Nvidia Cuts Vera Rubin HBM4 Bandwidth 13% [2026] - shattered.io

The cuts reach every configuration Nvidia lists, from the single GPU to the 72-GPU rack. The eight-GPU HGX Rubin NVL8 board takes the biggest hit at 26%.

Memory capacity stays at 288 GB per GPU and 20.7 TB per NVL72 rack, and the flagship Rubin GPU’s advertised 50 PFLOPS of sparse NVFP4 inference performance also remains unchanged.

The reductions mean the published figure now matches the lower of the two numbers Nvidia once floated, rather than the preliminary architectural targets it had previously advertised.

What changed on the Vera Rubin spec sheet

The single Rubin GPU lost about 13% of its HBM4 bandwidth, and the NVL72 rack fell from 1,580 TB/s to 1,400 TB/s.

The two-GPU Vera Rubin superchip saw its combined HBM4 bandwidth drop from 44 TB/s to 38.5 TB/s and its NVLink figure fall from 7.2 TB/s to 6 TB/s.

Superchip networking bandwidth rose from 0.8 TB/s to 0.9 TB/s, while scale-out networking at the rack level climbed from 28.8 TB/s to 32.4 TB/s.

The specifications published during Vera Rubin’s development were preliminary architectural targets. The revised figures reflect what the system sustains in qualified volume production.

Nvidia also argued that individual component specifications do not determine application performance, and it described Vera Rubin as built for agentic workloads that reason, call tools and retrieve information thousands of times per task.

The company then pointed to third-party testing, saying published benchmarks show 30x more performance per watt and 30x more tokens per dollar compared with the previous-generation GB300 NVL72.

Why the HBM4 target slipped

TechPowerUp reported in March 2026 that SK hynix and Samsung were struggling to reach Nvidia’s 22 TB/s HBM4 target and that early shipments looked closer to 20 TB/s.

The target also climbed fast before it fell, rising from 13 TB/s at GTC 2025 to 20.5 TB/s by September 2025 and then to 22 TB/s at CES 2026.

Raising a bandwidth target from 13 to 22 TB/s in about a year puts hard pressure on a memory supply chain, and Wccftech cited reporting that HBM4 suppliers struggled to meet the original targets.

Lower required pin speeds mean more stacks pass qualification, which is a number that decides shipment dates for a part that sells in the tens of thousands of units per quarter.

Supplier mix adds to the pressure, with Nvidia planning to source Vera Rubin HBM4 from SK hynix and Samsung without Micron, and with only two suppliers and a hard schedule, Nvidia had little room to hold out for faster memory.

Market impact and next steps

Samsung SDS plans a Vera Rubin GPU-as-a-service launch for March 2027, and hyperscalers have been pricing capacity around Nvidia’s headline figures for months.

The NVL8 board matters because it is the form factor most enterprises and smaller cloud providers actually buy, and a 26% cut on the product that ships to the broadest market is a bigger commercial issue than the 11% trim on the headline rack.

Reports trace the edits to product pages and to archived snapshots, with one account saying the revisions went live between September 1 and 11, 2026, and stayed up for more than a month before enthusiasts and outlets noticed.

Readers should treat the scale-out networking figure of 32.4 TB/s as reported rather than confirmed until Nvidia’s own documentation is archived.