The numbers that ship are not the numbers that launched

Nvidia did not announce a Rubin downgrade. It edited one. The current Vera Rubin NVL72 page lists each Rubin GPU at 288 GB of HBM4 with 19.2 TB/s of bandwidth and 3 TB/s of NVLink bandwidth. The figures Nvidia promoted earlier in the year, in its January platform explainer and its July technical article on the Rubin GPU, were 22 TB/s and 3.6 TB/s. ServeTheHome's comparison of the preliminary and shipping tables puts the two sets side by side: memory bandwidth down 12.7%, NVLink down 16.7%, and one figure moving the other way, scale-out networking up from 0.4 to 0.45 TB/s per GPU.

Archived versions of Nvidia's product pages place the specification changes between September 1 and September 11, 2026, according to Wccftech's comparison of Wayback Machine snapshots. Nvidia did not publish a separate announcement explaining the revisions. The revised numbers then sat on the site for weeks while launch coverage, including plenty of it written this month, kept quoting 22 TB/s. That gap between the datasheet and the coverage is the practical problem. Infrastructure teams that used the preliminary figures in capacity or performance planning should revisit those models against the shipping specifications. The difference is particularly relevant when estimating memory-bound workloads and inter-GPU communication.

What did not change matters too

Nvidia's published shipping specifications retain the same 288 GB memory capacity and theoretical compute figures per GPU. The reductions are concentrated in HBM4 bandwidth and NVLink bandwidth, while scale-out networking bandwidth has increased. The distinction matters most for workloads that are constrained by memory bandwidth or GPU-to-GPU communication. Large-model inference and distributed training can be sensitive to both, although the real-world effect depends on the model, workload and software. The published specification changes establish lower theoretical bandwidth ceilings; they do not quantify the performance loss for a particular customer workload.

The early figures were explicitly preliminary. Nvidia's original materials stated that all values were "up to and subject to change." These are therefore revisions to pre-launch specifications, not proof that Nvidia failed to deliver a contractual shipping commitment. The important distinction for buyers is that the current shipping specifications are lower than the figures used in earlier planning and coverage. And the revised pin rate, 9.375 Gbps, still runs above JEDEC's 8 Gbps HBM4 target and above the 7.6 Gbps ServeTheHome reports for AMD's MI455X stacks. Rubin after the cut is not slow memory. It is slower than Rubin's own launch sheet said, which is a different and narrower claim.

The eight-GPU board takes the real hit

The rack follows the per-GPU figures down in proportion: 1,400 TB/s of memory bandwidth instead of 1,580, and 216 TB/s of NVLink instead of 260, partly offset by scale-out networking rising to 32.4 TB/s. The HGX Rubin NVL8, the eight-GPU board sold with either an x86 CPU or Nvidia's Vera CPU, is cut harder than the GPU itself. Its memory capacity falls from 2.3 TB to 2.0 TB, about 250 GB per GPU instead of a full 288, which is consistent with approximately seven 36 GB HBM4 stacks per GPU rather than eight, although Nvidia has not explicitly confirmed the physical configuration. Its aggregate memory bandwidth falls 26%, from 176 TB/s to 130 TB/s, and its training figures are trimmed as well, NVFP4 from 280 to 270 PFLOPS and FP8 from 140 to 130. A buyer comparing an NVL8 against the rack on launch numbers would get the proportions wrong in both directions.

Nvidia has not said why the figures moved. ServeTheHome suggests the original memory targets may have proved difficult to deliver at production scale, potentially involving memory-controller speeds, stack performance or yields. That remains the site's assumption; Nvidia and its memory suppliers have not confirmed a cause. The documentation is still inconsistent, too. Nvidia's current HGX page lists the revised NVL8 system figures of 2 TB and 130 TB/s, but its individual Rubin GPU table still shows 288 GB, 22 TB/s and 3.6 TB/s of NVLink bandwidth. Buyers have to distinguish the revised system specifications from the stale individual-GPU entries.

Sources

  1. [1] ServeTheHome, November 2026 — "NVIDIA Updates Rubin GPU Specs as First Production Hardware Arrives" (preliminary-vs-shipping tables for the Rubin GPU, Vera Rubin NVL72 and HGX Rubin NVL8; preliminary-figures disclaimer; JEDEC and MI455X context; NVL72 general-availability note; page-displayed date October 31, 2026, UTC timestamp November 1, 03:00; read first-hand)Read source
  2. [2] NVIDIA — "NVIDIA Vera Rubin NVL72" product page (current specifications verified first-hand: per-GPU 288 GB | 19.2 TB/s | 3 TB/s NVLink; Superchip 576 GB | 38.5 TB/s; rack 20.7 TB | 1,400 TB/s | 216 TB/s)Read source
  3. [3] NVIDIA — "NVIDIA HGX Platform" page (current specifications verified first-hand: NVL8 system table lists the revised 2 TB | 130 TB/s | 24 TB/s figures, while the individual Rubin GPU table still shows 288 GB | 22 TB/s | 3.6 TB/s)Read source
  4. [4] Wccftech — "NVIDIA Rewrote Vera Rubin's Spec Sheet Between September 1 And 11 Without A Word" (Wayback Machine dating of the edit; HGX per-GPU table inconsistency)Read source
  5. [5] VideoCardz — "NVIDIA quietly cuts Vera Rubin specs, HBM4 bandwidth and NVLink 6 speeds now lower" (percentage calculations; HGX page still listing original per-GPU figures)Read source