Nvidia Cuts Vera Rubin Memory Bandwidth to 19.2 TB/s per GPU

Nvidia’s own Vera Rubin NVL72 product page now lists 19.2 TB/s of HBM4 memory bandwidth per Rubin GPU, down from 22 TB/s on the same page as archived by the Internet Archive on 12 August 2026. Analyst Jukan flagged the change on X on 25 October 2026, writing that “NVIDIA has officially lowered the HBM4 pin speed.”

The rack-level figure fell from 1,580 TB/s to 1,400 TB/s. The compute figures still listed, including 50 PFLOPS of NVFP4 inference per GPU, and the 288 GB of HBM4 per GPU are the same as in August, though the INT8, FP32 SGEMM and FP64 DGEMM rows no longer appear and NVFP4 inference is now footnoted as a sparse specification. The August page described its values as “Preliminary information” and “subject to change,” and the current page gives no reason for the new figures.

What Changed on Nvidia’s Page

We compared the specification table on Nvidia’s Vera Rubin NVL72 page as archived on 12 August 2026 with the live page read on 26 October 2026. The memory and NVLink bandwidth figures are lower at every level of the system, from one GPU to the full 72-GPU rack.

Source: NVIDIA Vera Rubin NVL72 product page, Internet Archive snapshot of 12 August 2026 and live page read 26 October 2026. Percentages are our arithmetic.

Internet Archive snapshots of the page from 2 October 2026 onward already show the lower figures, so the change was made between 12 August and 2 October. The Korean-language version of the page, which Jukan posted, shows the same 19.2 TB/s figure.

What Was Posted on X

Jukan (@jukan05) posted on 25 October 2026 that Nvidia had lowered the HBM4 pin speed from 11.7 to 9.3 Gbps, with a screenshot of the Korean-language Vera Rubin spec table showing 19.2 TB/s per Rubin GPU. The pin-speed figures are Jukan’s; Nvidia’s page lists bandwidth, not pin speed, and this piece compares that page with its 12 August version.

The Pin-Speed Claim

Nvidia’s page lists bandwidth, not pin speed. The pin-speed figures come from Jukan’s post, which reads “11.7-> 9.3Gbps.” Pin speed is how fast data moves through each connection between a memory stack and the GPU, so a lower required speed is, in principle, easier for memory chip makers to meet, which is the point IT Home’s report makes below.

What It Means for Memory Makers

IT Home, as summarised by TechFlow on 26 October 2026, compared Nvidia’s October documentation with a June version saved by a third party. The report says that “lowering the HBM4 memory bandwidth threshold is expected to allow more memory products to meet the Rubin GPU specifications, thereby increasing final production volume.”

The same report lists other changes it found: HBM capacity in the HGX NVL8 system down from 2.3 TB to 2 TB, its HBM bandwidth down from 176 TB/s to 130 TB/s, and NVLink Switch bandwidth down from 28.8 TB/s to 24 TB/s. SK Hynix reports third-quarter results on 28 October, as our SK Hynix dividend piece reported.

Next to the 30x Claim

On 21 October, at GTC Berlin, Jensen Huang said Vera Rubin is another 30 times beyond Blackwell, and Nvidia’s release that day said Vera Rubin NVL72 systems “deliver 30x higher throughput per megawatt,” as our Huang 30x piece reported. By then, according to the archived snapshots, the product page had been listing 19.2 TB/s per GPU for at least 19 days.

The Structural Read

As we read it, the cut is consistent at every level Nvidia lists, from 22 to 19.2 TB/s per GPU, 44 to 38.5 TB/s per superchip and 1,580 to 1,400 TB/s per rack, so it reads as one specification change rather than a typo in one cell.

As we read it, the supplier side is the one IT Home spelled out: its report says a lower HBM4 bandwidth threshold is expected to let more memory products meet the Rubin specifications.

As we read it, the timing matters for how Nvidia’s claims are read: the archived page already showed 19.2 TB/s per GPU before the 21 October GTC Berlin keynote where Jensen Huang put Vera Rubin at 30 times Blackwell.

Three Implications

FOR RUBIN BUYERS Nvidia’s page now lists 19.2 TB/s of HBM4 bandwidth and 3 TB/s of NVLink bandwidth per Rubin GPU, against 22 TB/s and 3.6 TB/s on the 12 August page.

FOR MEMORY MAKERS IT Home reported that the lower threshold is expected to let more memory products meet the Rubin specifications; SK Hynix reports third-quarter results on 28 October.

FOR THE PIN-SPEED FIGURES The 11.7 and 9.3 Gbps figures are from Jukan’s post; Nvidia’s page lists bandwidth, not pin speed.

The Business Engineer Lens

This story maps onto the Business Engineer framework The AI Memory Chokepoint.

The analysis puts it this way: “Memory is becoming the new compute.”

As we read it, Nvidia’s page change is a small, concrete instance of that claim: the compute figures still listed held while the memory bandwidth figure moved. The analysis was published in December 2025 and does not discuss this change.

What Is Not Established

Nvidia’s page does not say why the bandwidth figures changed, whether the change affects any performance claim, or which memory suppliers it applies to. The 11.7 and 9.3 Gbps pin speeds are Jukan’s figures, not Nvidia’s. The HGX NVL8 and NVLink Switch figures come from IT Home via TechFlow, and we did not check them against Nvidia’s documents. We did not contact Nvidia.

The Bottom Line

Nvidia’s Vera Rubin NVL72 page now lists 19.2 TB/s of HBM4 bandwidth per GPU, down from 22 TB/s in August, and 1,400 TB/s per rack, down from 1,580 TB/s, while the compute figures it still lists are unchanged. IT Home reported that the lower threshold is expected to let more memory products qualify.