Nvidia edited the Vera Rubin product pages sometime between September 1 and September 11, according to Wayback Machine comparisons first reported by IT Home. Per-GPU HBM4 bandwidth fell from 22 TB/s to 19.2 TB/s. NVLink 6 dropped from 3.6 TB/s to 3 TB/s.
Memory capacity held at 288 GB. The sparse inference headline figure of 50 PFLOPS stayed put. At rack scale the changes compound: aggregate memory bandwidth slid from 1,580 TB/s to 1,400 TB/s and the NVLink total from 260 TB/s to 216 TB/s.
The eight-GPU HGX Rubin NVL8 absorbed the heaviest cut. Its HBM4 bandwidth fell 26 percent, from 176 TB/s to 130 TB/s, and capacity from 2.3 TB to 2 TB. Dense training ratings slipped to 270 PFLOPS of NVFP4.
An Nvidia representative told Wccftech the earlier numbers were preliminary architectural targets and the revised ones describe systems sustaining qualified volume production. That is a reasonable reading, and it is also the quietest possible way to make it.
Timing carries the weight. The edit landed as fall production shipments began, weeks after Nvidia said racks were already running at CoreWeave, Google Cloud and Microsoft Azure. The revised figures describe hardware leaving the door, not a roadmap.
One inconsistency survives. The per-GPU table still pairs 288 GB with 22 TB/s, numbers that no longer reconcile with the system totals printed beside them.
