There is an outlook that even if Nvidia reduces the amount of high bandwidth memory (HBM) in its next-generation artificial intelligence (AI) accelerator Rubin Ultra, total HBM demand next year could actually exceed earlier expectations. The analysis is that lowering the HBM capacity per chip would ease supply constraints, allowing Nvidia to ship more AI accelerators.
On the 10th (local time), Timothy Arcuri, an analyst at global investment bank UBS, said in a report that Nvidia appears to be considering a "de-spec" that would lower some of Rubin Ultra's specifications, and offered this outlook. Arcuri said, "A de-spec could reduce HBM usage per chip, but it could increase the number of chips circulating across the entire supply chain." As a result, total HBM consumption in 2027 could exceed previous projections.
If Nvidia lowers the HBM amount, the number of AI accelerators that can be produced with a limited HBM supply increases. UBS analyzed that if the increase in accelerator shipments outweighs the decrease in memory usage per chip, total HBM consumption could instead rise.
The HBM price outlook also rose. UBS said supply for sixth-generation HBM (HBM4) and seventh-generation HBM (HBM4E) is tightening more than initially expected. As memory makers again widen the pricing premium attached to HBM, the average selling price (ASP) of HBM is forecast to rise about 79% from a year earlier. That is 12 percentage points higher than the previous forecast of a 67% increase.
Earlier, U.S. information technology (IT) outlet The Information reported that Nvidia is testing several prototypes of Rubin Ultra with less HBM capacity than originally planned. According to the report, some test plans reportedly include lowering memory capacity to 192 gigabytes (GB) or 256 GB, or using HBM4 instead of HBM4E. It is known to be a contingency in case sufficient volumes of HBM4E cannot be secured in time for the Rubin Ultra launch.
This is the backdrop for analysis that even if a reduction in Rubin Ultra's HBM amount is reflected in the final product, the impact on HBM suppliers such as Samsung Electronics, SK hynix, and Micron could be limited. If HBM capacity per product decreases but AI accelerator shipments increase and HBM prices also rise, the overall market size could expand.