AI 'Continuous Learning' May Spark a New Memory Cycle! Citi: Demand for HBM, Server DRAM, and eSSD Rising Together; Shortages May Persist Until 2031

Citi said in its latest global semiconductor industry report that as artificial intelligence moves beyond traditional training and inference models into the era of "continuous learning," the global memory chip market may usher in a new round of structural demand growth. Unlike current AI models, which mainly rely on one-time training, continuous learning requires models to constantly absorb new data, update their own knowledge, and at the same time retain and recall previously learned information. This means HBM, server DDR5, and enterprise SSDs (eSSD) may see simultaneous demand expansion, pushing the global memory market supply shortage to intensify further in 2028 and even persist until 2031. Citi believes continuous learning will become one of the key themes of AI development over the next five years. Under this model, AI models are no longer relatively static after training is completed, but can continuously conduct incremental training based on new tasks, data, and knowledge while avoiding forgetting previously acquired information. The resulting continuous model updates and historical data access requirements will significantly increase AI systems' consumption of computing and storage resources. The bank expects that starting in 2027, storage demand from AI training and inference may accelerate simultaneously. Among these, HBM will benefit from continuous training and model updates, server DDR5 and SoCAMM2 will benefit from AI inference and increased CPU workloads, while eSSD will take on long-term storage demand for large amounts of historical data and data such as KV Cache. Under this trend, Citi lists Samsung Electronics, SK Hynix (SKHY.US), Micron Technology, Inc. (MU.US), SanDisk (SNDK.US), and Kioxia as top picks in the storage sector; in the semiconductor equipment and materials sector, it is bullish on Montage Technology, Applied Materials (AMAT.US), Lam Research (LRCX.US), TES, Eugene Technology, and TechWing. AI Enters the "Continuous Learning" Era, the Logic of Storage Demand Is Changing Citi points out that current AI models have an important limitation: although models can retain knowledge learned during training, they do not automatically integrate users' subsequent interactions and continuously generated new data directly into model parameters. Continuous learning attempts to solve this problem. Models need to continuously receive training on new tasks and knowledge while retaining information they have already mastered. One of the biggest technical challenges is avoiding so-called "catastrophic forgetting," in which a model loses its original capabilities while learning new knowledge. This change may reshape the entire AI storage industry chain. From 2022 to 2024, as the scale of AI training models continued to expand, HBM demand was the first to experience explosive growth, while demand for traditional DRAM and NAND was relatively weak. But since the second half of 2025, as AI inference tasks have become more complex, demand for server DDR5 and eSSD has begun to grow rapidly. Citi expects that from 2027 to the first half of 2028, the first phase of continuous learning will begin. During this period, AI servers will need both continuous training and model updates and the execution of increasingly complex inference tasks, so demand for HBM, server DRAM, SoCAMM2, and eSSD may grow rapidly in sync for the first time. After the second half of 2028, personal AI and Physical AI may further become new demand DRIVE. As applications such as edge AI, local data management, and Siasun Robot&Automation gradually become widespread, terminal devices will need to have persistent memory and continuously adapt to their surroundings. At that point, local and edge storage demand may also join this growth cycle. It is worth noting that AI Token usage has already shown rapid expansion. Data on page 3 of the report shows that from January 2025 to August 2026, the compound monthly growth rate of monthly AI Token usage was about 31%, and the year-over-year increase in August 2026 reached 2,434%. Citi believes this trend further supports the judgment that AI workloads and storage intensity will rise over the long term. HBM Demand May Grow 62% in 2027 and Another 69% in 2028 Among all storage products, HBM remains one of the most direct beneficiaries of AI training and continuous learning. Although the market has recently been concerned about AI chips reducing HBM configuration specifications and AI security issues, Citi believes this does not mean overall HBM demand is beginning to weaken. On the contrary, AI chip manufacturers are shifting from "Scale-up" to "Scale-out" architectures, that is, expanding overall computing power by deploying more accelerators, which will still drive continued growth in system-level HBM demand. Citi expects HBM bit demand to grow 62% year over year in 2027 to 75.2 billion Gb, and to grow another 69% year over year in 2028 to 127 billion Gb, equivalent to about twice its previous forecast. In addition to NVIDIA Corporation, demand for ASIC chips from Broadcom Inc. (AVGO.US), Alphabet Inc. Class C, and others will also become important sources of growth. More importantly, even if memory manufacturers actively expand production, HBM supply may still fail to catch up with demand. Citi expects HBM demand to increase from 46.4 billion 1Gb-equivalent units in 2026 to 75.2 billion in 2027, while supply will increase from 36.1 billion to 59.3 billion. The bank expects HBM supply will still be significantly insufficient in 2027, with a supply gap of about 21%; by 2028, as ASIC shipments grow further, this gap may widen further to about 36%. Therefore, Citi views the recent so-called HBM "spec reduction" more as a response to insufficient supply to improve resource efficiency and maximize AI accelerator shipments, rather than a signal of weak demand. AI Inference Becomes a Second Growth Engine; Server DRAM Demand May Surge 51% in 2027 If HBM mainly benefits from AI training, then server DRAM is expected to become an important beneficiary of AI inference and the expansion of continuous learning. Citi expects that under the push from continuous learning, server DRAM demand will grow by more than 50% in 2027, while eSSD demand may also grow by about 50%. As AI inference tasks continue to increase in complexity, CPU demand and server