BlockBeats news, on August 4, Citrini analyst Jukan posted that NVIDIA is discussing adjustments to the HBM configuration for its next-generation Rubin Ultra GPU, potentially changing the main version from HBM4E 12Hi 384GB to HBM4 8Hi 192GB, with an HBM4E 8Hi version to follow.
Analysis suggests that if NVIDIA's CoWoS capacity allocation reaches 1.2 million units in 2027, its accelerator production is expected to be approximately 9.9 million units, including around 1.6 million Rubin Ultra units. If a lower-specification HBM solution is adopted, NVIDIA's HBM demand in 2027 is projected to decline from 24.1 billion Gb to 21.6 billion Gb, a reduction of about 10%, potentially lowering overall HBM demand by approximately 4%.
Jukan believes that this "downgrade" is not due to weakened demand, but rather NVIDIA's strategy to increase accelerator output amid limited HBM supply. Since DRAM manufacturers are unable to expand HBM production at a pace matching TSMC’s advances in process technology and packaging capacity, HBM supply has become a bottleneck for AI chip production.
In addition, DRAM manufacturers also find it difficult to significantly adjust capacity allocation. Currently, consumer electronics manufacturers of smartphones and PCs are already facing the impact of memory shortages, and further reducing traditional DRAM supply could intensify pressure in the end markets.
Analysis suggests that NVIDIA, by adjusting its HBM configuration, is expected to increase AI accelerator output while alleviating HBM supply constraints, even with limited memory resources.