Citi sees memory chip shortfall widening through 2031

Citi's research note, published Monday, frames continual learning as the driver behind a multi-year memory squeeze. That approach lets AI models take on new tasks and information while keeping access to what they already learned, which raises requirements for both memory and storage. Citi expects high-bandwidth memory bit demand to climb 62% to 75.2 billion gigabits in 2027, then another 69% to 127.0 billion gigabits in 2028, and it sees enterprise solid-state drives drawing extra demand as AI systems need more room to retain prior knowledge.

The supply side is where the imbalance shows. Citi forecasts DRAM supply growth of 19% in 2027 and 22% in 2028, below projected demand growth of 30% and 35%, yielding supply-demand ratios of -8.7% and -9.7%. NAND shows a similar pattern: demand growth of 29% and 33% against supply increases of 21% and 25%, for ratios of -6.1% in 2027 and -5.5% in 2028, compared with an estimated -0.8% in 2026.

Citi's equity research leans into that view, naming Samsung Electronics, SK Hynix, Micron, Sandisk and Kioxia as preferred memory semiconductor stocks, and Montage, Applied Materials, Lam Research, TES, Eugene Tech and TechWing in equipment and materials. The projections reflect Citi's expectation that AI-related memory consumption will grow faster than available supply, with the imbalance potentially extending through 2031. Whether that plays out appears to hinge on how quickly continual learning becomes a standard part of AI development, since Citi's demand figures rest on that shift rather than on today's workloads.