Nvidia is reviewing a downgrade of specifications for its next-generation AI chip, ‘Rubin Ultra,’ scheduled for release early next year, due to a High Bandwidth Memory (HBM) supply shortage. Market research firm TrendForce stated on Aug. 6 that Nvidia has not finalized the chip’s specifications and is discussing a reduction in HBM specifications. Initially, Nvidia planned to equip the chip with the 12-layer HBM4E, but is now reviewing downgraded alternatives such as the 8-layer HBM4E or the 12-layer HBM4. The review is driven by DRAM production capacity limitations expected to continue until next year, as wafer capacity for HBM is limited and uncertainties remain over the verification schedule and mass production yield of the 12-layer HBM4E. Major cloud service providers are also considering reducing HBM capacity for their in-house custom AI ASICs. The memory industry dismisses ‘peak out’ concerns; Samsung Electronics said the supply-demand gap next year will grow larger than this year, and SK Hynix projected memory demand will continue expanding due to solid big tech investments. The prolonged imbalance is maximizing negotiating power for suppliers like Samsung Electronics and SK Hynix, who are expected to exercise strong pricing power next year. TrendForce forecast additional HBM price increases, noting AI chip manufacturers will face the double burden of securing volume and rising procurement costs.