Nvidia’s Rubin Ultra HBM cuts: reported 192GB–256GB designs and what they mean

Nvidia is reportedly exploring lower-memory versions of its Rubin Ultra AI accelerator as it faces constraints in securing high-bandwidth memory (HBM). Reports describe prototypes with 192GB or 256GB per GPU, but the company has not announced a final specification. If those configurations proceed, they would fall below standard Rubin’s reported 288GB and well short of Rubin Ultra’s original roughly 1TB ambition. 145

What Rubin Ultra configurations are being tested?

TrendForce says Nvidia has expanded its evaluation beyond a baseline of 12-high HBM4E stacks. The alternatives include 8-high HBM4E, 12-high HBM4 and 8-high HBM4; the final design remains undecided. 2

Separately, reports describe Rubin Ultra prototypes with 192GB or 256GB of HBM per GPU, with some using HBM4 rather than the newer HBM4E. Those capacity reports and TrendForce’s list of stack options describe designs under evaluation—not a settled product configuration. 15

How do the reported capacities compare?

Configuration	Reported HBM capacity per GPU	What the reports say
Standard Rubin	288GB	The comparison point cited in reporting on Rubin Ultra. 1
Rubin Ultra prototypes	192GB–256GB	Capacities reported for test versions; not confirmed final specifications. 5
Original Rubin Ultra ambition	Roughly 1TB	The earlier target described in reports about the planned design. 4

At 192GB, a reported Rubin Ultra configuration would have one-third less memory than standard Rubin. A 256GB version would also be below Rubin’s reported 288GB. Both figures are far below the roughly 1TB ambition, but they should not be treated as a confirmed before-and-after comparison for a shipping product: the earlier target and newer prototype reports may reflect different design assumptions. 145

Why could a shortage change the design?

TrendForce attributes the evaluation of alternatives to supply-side constraints, including an expected DRAM shortage in 2027 and uncertainty around qualifying 12-high HBM4E. Considering other stack heights and HBM generations could give Nvidia more options if the newest memory is difficult to secure in sufficient quantities. 2

Using less HBM could lower the memory component requirement for an individual accelerator, but the available reporting does not establish a final chip price, a customer price cut or a guaranteed increase in the number of GPUs Nvidia can ship. The reported design changes point to supply flexibility; their effect on total system and cloud costs remains uncertain.

What lower memory could mean for AI workloads

Memory capacity affects how much model data and working state can be held on a GPU. With less HBM per device, some memory-intensive jobs could need to be spread across more accelerators or use a different deployment approach. That may offset some of the benefit of fitting more lower-memory GPUs into a constrained supply pool. 56

This is an implication of the reported capacities, not a measured performance result for a final Rubin Ultra product. The sources do not establish how a shipping configuration would perform across specific models or workloads.

Is a consumer GPU delay connected?

A separate report says Rubin-based RTX 60-series gaming cards might arrive in 2028 rather than the previously expected late-2027 window. That report cites a hardware leaker, and Nvidia said it does not comment on rumors or unannounced products. It does not show that the Rubin Ultra HBM changes caused a consumer-GPU delay. 18