📌 The HBM Bottleneck Is Changing AI Chip Specifications and Allocation
Nvidia is evaluating lower HBM configurations for Rubin Ultra. What is worth watching is not whether AI memory demand has peaked, but rather that memory supply and certification progress may already be influencing the product specifications of next-generation chips in reverse. BofA’s view is that configurations as low as 192GB are mainly due to tight HBM supply and HBM4e certification constraints, rather than a permanent reduction in the eventual memory requirement of approximately 1TB.
This distinction directly affects the pace at which revenue is realized across the industry chain. If the constraints stem from supply and certification, whether GPU makers can ship at higher specifications depends on the amount of qualified HBM they can obtain; memory makers, meanwhile, face tighter allocation of high-end products, but order conversion will still be constrained by validation progress. The short-term reduction in configurations may therefore slow the amount of memory installed per chip, but cannot by itself prove that long-term demand is weakening.
The demand side is also becoming more diversified. Meritz Securities estimates that Broadcom’s HBM demand may reach 35 billion to 40 billion Gb in 2028, higher than Nvidia’s 30 billion Gb; Google would account for 20 billion Gb, and AMD for 10 billion Gb. If this forecast materializes, HBM will no longer be determined primarily by a single GPU customer, as custom-chip makers and cloud providers jointly compete for limited supply.
The key variables remain HBM capacity expansion, HBM4e certification, and the actual production ramp of these new customers’ products. They will determine whether the bottleneck temporarily lowers configurations or continues to alter the ability of different chip makers to obtain high-end memory.
▌ Sources