
16 Nov 2026
NVIDIA's Rubin HPC packs 50,000+ CUDA cores for the simulations AI agents run
NVIDIA used the SC26 supercomputing conference on Monday to unveil Rubin HPC, a GPU tuned to keep a record number of cores fed with scattered scientific data. The pitch: AI agents will run so many simulations that simulation becomes the bottleneck.
Not every new NVIDIA chip is for chatbots. Rubin HPC is a bet that the next compute crunch is AI agents running physics, chemistry and engineering simulations at scale. It also signals the roadmap is moving on schedule despite the turmoil at the top.
On Monday, Nov. 16, at SC26, the annual supercomputing conference, NVIDIA announced Rubin HPC. It is a new GPU in the company's Vera Rubin platform, built to speed up simulation and data processing for science, engineering, and analytics. A GPU is the graphics chip that does the heavy math. Those lines are NVIDIA's.
The specs NVIDIA disclosed: more than 50,000 CUDA cores on a single GPU, which NVIDIA calls a record; a "sparse data acceleration" capability that keeps those cores fed with data; and seventh-generation NVLink for connections from one GPU to another. CUDA cores are the GPU's general-purpose math units. NVLink is NVIDIA's high-speed chip-to-chip link. Sparse data, here, means the values a program needs are scattered through a huge file, not lined up in one neat block. Those figures and that name are NVIDIA's.
The design pitch. Scientific workloads keep reaching into huge datasets for scattered values. NVIDIA's examples are climate models, engineering simulations, genomic analysis, and financial risk. Genomic analysis is the work of reading genetic data. Rubin HPC, NVIDIA says, scales how data moves together with the computing, instead of adding computing alone. NVIDIA says that speeds sparse matrix operations, stencil computations, sorting, and reductions. A sparse matrix is a grid of numbers that is mostly empty. A stencil computation updates each point from its neighbors, the way a weather model does. A reduction folds many numbers into one, such as a sum. Those lines are NVIDIA's.
The agent angle, as NVIDIA states it. Agents exploring candidate materials, molecules, and designs drive more demand for simulation. An agent, here, is software that can take a next step, not only draft a reply. NVIDIA paired Rubin HPC with expanded CUDA-X libraries those agents can call. CUDA-X is NVIDIA's set of libraries for that kind of math. That pairing is NVIDIA's.
Ian Buck, NVIDIA's vice president of hyperscale and HPC, said: "For agents tackling science and engineering, faster simulation means faster discovery." HPC is high-performance computing, the work of the big scientific machines. That quotation is his, in NVIDIA's announcement.
Availability: NVIDIA says Rubin HPC is expected in 2028. The company gave no price.
Same-day context: NVIDIA also introduced Nemotron for Science at SC26, an open model plus a chemistry dataset for science agents, and the chip news lands the day before NVIDIA's Tuesday earnings, with the company under an interim chief executive after Jensen Huang's resignation.
The picture is NVIDIA's official image of the Rubin HPC GPU package on a black background. The credit is NVIDIA.
In plain terms, NVIDIA used a supercomputing conference to show a chip built for scientific simulation, not for chat. It says the chip has more than 50,000 CUDA cores, a record on the company's count, plus a way to keep those cores supplied with scattered data, and a new generation of its chip-to-chip link. The company says AI agents will want more of this kind of simulation. The chip is not for sale. NVIDIA says 2028, and it did not name a price.
RELATED
Sources
- NVIDIA
blogs.nvidia.com
- NVIDIA SC26 news
blogs.nvidia.com
- Tom's Hardware
tomshardware.com



















