NVIDIA says its Vera Rubin platform led every benchmark in the latest MLPerf Training round, spanning language models, image generation, recommendation systems and reinforcement-learning post-training. The results, released on November 16, 2026, mark Vera Rubin’s debut in the tests. NVIDIA says the system delivered up to twice the performance per GPU of its GB300 NVL72 system.
In one training test, Vera Rubin scaled from 256 to 1,024 GPUs with about 97% efficiency, according to NVIDIA. Quadrupling the GPUs cut the reported training time from 16 minutes to about four. NVIDIA also submitted results across every training benchmark in this round; the company says it was the only platform to do so.
What the agentic-AI result measures
NVIDIA also made its first submission to MLPerf Endpoints, which tests AI systems on coding and business workflows that involve multiple steps, such as using tools and checking results. In a Kimi K3 test, NVIDIA reports more than 10 times the throughput per watt for Vera Rubin NVL72 compared with GB300 NVL72, at a shared minimum of 50 tokens per second per user. Throughput per watt measures how much work a system produces for its power use.
That comparison is not a simple like-for-like hardware count: the Vera Rubin result used 32 accelerators, while the GB300 result used 48. The Vera Rubin result is also provisional and still awaiting verification and peer review by MLCommons. The reported figures list average user interactivity of 50.4 tokens per second for Vera Rubin and 56.2 for GB300.
A benchmark result, not a product launch
For large training runs, adding GPUs helps only if they can work together efficiently. NVIDIA points to its NVLink connections inside systems and its InfiniBand and Ethernet networking between systems as part of how it scales workloads. The round also included a 10,240-GPU DeepSeek-V3 run on NVIDIA GB300 systems; Microsoft Azure submitted a run at the same scale using GB200 systems.
Twenty organizations submitted MLPerf Training results using NVIDIA technology. The benchmarks offer a way to compare performance on defined tasks, but the announcement does not say when Vera Rubin systems will be available to customers or what they will cost.
Why it matters
For researchers and companies training AI models, the reported scaling result suggests that a large Vera Rubin system could shorten a defined training run when its GPUs work efficiently together. But the agentic-AI comparison remains provisional, and NVIDIA has not given customer availability or pricing details, so the results do not establish when or at what cost people can use the hardware.