AI "continuous learning" may start a new storage cycle! Citi: Demand for HBM, server DRAM, and eSSD is surging, shortage may continue until 2031

According to Odaily, Citigroup stated in its latest global semiconductor industry report that as artificial intelligence advances from traditional training and inference modes into the "continual learning" era, the global memory chip market could experience a new round of structural demand growth. Unlike current AI models that primarily rely on one-time training, continual learning requires models to continuously absorb new data, update knowledge, and retain and utilize previously learned information. This suggests HBM, server DDR5, and enterprise-level SSDs (eSSD) may all see simultaneous demand expansion, intensifying supply shortages in the global memory market in 2028, potentially lasting until 2031.

Citigroup believes continual learning will become one of the key themes in AI development over the next five years. In this model, AI models do not remain relatively static after training but are able to undergo incremental training based on new tasks, data, and knowledge, while avoiding the loss of previously obtained information. The resultant ongoing model updates and requirements for accessing historical data will significantly increase AI systems' consumption of computing and storage resources.

The firm expects that starting from 2027, AI training and inference demands for storage are likely to accelerate in tandem. HBM will benefit from ongoing training and model updates, server DDR5 and SoCAMM2 from increased AI inference and CPU workloads, while eSSD will support the long-term storage needs of large volumes of historical data and KV Cache.

Given this trend, Citigroup names Samsung Electronics, SK hynix (SKHY.US), Micron Technology (MU.US), SanDisk (SNDK.US), and Kioxia as its top picks in the memory space; for semiconductor equipment and materials, the bank favors Montage Technology, Applied Materials (AMAT.US), Lam Research (LRCX.US), TES, Eugene Technology, and TechWing.

AI Enters the “Continual Learning” Era: Changing Logic of Storage Demand

Citigroup points out a significant limitation of current AI models: although these models can retain the knowledge learned during training, they do not automatically integrate subsequent user interactions or continuously generated new data into model parameters.

Continual learning seeks to address this issue. Models need to continuously be trained on new tasks and knowledge while retaining previously acquired information—the greatest technical challenge being to avoid so-called "catastrophic forgetting," where models lose previous capabilities while learning new things.

This transition may reshape the entire AI memory industry chain.

Between 2022 and 2024, with the expansion of AI training model sizes, HBM demand saw explosive growth first, while traditional DRAM and NAND demand remained relatively weak. But from the second half of 2025, as AI inference tasks become more complex, server DDR5 and eSSD demand started to grow rapidly.

Citigroup predicts the first stage of continual learning will arrive in the first half of 2027-2028. During this period, AI servers will require not only ongoing training and model updates but also will need to handle increasingly complex inference tasks, leading to the first synchronous rapid growth in demand for HBM, server DRAM, SoCAMM2, and eSSD.

After the second half of 2028, personal AI and Physical AI may become new drivers for demand. With the proliferation of edge AI, local data management, and robotics, end devices will need persistent memory and the ability to continually adapt to their environment, possibly adding local and edge storage into this growth cycle.

It is worth noting that AI Token usage has already shown rapid expansion. Page 3 of the report shows that from January 2025 to August 2026, the compound monthly growth rate (MoM) of AI Token usage is about 31%, with a year-on-year increase of 2,434% as of August 2026. Citigroup believes this trend further supports the judgment of a long-term rise in AI workloads and storage intensity.

HBM Demand May Increase 62% in 2027, 69% in 2028

Among all storage products, HBM remains one of the primary beneficiaries of AI training and continual learning.

Despite recent market concerns regarding AI chips reducing HBM specifications and AI security issues, Citigroup asserts that this does not indicate an overall weakening of HBM demand. On the contrary, AI chipmakers are shifting from "scale-up" to "scale-out" architectures, expanding total computing power by deploying more accelerators, which will continue to drive system-level HBM demand upward.

Citigroup expects that HBM bit demand will grow 62% year-on-year to 75.2 billion Gb in 2027 and further by 69% to 127 billion Gb in 2028, about twice previous forecasts. In addition to NVIDIA, Broadcom (AVGO.US), Google, and other ASIC chip demands will also become important sources of growth.

More importantly, even if memory manufacturers aggressively expand capacity, HBM supply may still fail to keep up with demand.

Citigroup projects that HBM demand will rise from 46.4 billion 1Gb equivalent units in 2026 to 75.2 billion in 2027, while supply will only go from 36.1 billion to 59.3 billion, leaving a supply gap of around 21% in 2027; by 2028, with further increases in ASIC shipments, this gap could widen to about 36%.

Therefore, Citigroup views the recent so-called HBM "spec reduction" more as a strategy to improve resource efficiency and maximize AI accelerator shipments under supply constraints, rather than a signal of weakening demand.

AI Inference Becomes Second Growth Engine; Server DRAM Demand May Surge 51% in 2027

If HBM primarily benefits from AI training, server DRAM is expected to be a major beneficiary of the expansion in AI inference and continual learning.

Driven by continual learning, Citigroup estimates server DRAM demand will grow by more than 50% in 2027, while eSSD demand could also rise by about 50%. As the complexity of AI inference continues to increase, both CPU demand and server memory capacity will further expand.

Specifically, the firm expects server DRAM demand to increase from 226.3 billion 1Gb equivalent units in 2026 to 341.7 billion in 2027, a year-on-year increase of about 51%. At that time, servers will account for roughly 67% of global DRAM demand, remaining the largest application market.

