# Nemotron for Science puts agents inside the HPC loop _NVIDIA opened SC26 with local science agents, new Vera CPUs, Rubin HPC and quantum work with IBM._ **Published:** 2026-11-17 **Source:** https://www.startuphub.ai/ai-news/technology/2026/nemotron-for-science-puts-agents-inside-the-hpc-loop --- Ian Buck kicked off NVIDIA SC26 week in Chicago with agents that plan experiments, run code and check results on systems, [NVIDIA Blog](https://blogs.nvidia.com/blog/ian-buck-sc26/) reports. The claim is simulation stays on top, with AI, agents and quantum alongside it. Nemotron for Science pairs the Nemotron 3.5 Super VL open model with a new open chemistry dataset and post-training recipes, [NVIDIA](https://www.startuphub.ai/ai-news/public-companies/2026/spacex-openai-ipos-nvidia-earnings-fed-rate-hikes) says. The NVIDIA newsroom version adds scientific evaluations to that bundle for domain agents, the company says. Labs can run it on a DGX Spark, a workstation or a supercomputer, NVIDIA says. Buck, vice president of hyperscale and high performance computing, framed the payoff as local ownership: "You own your data. You own the model." Argonne put that loop to work. An agent built with the toolkit took one prompt, planned hypotheses and chose simulations, NVIDIA says. It then flagged two new binding sites. KTH Royal Institute of Technology fixed 81% of broken GROMACS setups in one overnight run on a single DGX Spark, NVIDIA says. NVIDIA ported a Weather Research and Forecasting routine to GPUs for a gain of more than 5x on about 150 million tokens, NVIDIA says. CPUs carry the agent loop by running tools between model steps, Buck said. DKRZ picked Vera CPUs for its HLRE-5 system expected in early 2028, NVIDIA says, joining systems at Texas Advanced Computing Center. Rubin HPC adds FP64 to FP4 support, sparse acceleration and seventh-generation NVLink, NVIDIA says, with more than 50,000 CUDA cores or about three times Hopper. StartupHub.ai data shows AMD, Intel and IBM among the tracked vendors in this comparison set, with no verified funding change attached to this announcement as of . That leaves AMD and Intel defending on-premise share with local inference stories until 2028 hardware lands. On quantum, a JUPITER run cut operations for one logical step tenfold using magic state cultivation, NVIDIA says. NVIDIA, IBM and Fermilab will pair DGX B200 systems with IBM quantum computers including Nighthawk, the companies say. A TCBP-JOBS decoder cut research compute from 12 weeks to 45 minutes, NVIDIA says, toward a latency goal from about 300 microseconds to 10. Photonics switches at Lambda show 15x better reliability on bit errors and link flaps, NVIDIA says. None are independent benchmarks yet. Reproducible GROMACS and WRF reruns outside NVIDIA systems would change the read. The open question is how many labs hit 81% repairs and 5x ports without that hardware. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.