Nvidia is expected to see an increase in overall demand for high-bandwidth memory (HBM) next year, despite reducing the memory capacity of its next-generation AI chip, the Rubin Ultra. While the amount of HBM per chip may decrease, Nvidia is anticipated to supply more products, leading to an overall increase in HBM usage.
On August 10, Timothy Arcuri, a UBS analyst, stated in a report, "It appears that Nvidia is pursuing a 'de-spec' strategy for the Rubin Ultra specifications." He added, "While the HBM per chip may be reduced, the number of chips distributed across the supply chain could increase, suggesting that total HBM usage in 2027 may exceed previous forecasts."
UBS also raised its price outlook for HBM, noting that the supply of HBM4 and the next-generation HBM4E is tighter than expected. The firm explained that memory manufacturers are charging higher prices for HBM, leading to an increase in the average selling price forecast from 67% to 79% year-over-year.
Earlier, IT news outlet The Information reported that Nvidia is testing several prototypes of the Rubin Ultra with less HBM than originally planned. Some products are reportedly being considered with memory capacities reduced to 192GB or 256GB, or using HBM4 instead of HBM4E.
The consideration to reduce memory capacity is driven by concerns over HBM supply shortages. Nvidia is testing various specifications in anticipation of potentially not securing enough HBM4E in time for the Rubin Ultra's release. The final specifications have yet to be confirmed.
If UBS's predictions hold true, the reduction in HBM capacity for the Rubin Ultra may not significantly impact memory manufacturers such as SK Hynix, Samsung Electronics, and Micron. Although the HBM usage per product may decrease, the overall shipments of AI semiconductors and HBM prices could rise together.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.