Bernstein's report highlights a shift in AI hardware demand from high-bandwidth memory (HBM) towards conventional dynamic random-access memory (DRAM) and storage solutions. The analysis indicates that while HBM remains essential for the training phase, the inference phase - specifically the memory-bound decode stage - is driving a need for increased DRAM capacity. This shift is underpinned by the growing requirements of key-value (KV) caches, which are expected to outgrow model weights in large deployments. Consequently, companies like Samsung, SK hynix, and Micron are being recommended for their exposure to these emerging memory trends.
Bernstein Shifts AI Hardware Bets from HBM to DRAM and Storage
