# Vera Rubin Inference Results Leave Nvidia’s 75% Training Claim Untested By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/40939 Published: 2026-10-26T10:51:33.000Z Updated: 2026-10-26T10:51:33.000Z Section: Technology > At about 190 output tokens per second per user, Vera Rubin reached 5.7 times GB300’s total throughput per megawatt of GPU and memory power, using a four-GPU GB300 comparison. Oct. 15 AA-AgentPerf results put Vera Rubin NVL72 at 5.7 times GB300’s total throughput per megawatt at a specific speed, while leaving Nvidia’s separate training estimate untested. The finding measures inference — generating outputs from a trained model — rather than a completed training run. The benchmark exercised coding-agent tasks with DeepSeek V4 Pro. At the selected operating point, Vera Rubin produced about 190 output tokens per second for each user. GB300 rack configurations did not reach that rate, so the comparison used a four-GPU GB300 setup. The 5.7-times result reflects that configuration and speed, rather than a full-rack comparison. The power measure includes the GPUs and their memory but excludes CPUs, networking and cooling. The per-user speed leaves out time to first token, or the interval before a model begins responding. The total-throughput calculation also includes cached input tokens. When the systems were measured at their respective throughput peaks, Vera Rubin’s total throughput per megawatt was 1.14 times GB300’s. Nvidia’s Jan. 5 technical comparison addressed model training, a different stage of AI computing. It modeled training a 10-trillion-parameter mixture-of-experts model on 100 trillion tokens over one month. Nvidia estimated that workload would require 16,000 Vera Rubin GPUs, compared with 64,000 Grace Blackwell GPUs — a 75% lower GPU count for the specified scenario. The figure is a fixed-workload estimate, not a reported customer training run. A mixture-of-experts model routes work among specialized components, requiring GPUs to exchange information during training. Nvidia attributes part of its projected GPU reduction to denser computing and faster interconnections. The inference benchmark does not measure training speed or the resources required to complete the modeled job. These results describe serving performance at one common per-user speed and at each system’s own peak. They do not show that a customer trained the specified model with 75% fewer GPUs or establish total data-center energy use or cost for that run. CoreWeave opened production access to Vera Rubin NVL72 on Sept. 30 through its cloud service, in a [production rollout of the platform](). ## Links in this article - [production rollout of the platform](https://www.tokenpost.com/news/technology/25883)