Nvidia’s next flagship accelerator may ship with far less memory than originally planned, the clearest sign yet that the HBM shortage is rewriting chip designs before they reach production.
The Information reports that Nvidia has tested at least three lower-memory versions of Rubin Ultra, including configurations with 192GB or 256GB of memory and one that steps back to HBM4 rather than the HBM4E announced for the chip. TrendForce likewise says Nvidia has widened its evaluation beyond the baseline 12-Hi HBM4E design to include 8-Hi HBM4E, 12-Hi HBM4 and 8-Hi HBM4 options.
The original blueprint, shown at GTC, called for 1TB of HBM4E across 16 stacks per tray. The Vera Rubin GPU shipping today carries 288GB of HBM4, so some of the tested variants would come with less on-package memory than the generation Rubin Ultra is designed to succeed.
TrendForce attributes the rethink to a supply picture that stays constrained: DRAM output will remain limited through 2027, HBM4E validation timelines are still uncertain, and the projected 50 to 60 percent growth in HBM bit shipments next year will not keep up with demand, which leaves memory vendors holding pricing power.
Trimming stack height means less memory per GPU but more chips from the same wafer input, a tradeoff Nvidia appears willing to explore as it juggles HBM allocations. The company has not confirmed any specification change, saying only that it continually optimizes compute, networking and memory configurations.
