
On November 17, 2026, NVIDIA vice president Ian Buck opened the company's SC26 week in Chicago with three announcements: Nemotron for Science, a Rubin-based HPC GPU, and an expanded Vera CPU platform. The science agent toolkit lets researchers plan experiments, execute code and verify results on their own systems rather than in a vendor cloud. Liu Gong believes the move is less about selling more accelerators than about embedding an agent framework into the HPC software stack, so software ties
On the morning of November 17, 2026, NVIDIA kicked off its week at SC26 at McCormick Place in Chicago, with Ian Buck, vice president and general manager of compute products, delivering the special address. The talk centered on three releases. Nemotron for Science, built on the Nemotron model family, is an agent toolkit that can plan experiments, run code and check results, and it runs on researchers own systems instead of a vendor-hosted cloud. The Rubin HPC GPU follows the Blackwell generation with the Rubin architecture aimed at high-performance computing workloads, pairing HBM4 high-bandwidth memory with NVLink-C2C links to NVIDIA homegrown CPU. The expanded Vera CPU platform scales the Arm-based processor to larger node and rack configurations. Buck framed the combination as infrastructure for what he called the age of agents in science, and the rest of NVIDIA SC26 schedule is
The SC conference series has long served as the launch window for HPC hardware and networking, and for three years running its agenda has been dominated by AI. NVIDIA has kept a roughly two-year cadence for data-center GPUs since Hopper: Blackwell reached volume in 2024, and Rubin lands in 2026 on the public roadmap, which lines up with SC26 and with a new procurement cycle at national HPC centers. Two forces make a science agent viable now. First, scientific AI is moving from discriminative models to executable workflows; materials, biology and climate teams no longer settle for a prediction and want experiment plans generated and verified automatically. Second, data-sovereignty rules have
Liu Gong's take: within the next 12 months, a locally runnable science agent will stop being bundled software and become a scored line item in HPC tenders, and NVIDIA is after the right to define that standard rather than a few extra board sales.
From a technical standpoint, the operative word in Nemotron for Science is verification, not generation. Planning an experiment, executing code and checking the result form a closed loop that must happen on one machine, which separates it from a chat-style cloud assistant. In industrial visual inspection, Liu Gong has seen that the expensive part of a production line is never inference speed but re-judgment and traceability: after a vision model goes live on an SMT line, engineers spend roughly 60% of their hours attributing false positives and misses. Science works the same way, where an agent that runs the code and reports why an experiment failed is worth more than one that writes an elegant plan. Whoever standardizes the verification interface holds the entry point to the next HPC software stack.
On the business side, this is ecosystem reinforcement. NVIDIA public roadmap puts Vera Rubin NVL144 at roughly 3.6 exaflops of FP4 inference, about 3.3 times GB300 NVL72, while FP64 gains are far smaller. The pitch is shifting from double-precision peak to low-precision agent throughput, a paradigm shock for national labs that still benchmark procurement on FP64 and a hard gap for rivals competing on FP64 alone. The 88-core Vera CPU then squeezes x86 vendors further out of HPC nodes, moving NVIDIA from selling accelerators to selling whole architectures.
For the industry, the first pressure shows up in scheduling. Slurm-style schedulers bill by node and core-hour, while agents bill by a completed task loop; changing the billing unit changes budget models and acceptance criteria at HPC centers. Liu Gong predicts that by the first half of 2027, the first pilot contracts settled on agent task completion rates will appear at major HPC centers. The risks are equally clear: permission boundaries, code-execution sandboxes and reproducibility have no third-party certification yet, and liability for a wrong experimental conclusion will be the buyer most practical concern.
- NVIDIA public roadmap lists Vera Rubin NVL144 at about 3.6 exaflops of FP4 inference, roughly 3.3 times the ~1.1 exaflops of GB300 NVL
- – Rubin pairs HBM4 memory at about 13 TB/s bandwidth, roughly 1.6 times the HBM3E used in GB
- – Vera is an 88-core custom Arm design, about 22% more cores than the 72-core Grace it replaces, linked to Rubin GPUs through NVLink-C2C.
- Accelerator-based systems now make up close to 40% of recent TOP500 lists, while homogeneous CPU-only machines keep shrinking, the installed base NVIDIA is building on.
- Nemotron for Science runs on-premises, which removes the data-export approval step that cloud science agents require, cutting compliance from months to
For the industry, the short-term effect lands within three months: leading HPC centers and supercomputing labs will run small-scale evaluations, wiring Nemotron for Science into existing Slurm queues behind sandboxes, while bundled GPU-and-CPU purchasing gains leverage and pure accelerator orders lose pricing room. Over a one-year horizon, HPC tenders will likely list agent runtime and data-localization as separate line items, making the software stack a scoring dimension alongside peak throughput.
For ordinary users, there is little direct change in the first three months, since the shift hides behind cloud research APIs. The one-year view is more concrete: if drug screening, battery
Questions worth exploring:
- Verifiability design for science agents: making result checking a third-party auditable interface instead of model self-certification.
- The 88-core Arm path for Vera: how a homegrown CPU carves share out of an x86-dominated HPC installed base.
- On-premises agents and data sovereignty: how export controls and research data compliance reshape the product form of science AI.
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from NVIDIA Blog (published 2026-11-17 14:00:10). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.
