$ today_in_10_seconds
Nvidia just pitched a supercomputing GPU as agent infrastructure. Rubin HPC packs more than 50,000 CUDA cores and is aimed at scientists running agent swarms that hypothesize, simulate, and refine in a loop. It doesn't ship until 2028, and the specs that matter most are still missing.
$ lead_story
Nvidia Unveils Rubin HPC, a 50,000-Core GPU for Agentic Science
At SC26, Nvidia announced Rubin HPC, which it says is meant to accelerate agentic workflows and simulation for scientists, engineers, and analysts working with large data sets. The core count roughly doubles today's chip: the current Rubin GPU has 224 SMs and 28,672 CUDA cores, so a jump of this size likely requires more dies on the package. The agentic framing is the interesting part. Nvidia is selling this to researchers marshaling agent swarms, using CUDA-X libraries to shorten the loop from hypothesis to simulation to refinement. If agents are the ones running experiments, simulation throughput becomes tool-call latency, and that's what this chip is built to cut.
// why_it_matters: In agentic science, the simulation is the tool call, and Nvidia wants to own it.
$ also_worth_attention
The software play: new CUDA-X libraries for science
Nvidia also introduced several new additions to CUDA-X, the library suite it ships to ease scientific research projects. For agent builders, this is the more practical news. Optimized, domain-specific libraries are the tools a science agent calls, and better tools matter more than more cores for most teams in 2026.
link →What Nvidia didn't say
Memory, FP64, power, and pricing are all undisclosed. That's a real gap: AMD and Intel are touting native FP64 rebalancing at the same conference, and double precision is what serious simulation runs on. A high core count without an FP64 story is a marketing headline, not a purchasing decision.
link →The interconnect and the timeline
Nvidia claims 3x the performance of Blackwell and will link chips with NVLink 7, which past disclosures suggest keeps 3.6 TB/s of per-GPU bandwidth while enlarging the scale-up domain. But Rubin HPC arrives sometime in 2028, so anything you build for agentic simulation this year runs on current hardware.
link →
$ bigger_picture
The trend behind the headlines
Chipmakers are now labeling silicon by agent workload. Nvidia has also positioned its Vera CPU for agentic workloads, and Google has reworked TPU silicon into specialized systems for agents and security. Rubin HPC adds a third category: hardware for the tools agents call, not the model doing the reasoning. In scientific work, an agent might spend a few seconds generating tokens and hours waiting on a simulation. That makes tool execution, not inference, the bottleneck. The infrastructure race is moving from 'fast tokens' to 'fast everything the agent touches.'
NVIDIA Warp
NVIDIA Warp is an open-source Python framework for writing GPU-accelerated, differentiable simulation code. It's a good fit if you want an agent to write, run, and iterate on physics or geometry simulations today without hand-writing CUDA, and it runs on current Nvidia GPUs.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime