# Nvidia's Vera Rubin NVL72 Posts First MLPerf Inference Numbers, Claiming 3.7x Over Current Systems

The results arrive in MLPerf's preview category, meaning the hardware is not yet generally available and the comparison baseline is Nvidia's own previous generation.

- Published: 2026-09-17T06:17:58.659Z
- Canonical: https://polylog.news/ai/2026-09-17/nvidia-s-vera-rubin-nvl72-posts-first-mlperf-inference-numbe
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [NVIDIA Blog](https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference/), [MLCommons](https://mlcommons.org/2026/09/mlperf-inference-v6-1-results/), [Nebius](https://nebius.com/blog/posts/mlperf-inference-v6-1-results)

Nvidia submitted its Vera Rubin NVL72 rack-scale system to [MLPerf Inference v6.1](https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference/), the industry benchmark run by MLCommons, and reported up to 3.7 times the throughput of the GB300 NVL72 on Qwen3-VL and up to 2.5 times on DeepSeek-R1, across the offline, server and interactive scenarios. The submissions ran vLLM together with Nvidia's Dynamo inference framework.

Two details matter more than the headline multiple. First, Vera Rubin entered the preview category, which MLPerf reserves for hardware that is not yet commercially available, so no independent buyer has reproduced the numbers on a purchased system. Second, the baseline is Nvidia's own prior generation rather than a competing accelerator, which measures generational progress and not competitive position.

The rest of the round is arguably the more durable signal. Nvidia says software optimization alone delivered up to 1.6 times higher performance over v6.0 on existing hardware, and a 288-graphics-processing-unit submission spanning four GB300 racks held 99 percent scaling efficiency against a single-rack baseline. [MLCommons reported record participation](https://mlcommons.org/2026/09/mlperf-inference-v6-1-results/) for this round, and cloud providers submitted their own results, with [Nebius claiming five first-place finishes](https://nebius.com/blog/posts/mlperf-inference-v6-1-results) across 20 category entries plus a Vera Rubin preview entry of its own.

For anyone modeling serving costs, the software gain is the figure to rely on today. It applies to installed fleets now, while the Vera Rubin figures describe a platform buyers cannot yet deploy.

## What this means

Throughput per rack sets the denominator in the cost per million tokens that every inference provider quotes. If the 1.6 times software gain on shipped GB300 hardware holds in production, serving margins improve for existing fleets without new capital spending, which favors operators with large installed bases and weakens the case for rushing to the next platform. Neoclouds that priced multi-year contracts against fixed per-token assumptions carry the exposure, because their customers will expect the efficiency gains to be passed through. The open question is whether Vera Rubin's advantage holds up on real workloads. If generally available systems reproduce something near 3.7 times, Nvidia's pricing power extends another cycle. If the gain compresses to the low double digits in percentage terms on mixed traffic, buyers have a reason to keep depreciating current racks longer.

## What to watch

- Whether any customer outside Nvidia publishes Vera Rubin throughput on generally available hardware, which would convert a vendor preview into a verified capability.
- Whether AMD, Google and Amazon submit comparable large-model results in the same round, since their absence makes the leaderboard a measure of Nvidia against itself.
- Whether inference providers cut published per-token prices in the weeks after this round, which is the clearest evidence that benchmark gains reach customers.