Training DeepSeek-V3 drops from 16 minutes to about 4 when the GPUs go from 256 to 1024. According to NVIDIA, the scaling holds at about 97% efficiency.
Fact
The facts: on 16 November 2026 NVIDIA announced its results from MLPerf Training v6.1 and from its first MLPerf Endpoints submission. According to NVIDIA, the company swept every benchmark in Training v6.1 and was the only platform with results on every one of them, and in Endpoints it delivers more than 10x the throughput per watt of the previous generation on agentic workloads. In its debut, Vera Rubin NVL72 delivers up to 2x more performance per GPU than GB300 NVL72. NVIDIA also ran the benchmark on 10240 GPUs with GB300 systems, the largest scale in this round, and Microsoft Azure matched that scale with 10240 GPUs with GB200 systems. 20 partner organizations took part in Training. The Vera Rubin results in Endpoints are provisional: they are pending MLCommons Association verification and are subject to change pending peer review.
In Endpoints the numbers are more specific. On the Kimi K3 benchmark, a large open model built for agentic workflows, NVIDIA delivers more than 10x the throughput per watt of GB300 NVL72 at the same response speed. One system has 32 accelerators, the other 48. The smaller system gives more per watt, according to NVIDIA itself.
Tokens per watt decide how many agents one data center can serve.
Scale is the second half. NVIDIA ran the benchmark on 10240 GPUs with GB300 systems, the largest in this round, and Microsoft Azure matched that number with GB200. The record is not the point: adding GPUs shortens training only if they work together. That is why NVIDIA leans on the links between them, not just on the chips themselves.
The debut is a fact. MLCommons' signature under the Endpoints results is not there yet.
Official primary sources
NVIDIA - NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance, 16.11.2026The visual is generated code-art, no third-party images.