NVIDIA explains Vera Rubin’s revised specs and performance claims

NVIDIA characterized the earlier Vera Rubin figures as preliminary architectural targets and the revised figures as specifications for systems that sustain qualified, volume production, according to a report published by Wccftech on October 29, 2026. The reported revisions span four configurations, and NVIDIA’s product page lists several of the updated GPU and NVL72 figures. NVIDIA’s Vera Rubin NVL72 specifications

Why NVIDIA says Vera Rubin’s figures changed

An NVIDIA representative told Wccftech that figures published during development were preliminary targets, while final specifications reflect what the system sustains in qualified, volume production. The representative also said individual component specifications do not determine application performance, which NVIDIA frames as a data-center-scale result involving compute, networking, memory, software and models.

NVIDIA announced on May 31, 2026, that Vera Rubin was ramping into full production. In the later statement reported by Wccftech, the representative said Vera Rubin NVL72 was in full production and shipping. Those are separate statements made at different points in the platform’s production timeline. NVIDIA’s May 31 production announcement

A Bug Labs video published October 14, 2026, shows architectural diagrams and discusses the 22 TB/s and 19.2 TB/s HBM4 bandwidth figures that appeared on NVIDIA pages before the later clarification.

The reported specification changes across four configurations

The earlier and revised values below are the configuration-level figures reported by Wccftech. NVIDIA’s product page lists the revised Rubin GPU, Vera Rubin Superchip and NVL72 specifications, including 19.2 TB/s of HBM4 bandwidth per Rubin GPU and 1,400 TB/s for the NVL72. The HGX Rubin NVL8 figures are also included in the reported comparison.

The figures do not all move in the same direction. Rubin GPU and NVL72 HBM4 capacity are unchanged in the reported comparison. Networking bandwidth rises for the Superchip, NVL72 and HGX Rubin NVL8, while several HBM4 and NVLink bandwidth figures and the two NVL8 training figures are lower. These are specification changes; application performance depends on more than a single component figure.

What NVIDIA’s performance claims compare

NVIDIA’s representative told Wccftech that third-party agentic benchmarks use production specifications and cited 30× more performance per watt and 30× more tokens per dollar versus GB300 NVL72. These are two different metrics, and both claims use GB300 NVL72 as the comparison system.

NVIDIA’s product page describes separate comparisons against GB200 NVL72: one-tenth the cost per million tokens and up to 10× more tokens per megawatt. The cost-per-million-tokens comparison specifies Kimi-K2-Thinking with 32K input and 8K output sequence lengths. The same page says Vera Rubin NVL72 paired with Groq 3 LPX can deliver up to 35× more throughput per megawatt for trillion-parameter models. NVIDIA’s Vera Rubin NVL72 product page

The 30× claims, the cost and tokens-per-megawatt comparisons, and the 35× throughput claim use different measures, reference systems or configurations. NVIDIA’s product page ties the 35× figure to the Groq 3 LPX pairing and trillion-parameter models; it is not the same comparison as the reported 30× figures versus GB300 NVL72.