NVIDIA unveils AI agents and Rubin HPC GPU at SC26
At SC26 in Chicago, NVIDIA announced AI agents for scientific research, a Rubin HPC GPU and a wider Vera CPU platform. The company also outlined quantum computing work with IBM and Fermilab.

Key takeaways
- NVIDIA announced Nemotron for Science, a package for running AI agents on researchers' own systems.
- The company gave examples at Argonne, KTH and the SETI Institute, including an agent that repaired 81% of GROMACS test cases.
- NVIDIA also announced the Rubin HPC GPU, a wider Vera CPU platform and quantum work with IBM and Fermilab.
- Rubin HPC and DKRZ's Vera-based HLRE-5 system are both expected in 2028.
In this article
What NVIDIA announced
NVIDIA used its SC26 event in Chicago on Monday to present new NVIDIA AI agents for scientific research, along with a Rubin HPC GPU and a wider Vera CPU platform. Ian Buck, a vice president at NVIDIA with responsibility for hyperscale and high-performance computing, led the address, according to the NVIDIA Blog. The company also announced quantum computing work with IBM and Fermilab.
NVIDIA's central claim is that agents can sit between researchers and their machines. The software is meant to design trials, execute code and verify outcomes on a researcher's own systems. Buck described agents as a major change in how people work with supercomputers, instruments and devices. The NVIDIA Blog reports that the tools are intended to run locally, not only in a remote cloud service.
Nemotron for Science is the model package at the centre. It pairs the open Nemotron 3.5 Super VL model with a new chemistry dataset and tuning recipes for chemistry and life sciences. NVIDIA says researchers can run it on a DGX Spark desktop, a workstation or their own supercomputer.
Agents that plan and check experiments

NVIDIA gave several examples of agents in use. At Argonne National Laboratory, a toolkit-built agent received one instruction, formed and tested hypotheses, selected simulations and found two new binding sites, according to the company. At KTH Royal Institute of Technology, an agent repaired broken GROMACS setups. During an overnight session on one DGX Spark, it repaired 81% of the cases it was given.
At the SETI Institute's Allen Telescope Array, an agent was given a single line of instruction to chart a methanol spectral line. It then adjusted for Doppler shift by itself, NVIDIA says. The company also introduced an Nsight AI Porting Agent, which moved a Weather Research and Forecasting model routine onto GPUs. NVIDIA says the port ran more than five times faster and used about 150 million tokens, the units that models process.
Buck explained that agents depend on CPUs between model steps. A processor has to launch tools, carry out operations and collect results while the model decides what to do next. That argument underpins NVIDIA's wider CPU push.
Vera CPU and Rubin HPC GPU
NVIDIA is expanding its Vera CPU platform for high-performance computing. DKRZ, the German Climate Computing Center, has chosen Vera CPUs for its HLRE-5 supercomputer, which is expected in early 2028. The system joins other Vera machines already announced at the Texas Advanced Computing Center.
The Rubin HPC GPU sits in NVIDIA's Vera Rubin platform. NVIDIA positions it for scientific work, engineering, data analytics and AI. It handles precision formats from FP64 through FP4, includes acceleration for sparse data and uses seventh-generation NVLink. It carries over 50,000 CUDA cores, roughly triple the count in Hopper, and is expected in 2028.
Quantum work with IBM and Fermilab
NVIDIA, IBM and Fermilab announced plans for a quantum computing co-design platform. The platform is intended to support research into error correction and applications. NVIDIA and IBM plan to pair DGX B200 systems with IBM quantum computers, including machines powered by IBM Quantum Nighthawk.
Buck also described quantum decoding work. Work at Jülich's JUPITER system demonstrated that magic state cultivation can reduce the effort for one logical operation by a factor of 10, according to Buck. The technique is being added to CUDA-Q. A new CUDA-Q decoding algorithm, TCBP-JOBS, shortens a particular error-correction research calculation from 12 weeks to 45 minutes, according to NVIDIA. The company says its goal is to bring decoding delay down from roughly 300 microseconds to 10.
Separately, Lambda is running an early installation of NVIDIA Photonics switches that use co-packaged optics. NVIDIA says the switches are 15 times more reliable when it comes to bit errors and link flaps, short connection interruptions.
Why the science push matters
Much of the address framed simulation as the enduring base of scientific computing. Buck said AI, quantum methods and agents would sit alongside simulation rather than replace it. He also looked back to his first public CUDA presentation, given at an SC06 tutorial in Tampa, Florida, two decades ago. NVIDIA now offers more than 1,000 CUDA-X libraries, he said.
The announcements matter because they show NVIDIA trying to own more of the scientific computing stack, from CPUs and GPUs to models, agent software and quantum tools. For research teams, the appeal is local control: agents that can work on their own systems and data. The performance figures are company claims and have not been independently confirmed. NVIDIA has given 2028 as the expected date for Rubin HPC and for DKRZ's HLRE-5 system.
Sources
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



