Citi bets on AI-driven memory upcycle, calls pullback a buying opportunity

Citi argues the AI-driven memory upcycle is still in its early stages and could outperform the 2001–07 NAND upcycle, backed by three- to five-year LTAs and a persistent shortage of high-bandwidth memory (HBM). After a pullback in memory names, Citi expects AI-driven demand to stay robust as GPU deployments grow, with HBM capacity per AI system projected to rise about 434% (from 20.7 TB to 110.6 TB) and GPU counts per system from 72 to 576. Citi also notes SK Hynix could announce shareholder returns ahead of its Q3 and keeps a Buy rating on SK Hynix with higher earnings estimates, framing memory pullbacks as buying opportunities for names like MU and SK Hynix.