memory capacity will both increase further. Specifically, the bank expects server DRAM demand to grow from 226.3 billion 1Gb-equivalent units in 2026 to 341.7 billion in 2027, an increase of about 51% year over year. By then, servers will contribute about 67% of global DRAM demand, continuing to be the largest application market. This trend is also linked to AI chips reducing HBM configuration per accelerator. When some memory workloads such as KV Cache are offloaded from HBM to server DRAM or external storage, this may instead further increase demand for server DRAM and eSSD. Citi points out that memory customers are extending some long-term agreements (LTAs) from three years to five years, which also reflects improving market visibility into memory demand for many years ahead. eSSD Demand May Explode, Expected to Grow Nearly 53% in 2027 The NAND market may also become an important beneficiary of the continuous learning era. Continuous learning means AI models not only need to constantly learn new information, but must also retain large amounts of data already acquired in the past, thus requiring greater long-term storage capacity. Citi expects global NAND demand to grow 29.1% year over year in 2027, higher than supply growth of 21.2%; among this, enterprise SSD demand is expected to grow 52.9% year over year, significantly higher than overall SSD growth of about 45%. One important factor driving this growth is KV Cache offloading. As some AI accelerators reduce HBM configuration, more and more KV Cache and other memory workloads may be transferred to external storage, so high-capacity QLC enterprise SSDs are expected to gain more demand. Citi also points out that AI servers are increasing storage configuration near GPUs, including the use of QLC SSDs, so that more data can be stored near accelerators, thereby improving model memory and data access efficiency. At the same time, consumer electronics NAND demand may be dragged down by relatively weak smartphone and PC markets, but Citi expects enterprise demand growth brought by AI inference to be enough to offset this impact. The report also mentions that AI data centers, including those in China, are increasingly considering using SSDs to replace HDDs, which may further expand potential eSSD demand starting in 2027. DRAM Supply-Demand Gap May Widen to 9.7% in 2028; Price Increase Cycle Expected to Continue The other side of rapid demand growth is that supply expansion in the memory industry remains constrained. Citi expects global DRAM bit demand to grow about 30.2% year over year in 2027, significantly higher than supply growth of 18.8%, pushing the market from roughly balanced in 2026 to a supply shortage, with a gap of about 8.7%; by 2028, the supply gap is expected to widen further to about 9.7%. Supply constraints mainly come from two factors. On the one hand, Samsung Electronics, SK Hynix, and Micron are allocating more DRAM capacity to HBM; on the other hand, the pace of migration to advanced processes has slowed, and new wafer fabs require long construction cycles, so additional bit supply cannot be released quickly. Citi expects average wafer capacity in the DRAM industry to grow only about 8% year over year in 2027, while full-year DRAM bit supply grows about 18.8%, so supply tightness may persist throughout 2027. Tight supply and demand also mean memory prices may remain strong. Citi expects that after DRAM blended average selling prices surge 242.4% in 2026, DRAM blended ASP may still rise 23.1% year over year in 2027. Although the increase is significantly slower than the extreme level in 2026, continued shortages of HBM and traditional DRAM will still support further price increases. NAND Also Enters a Phase of Supply Tightness; Shortage May Persist Until 2031 Supply pressure in the NAND market is also increasing. Citi expects NAND demand to grow about 29% year over year in 2027, while supply grows only about 21%, with a supply gap expected to reach about 6.1%; in 2028, demand is expected to grow 33%, continuing to exceed supply growth of 25%, and the market will remain in supply shortage, with a gap of about 5.5%. An important reason for this situation is that global memory manufacturers are currently prioritizing expansion of DRAM and HBM capacity, thereby limiting new NAND capacity. Citi expects that, driven jointly by continuous learning, high-density eSSD, NVIDIA Corporation KV Cache offloading, and demand for personal AI and physical AI after 2028, global memory supply tightness may worsen further in 2028 and continue until 2031. This means the current memory cycle may no longer be just a short-term cycle driven by traditional factors such as smartphone and PC restocking, but increasingly shows structural demand characteristics driven by changes in AI computing architecture. Memory Manufacturers Expand Production Aggressively; Equipment and Materials Companies Benefit in Tandem Facing long-term demand expansion, memory manufacturers are launching larger-scale capital expenditure, and this may also further spread AI storage investment opportunities to the semiconductor equipment and materials industry chain. Citi expects global DRAM and NAND capital expenditure to increase from about $54.9 billion in 2026 to $80.4 billion in 2027, up 46.5% year over year; among this, DRAM capital expenditure is expected to reach $58.6 billion, up 51.6% year over year, while NAND capital expenditure will reach $21.8 billion, up 34.2% year over year. Over the longer term, the report estimates global DRAM and NAND capital expenditure may further reach about $322.5 billion by 2031, including about $254 billion for DRAM and about $68.5 billion for NAND. Therefore, Citi believes this round of memory upcycle will not only benefit memory chip manufacturers, but is also expected to gradually transmit to semiconductor equipment and materials manufacturers. In the storage sector, the bank lists Samsung Electronics, SK Hynix, Micron Technology, Inc., SanDisk, and Kioxia as top picks; in equipment and materials, it is bullish on Montage Technology, Applied Materials, Lam Research, TES, Eugene Technology, and TechWing. Citi believes that as continuous learning and medium- to long-term personal AI development drive structural growth in memory demand, memory manufacturers will directly benefit from continued supply shortages, while equipment and materials companies are expected to benefit from large-scale capital expenditure and higher fab capacity utilization.