are memory-bound, making the bandwidth increase a central performance lever.
Three systems built on this platform were named at the conference. The Mission and Vision supercomputers at Los Alamos National Laboratory will be, according to Harris, the world's first agentic AI supercomputers. Mission houses 2,160 Rubin GPUs and 1,080 Vera CPUs, while Vision scales to 1,298 GPUs and 648 CPUs. A third system, Veritas, announced at ISC, deploys 576 Rubin GPUs and 288 Vera CPUs.
Software for agentic science
Nvidia also introduced three software tools designed to accelerate specific scientific domains. ALCHEMI targets chemical and material discovery, using a microservice called BGR to simulate millions of molecules and structures. DAQIRI connects next-generation scientific instruments directly to real-time AI inference. At CERN's ATLAS experiment, where less than 2 percent of collision data is typically stored, DAQIRI introduces a GPU-accelerated AI trigger pipeline that lets FPGAs handle low-latency routing while GPUs run deep learning models on far more data.
cuPhoton processes petabytes of camera and telescope data for astrophysics. In tests with 32 Grace Blackwell superchips simulating data from the Rubin Observatory, cuPhoton loaded and read images 15,000 times faster and accelerated signal processing and analysis by up to 8,000 times, Harris said.
Why agents in science?
When asked whether agentic AI is necessary for scientific research, Harris drew a distinction. "Agentic AI, or in fact any AI, is not required to do science," he told The Register. "But Nvidia believe agentic AI is already emerging as a powerful tool to do science at a scale that isn't possible when human scientists alone drive the process. Agents don't need to sleep, or eat, or take breaks. They can consume thousands or millions of technical papers and remember the details, and in some cases, they benefit from PhD-level understanding across diverse fields from astrophysics to zoology."
The vision is that human researchers will have teams of tireless AI agents running investigations continuously. These agents run on CPUs but access GPU-accelerated tools for peak performance, Harris added. Nvidia also pointed to Europe's HPC momentum, citing 35 new supercomputers brought online in the past year, all using its technology, including Jupiter, MareNostrum 5, Blue Swan, HammerHAI, and CINECA.
Why this matters for research scientists
For scientists in research roles, the shift Nvidia describes is not about replacing human curiosity but augmenting it with persistent, autonomous compute. The ability to chain AI reasoning, simulation, and instrument data into one workflow changes the bottleneck from human attention to infrastructure. Researchers planning experiments or analyzing large-scale data sets should watch how early deployments like LANL's Mission and Vision systems integrate these agentic pipelines, because the software and hardware stack that enables them will shape what's possible in grant proposals and lab architectures over the next two years. The AI Learning Path for Research Scientists offers a practical starting point for understanding how these autonomous workflows are built and deployed.
Your membership also unlocks:
