The core judgment of Citi’s (C) latest report is that AI is moving from one-time training to continuous learning, and the logic of memory demand is undergoing structural changes. HBM, server DRAM, and eSSD are expected to expand simultaneously, while supply expansion is constrained by capacity allocation and construction cycles, and the supply-demand gap may continue to widen until 2031. This round of memory prosperity is no longer just a traditional inventory restocking cycle, but increasingly shows structural characteristics driven by changes in AI computing architecture, and memory chip manufacturers, semiconductor equipment and materials suppliers may all benefit from it.
Citi expects that starting in 2027, demand for memory from AI training and inference is expected to accelerate simultaneously. Among them, HBM benefits from continuous training and model updates, server DDR5 and SoCAMM2 benefit from AI inference and increased CPU workloads, and eSSD undertakes the long-term storage needs for large amounts of historical data and data such as KV Cache.
Among all memory products, HBM remains one of the most direct beneficiary areas of AI training and continuous learning. Citi expects that HBM bit demand will increase by 62% year over year to 75.2 billion Gb in 2027, and further increase by 69% year over year to 127 billion Gb in 2028. Even if memory manufacturers actively expand production, HBM supply may still be unable to catch up with demand, with the supply gap at about 21% in 2027 and possibly widening to about 36% in 2028.
Server DRAM is expected to become an important beneficiary of the expansion of AI inference and continuous learning. Citi expects that server DRAM demand will grow by more than 50% in 2027, from 226.3 billion 1Gb equivalent units in 2026 to 341.7 billion units in 2027, and servers will contribute about 67% of global DRAM demand. At the same time, eSSD demand may also grow by about 50%. Memory customers are extending the terms of some long-term supply agreements from three years to five years, reflecting that market visibility into memory demand for many years to come is improving.
The NAND market may also become an important beneficiary of the continuous learning era. Citi expects that global NAND demand will increase by 29.1% year over year in 2027, higher than supply growth of 21.2%; among this, enterprise SSD demand is expected to increase by 52.9% year over year. One of the important factors driving this growth is KV Cache offloading. As some AI accelerators reduce HBM configurations, more and more KV Cache and other memory workloads may be transferred to external storage, and high-capacity QLC enterprise SSDs are therefore expected to gain more demand.
Tight supply and demand also mean that memory prices may remain strong. Citi expects that global DRAM bit demand will increase by about 30.2% year over year in 2027, significantly higher than supply growth of 18.8%, with a market gap of about 8.7%; by 2028, the supply gap is expected to further widen to about 9.7%. After DRAM blended average selling prices surged by 242.4% in 2026, DRAM blended ASP may still rise by 23.1% year over year in 2027. The NAND market supply gap is expected to reach about 6.1% in 2027 and about 5.5% in 2028. Citi expects that, driven jointly by continuous learning, high-density eSSD, Nvidia KV Cache offloading, and demand for personal AI and physical AI after 2028, global memory supply tightness may further deteriorate in 2028 and persist until 2031.
