Bernstein Identifies AI Memory Bottleneck and Issues Sector Ratings

Analysts at Bernstein report that artificial intelligence is creating a significant memory bottleneck, with demand expanding beyond high-bandwidth memory (HBM) into conventional DRAM, NAND flash, and general storage. The research indicates that while AI training remains compute-intensive and reliant on HBM, the decode stage of inference is memory-bound, which may limit user capacity due to key-value cache requirements.

To mitigate these pressures, several new memory tiers are emerging, including CXL memory and CMX context storage. Nvidia Inc. has also launched the Storage Next initiative to address these evolving AI memory requirements. Furthermore, the rise of agentic AI and retrieval-augmented generation is expected to intensify demand for system DRAM, SSDs, and HDDs.

Based on these trends, Bernstein issued Outperform ratings for Samsung Electronics, SK hynix, Micron Technology, SanDisk, Seagate, and Western Digital. Conversely, the firm rated Kioxia as Underperform.