NVIDIA Says Vera Rubin Debuts at the Top of MLPerf Training v6.1

NVIDIA says its Vera Rubin platform swept MLPerf Training v6.1 and cut energy per agentic task more than tenfold. The claims are NVIDIA's own account, and some inference results await MLCommons verification.

By Utopia Data & AI Team6 min read

The curved glass NVIDIA headquarters building in Santa Clara behind a grass slope, a flight of steps and a sign showing the NVIDIA logo and the number 2788
Licensed photo. NVIDIA's headquarters in Santa Clara, California, photographed on 4 August 2018. NVIDIA published the benchmark results on 16 November 2026. The image is contextual and shows no Vera Rubin hardware. Photo: Coolcaesar/Wikimedia Commons, CC BY-SA 4.0.[2][1]

In brief

  • NVIDIA said on 16 November 2026 that it swept every MLPerf Training v6.1 benchmark, and that its Vera Rubin NVL72 system delivered up to 2x more performance per GPU than GB300 NVL72 in its debut.[1]
  • In its first MLPerf Endpoints submission, NVIDIA reported more than 10x generational gains in throughput per watt for agentic AI, in a provisional Kimi K3 result pending MLCommons verification.[1]
  • The company said 20 organisations submitted MLPerf Training results on its platform, and that it was the only platform with results on every benchmark this round.[1]

What NVIDIA reported

NVIDIA has put its next data centre platform on the public scoreboard. On 16 November 2026 the company published a blog post saying Vera Rubin made its MLPerf debut with leading performance across training, post-training and agentic inference.[1]

The headline claims are a sweep of every MLPerf Training v6.1 benchmark and, in its first MLPerf Endpoints submission, more than 10x generational gains in throughput per watt for agentic AI. NVIDIA says it was the only platform with results on every benchmark this round.[1]

The post is NVIDIA's own account. It points readers to the MLCommons results pages and to two MLPerf technical blogs, which we did not read for this story. NVIDIA also says some of the inference results are provisional.[1]

Two cabinets of the JUPITER supercomputer in a white data hall, one printed with the JUPITER name and the Jülich logo, beside a stand listing partner logos
Licensed photo. Cabinets of the JUPITER supercomputer at Forschungszentrum Jülich, photographed on 30 June 2025. The MLPerf results concern other systems. The image shows what a large training machine looks like and is contextual. Photo: Forschungszentrum Jülich/Sascha Kreklau/Wikimedia Commons, CC BY 4.0.[3][1]

What the training results show

MLPerf Training measures how fast a system can train a range of AI models to a target level of accuracy. NVIDIA says its platform was the only one with results on every benchmark this round, which spanned large language models, image generation, recommender systems and reinforcement learning post-training.[1]

In its debut, Vera Rubin NVL72 delivered up to 2x more performance per GPU than the NVIDIA GB300 NVL72 system, which NVIDIA calls highly optimised. The company describes these as early results and says performance will keep climbing through software optimisation.[1]

How the scaling numbers read

On the DeepSeek-V3 benchmark, NVIDIA says Vera Rubin NVL72 scaled from 256 to 1,024 GPUs with about 97 percent efficiency. Quadrupling the GPUs cut the time to train from 16 minutes to about four. That is the pattern buyers look for, because added GPUs only help when they work together efficiently.[1]

NVIDIA also ran the benchmark on 10,240 GPUs with GB300 systems, which it describes as the largest scale in the round. It says Microsoft Azure matched that scale with 10,240 NVIDIA GPUs using GB200 systems. Submissions came from 20 organisations, including Azure, Google Cloud, Oracle, CoreWeave, Dell Technologies and Supermicro.[1]

A white NVIDIA sign with the green logo and the address 2800 and 2806 Scott Boulevard, standing among trees and hedges on a corporate campus
Licensed photo. An NVIDIA sign on its Santa Clara campus, photographed on 31 October 2020. The company's benchmark claims in this story are its own and the image is contextual. Photo: Will Buckner/Wikimedia Commons, CC BY 2.0.[4][1]

The agentic inference result and its caveat

MLPerf Endpoints tests systems on agentic coding and enterprise workflows and checks that the agents still finish their tasks correctly. In a provisional result for Kimi K3, NVIDIA says Vera Rubin NVL72 delivered more than 10x the throughput per watt of GB300 NVL72 at the same response speed.[1]

NVIDIA's footnote gives the inputs: 41,414.2 tokens a second at 100.89 kilowatts for Vera Rubin, against 2,210.4 tokens a second at 60 kilowatts for GB300. Our arithmetic puts those at about 410 and 37 tokens a second per kilowatt, a ratio near 11. The footnote says the Vera Rubin result awaits MLCommons verification and may change.[1]

Why Malaysia should care

Our analysis: Malaysia is adding data centre capacity where power is the limit, so throughput per watt matters more here than peak speed. NVIDIA's figures say nothing about Malaysian availability, pricing or hosting, and a local buyer would reach Vera Rubin through a cloud or data centre operator.

Malaysian data centre operators

Power limits capacity, so tokens per watt is a useful yardstick. NVIDIA's figure is provisional and covers one model.[1]

Practical move: Ask vendors for verified MLCommons results on the workloads you plan to host.

Malaysian companies buying AI cloud capacity

You will most likely reach Vera Rubin through a cloud provider. The post names no Malaysian pricing or region.[1]

Practical move: Ask your provider when Vera Rubin capacity reaches a nearby region and at what ringgit price.

AI engineering teams

Benchmarks test set workloads. Your model and data may behave differently.[1]

Practical move: Run a pilot on your own workload before you plan a move to new hardware.

What Malaysians can do now

  1. Read the MLCommons results pages before quoting any figure.
  2. Ask providers for verified results on your own model type.
  3. Compare tokens per watt only at the same response speed.

What we still do not know

What we still do not know

  • Whether MLCommons will verify the provisional Vera Rubin Endpoints result unchanged.
  • What Vera Rubin systems cost and when they ship in volume.
  • How rival platforms performed on the benchmarks they entered.

Sources

  1. NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading PerformanceNVIDIA Blog, 16 November 2026
  2. File:NVIDIA Headquarters.jpgWikimedia Commons, 4 August 2018
  3. File:JUPITER racks with logos of supporters and partners.jpgWikimedia Commons, 30 June 2025
  4. File:Nvidia sign.jpgWikimedia Commons, 31 October 2020

Facts come from NVIDIA's blog post of 16 November 2026, read directly on 17 November. All performance claims are NVIDIA's own. Calculations marked as ours use the numbers in NVIDIA's footnote. We did not read the MLCommons results pages, the technical blogs or rival submissions, and we did not run any benchmark. The three photographs are licensed contextual images and none shows Vera Rubin hardware. No generated image is used.

Last fact check: 17 November 2026. Factual correction? devteam.utopia@gmail.com.

More AI News

Back to AI NewsAll 201 stories