This trend also correlates with AI chips lowering individual accelerator HBM configuration. When some KV Cache and other memory workloads are offloaded from HBM to server DRAM or external storage, it may further boost server DRAM and eSSD demand.

Citigroup notes that memory customers are extending the terms of some long-term supply agreements (LTA) from three years to five, reflecting increased visibility for future memory demand in the market.

eSSD Demand May Explode, Estimated to Grow Nearly 53% in 2027

The NAND market may likewise become a major beneficiary in the era of continual learning.

Continual learning not only requires AI models to constantly learn new information but also to retain vast amounts of previously acquired data, thereby demanding larger long-term storage capacities.

Citigroup anticipates that global NAND demand will grow by 29.1% in 2027, outpacing the 21.2% supply growth. Specifically, enterprise SSD demand is expected to rise by 52.9% year-on-year, significantly outpacing the overall SSD growth rate of about 45%.

One important factor driving this growth is KV Cache offloading. As more AI accelerators reduce HBM configuration, an increasing amount of KV Cache and other memory workloads may be shifted to external storage, boosting demand for high-capacity QLC enterprise SSDs.

Citigroup also points out that AI servers are adding storage configurations closer to the GPU, including the use of QLC SSDs, allowing more data to be stored near accelerators, thereby improving model memory and data access efficiency.

Meanwhile, NAND demand for consumer electronics may be dragged down by relatively weak smartphone and PC markets; however, Citigroup expects the enterprise demand growth driven by AI inference to be sufficient to offset this impact. The report also mentions that AI data centers, including those in China, are increasingly considering replacing HDD with SSD, which could further expand potential eSSD demand from 2027 onwards.

DRAM Supply-Demand Gap Could Widen to 9.7% in 2028, Price Cycle Likely to Continue

On the other side of surging demand, capacity expansion in the memory industry remains constrained.

Citigroup projects that global DRAM bit demand will grow by about 30.2% in 2027, significantly higher than the supply growth of 18.8%, shifting the market from a near-balance in 2026 to a supply shortage with a gap of about 8.7%. By 2028, this supply gap may further widen to about 9.7%.

Capacity limitations are mainly due to two factors. On one hand, Samsung Electronics, SK hynix, and Micron are allocating more DRAM capacity to HBM; on the other hand, slowing migration to advanced processes and lengthy construction cycles for new fabs have prevented rapid release of additional bit supply.

Citigroup expects industry average wafer capacity for DRAM to grow only about 8% year-over-year in 2027, while bit supply growth stands at about 18.8% for the full year, suggesting tight supply throughout 2027.

Tight supply-demand dynamics also mean that memory prices are likely to remain strong.

Citigroup forecasts that after a 242.4% surge in average DRAM prices in 2026, DRAM average selling prices (ASP) may still see a further year-on-year increase of 23.1% in 2027. Although this is a slower pace compared to the extreme rise in 2026, continued undersupply of HBM and traditional DRAM will keep prices on an uptrend.

NAND Also Enters Tight Supply Phase, Shortages May Persist Until 2031

NAND market supply pressure is mounting as well.

Citigroup projects that in 2027, NAND demand will rise by about 29% while supply will only grow about 21%, resulting in a supply gap of around 6.1%. In 2028, demand is expected to rise by 33%, still outpacing supply growth of 25%, keeping the market in shortage with a gap of about 5.5%.

One key reason for this is that global memory manufacturers are currently prioritizing expansion of DRAM and HBM production, which limits additional NAND capacity.

Citigroup expects that under the combined influence of continual learning, high-density eSSD, NVIDIA KV Cache offloading, and post-2028 demand from personal and physical AI, global memory tightness could further intensify in 2028 and persist until 2031.

This means the current memory cycle may no longer be a short-term cycle driven by traditional factors like smartphone and PC restocking, but is increasingly characterized by structural demand driven by changes in AI compute architecture.

Memory Manufacturers Ramp Up Expansion; Equipment & Materials Companies Benefit Simultaneously

Facing long-term expanding demand, memory manufacturers are initiating larger-scale capital expenditure, which may also further extend AI memory investment opportunities across the semiconductor equipment and materials value chain.

Citigroup projects that global DRAM and NAND capital expenditure will rise from about $54.9 billion in 2026 to $80.4 billion in 2027, a 46.5% year-on-year jump; DRAM capex alone is expected at $58.6 billion, up 51.6%, and NAND capex at $21.8 billion, up 34.2%.

Looking further ahead, the report expects total global DRAM and NAND capex could reach about $322.5 billion by 2031, with approximately $254 billion in DRAM and $68.5 billion in NAND.

Thus, Citigroup argues this memory upcycle will benefit not only memory chip makers, but also gradually benefit semiconductor equipment and materials suppliers.

In storage, the bank names Samsung Electronics, SK hynix, Micron Technology, SanDisk, and Kioxia as top picks; in equipment and materials, Montage Technology, Applied Materials, Lam Research, TES, Eugene Technology, and TechWing are favored. Citigroup believes that as continual learning and medium- to long-term development of personal AI drive structural growth in memory demand, memory manufacturers will directly benefit from sustained supply shortages, while equipment and materials companies will profit from large-scale capital spending and higher fab utilization rates.