Nvidia has revised the technical specifications of its flagship Vera Rubin AI platform, lowering the stated bandwidth figures for HBM4 memory and the sixth-generation NVLink interface. The adjustments affected individual Rubin GPUs, the Vera Rubin NVL72 rack system, and the HGX Rubin NVL8 platform. The most significant reduction affected the total HBM4 memory bandwidth in the eight-GPU configuration. The changes were introduced on the official product pages in September 2026, but Nvidia did not publish a separate statement explaining the reasons for the revision.

What Changed at the Individual GPU Level

At the time of the platform's announcement in January 2026, the Vera Rubin NVL72 system specifications stated that each Rubin GPU featured 288 GB of HBM4 memory with a bandwidth of 22 TB/s. After the revision, the memory capacity remained unchanged at 288 GB, but the bandwidth was reduced to 19.2 TB/s, representing a 12.7% drop. At the same time, the NVLink 6 interface bandwidth also decreased: from 3.6 to 3 TB/s per GPU, down 16.7%. These figures determine the baseline performance of each accelerator in conjunction with memory and during data exchange between GPUs within a node.

System Losses in NVL72 and HGX Rubin NVL8

The Vera Rubin NVL72 rack system, combining 72 Rubin GPUs and 36 Vera CPUs, also lost aggregate performance metrics. While Nvidia previously cited a total HBM4 bandwidth of 1,580 TB/s and an NVLink bandwidth of 260 TB/s for it, these figures are now 1,400 and 216 TB/s, respectively. For the HGX Rubin NVL8 platform, which includes eight Rubin GPUs, the adjustments were even more noticeable: the stated HBM4 memory capacity decreased from 2.3 to 2 TB, and memory bandwidth dropped from 176 to 130 TB/s. NVLink Switch bandwidth also declined from 28.8 to 24 TB/s. Notably, the scale-out networking bandwidth for connecting multiple systems into clusters increased instead.

Model Training Performance Adjustments

In addition to raw memory and interface characteristics, Nvidia also adjusted neural network model training performance metrics. For the NVFP4 format, the value decreased from 280 to 270 petaflops, and for FP8/FP6 formats, from 140 to 130 petaflops. Although absolute numbers still remain at a level unavailable to competing solutions, a 3–7% decrease in the context of the company's multi-year promises may be perceived by analysts as a signal that targets set during the design phase were not fully achieved in mass production.

Documentation Discrepancies and Change Chronology

The current HGX Rubin NVL8 product page on Nvidia's website already displays the updated, lowered specifications, but the table for the individual Rubin GPU still shows previous figures: 22 TB/s for HBM4 memory and 3.6 TB/s for the NVLink interface. A separate page dedicated to DGX Rubin NVL8 also continues to claim 2.3 TB of memory and 176 TB/s bandwidth. Nvidia has not clarified whether these discrepancies are due to different hardware configurations or outdated documentation. An analysis of archived specification versions conducted by Wccftech showed that changes for the NVL72 system were made between September 1 and September 11, 2026.

Market Impact and Nvidia's Position

From a commercial standpoint, such adjustments are unlikely to significantly impact Rubin sales. According to reports, Nvidia has a long queue of enterprise customers awaiting the product release, and even the slightly reduced platform specifications remain well above competitors' offerings. Nevertheless, for the industry as a whole, revising the specifications of a flagship AI accelerator a few months before mass shipments creates a precedent that could affect partner trust and expectations surrounding subsequent architecture generations. For now, Nvidia remains silent regarding the reasons for the revision—whether due to HBM4 memory manufacturing constraints, thermal or electrical characteristics, or a reassessment of targets during final validation.