Will AI that "continuously learns" extend the memory shortage to 2031?

The structural leap in AI memory capabilities is pushing the global memory market into a multi-year supply shortage cycle.

According to a report by Citigroup on September 14, as AI moves from the simple training and inference stage to the "continuous learning" era, demand for HBM, server DDR5, and enterprise-grade solid-state drives (eSSDs) will experience explosive growth starting in 2027. Citigroup predicts that the DRAM supply-demand ratio (S/D ratio) will deteriorate to -8.7% and -9.7% in 2027 and 2028, respectively, while the NAND supply-demand ratio will also fall to -6.1% and -5.5% during the same period, with the supply-demand imbalance expected to continue until 2031.

While demand is surging, supply is constrained by slowing HBM capacity utilization and technology migration, with capacity expansion lagging far behind demand. Citigroup predicts that global memory capital expenditure will increase by 46.5% year-on-year to $80.4 billion in 2027, but due to the long lead times for capacity construction, even large-scale investment will struggle to fill the supply-demand gap.

Continuous Learning: A New Driving Paradigm for AI Memory Demand

Citi research indicates that "continuous learning" will be a core theme in the AI field over the next five years. Unlike current AI models that fix their parameters after training, continuous learning allows models to continuously absorb new knowledge while retaining previously learned information.

The core challenge of this approach lies in balancing the learning of new knowledge with the prevention of "catastrophic forgetting." To this end, AI systems need to continuously update their models while maintaining efficient access to historical data—both of which place significantly higher demands on memory.

Looking at the historical trajectory of AI workload evolution, the period from 2022 to the first half of 2025 was mainly the AI training cycle, with strong demand for HBM, while traditional DRAM and NAND were relatively flat. Starting from the second half of 2025, the increase in inference complexity will drive rapid growth in demand for server DDR5 and eSSD, and the market's attention to HBM will cool down.

Citigroup believes that the first phase of continuous learning will begin in 2027, with demand for HBM, server DRAM, SoC AMM2, and eSSD expanding in tandem. By the second half of 2028, with the rise of personal AI and physical AI, the demand landscape will further expand to edge device memory.

It is worth noting that the monthly usage of AI tokens shows a near-perfect exponential curve, with a month-on-month compound growth rate of 31% and a year-on-year growth rate of 2434% in August 2026. This provides strong underlying support for the continued high growth of overall memory demand.

HBM: Denormalization is a result of supply constraints, not weak demand.

Citigroup has provided a clear interpretation of the recent HBM "de-specification" phenomenon that has attracted industry attention: this is a resource efficiency optimization behavior by AI chip manufacturers in the context of limited HBM supply, rather than a signal of weakening end-user demand.

Citigroup projects that HBM bit demand will surge 62% year-on-year to 75.157 billion Gb in 2027 and further jump 69% to 126.99 billion Gb in 2028, double the previous forecast. The demand drivers include not only Nvidia but also ASIC manufacturers such as Broadcom and Google—Broadcom's HBM demand forecast expands from 9.002 billion Gb in 2026 to 41.003 billion Gb in 2028, while Google's forecast increases from 4.702 billion Gb to 14.003 billion Gb during the same period.

On the supply side, Citigroup expects HBM capacity (measured by TSV packaging capacity) to expand to 700,000 wafers/month in 2027 from 420,000 wafers/month in 2026. Even so, the supply-demand gap will remain high: the HBM S/D ratio is expected to be -21% in 2027, roughly the same as -22% in 2026, and will further deteriorate to -36% by 2028 due to accelerated ASIC shipments.

Citigroup believes that destandardization is essentially a pragmatic choice for chip manufacturers to maximize accelerator shipments under limited HBM supply, and should not be interpreted as a shift in demand.

DRAM: The supply-demand gap will reach a historic low in 2027.

Citigroup predicts that global DRAM demand will grow by 30.2% year-on-year in 2027, but supply growth will be only 18.8%, with the supply-demand ratio plummeting from +0.5% in 2026 to -8.7%, and further deteriorating to -9.7% in 2028.

On the demand side, the growth in server DRAM demand driven by continuous learning is particularly prominent. Citi predicts that server DRAM demand will jump from 226.3 billion units (1Gb equivalent) in 2026 to 341.7 billion units in 2027, representing a year-on-year increase of 51%, with servers continuing to account for approximately 67% of total DRAM demand. In contrast, PC demand is expected to remain relatively flat, while mobile phone demand is expected to recover moderately.

Supply-side constraints are also clear: on the one hand, HBM capacity occupancy continues to squeeze general-purpose DRAM wafer capacity, with the industry's average DRAM wafer capacity expected to grow by only about 8% to 2.295 million wafers per month in 2027; on the other hand, the slowdown in technology node migration limits the growth rate of bit output per wafer; in addition, the long cycle from project initiation to mass production of greenfield capacity means that even if capital expenditure is accelerating, supply response will lag behind.

Citigroup predicts that the average price of DRAM hybrids will increase by 23.1% year-on-year in 2027. After an abnormal increase of 242.4% in 2026, the upward trend in prices will continue, but the rate of increase will tend to normalize.

NAND: Strong demand for eSSDs coupled with restrained supply has significantly widened the supply-demand gap.

The supply and demand dynamics of the NAND market are also facing a substantial tightening. Citigroup predicts that NAND demand will grow by 29.1% year-on-year in 2027, while supply growth will be only 21.2%, and the S/D ratio will plummet from -0.8% in 2026 to -6.1%, remaining at -5.5% in 2028.

eSSD is the core driver of NAND demand growth. Citi predicts that eSSD demand will grow by 52.9% year-on-year in 2027 (compared to 45.0% for overall SSD), based on three main reasons: First, the decreasing capacity per HBM accelerator is forcing more KV caching workloads to migrate to external storage; second, the increasing complexity of AI inference workloads is driving the penetration of high-capacity QLC eSSDs in AI servers; and third, Nvidia's KV cache offloading trend and the trend of replacing HDDs with SSDs in Chinese AI data centers provide additional upward demand potential.

Supply-side constraints also warrant attention. Citigroup projects that the industry's average NAND wafer capacity will only increase slightly by 3.2% to 1.46 million wafers per month in 2027, with Samsung Electronics' NAND capacity even expected to decline by 4.7%, while SK Hynix and Micron's will remain largely flat. The main reason is that manufacturers are allocating more resources to greenfield expansion of DRAM/HBM, adopting a cautious approach to incremental investment in NAND.

Regarding NAND prices, Citi expects the average price of NAND hybrids to increase by 45.3% year-on-year in 2027, continuing the upward trend following the 236.2% increase in 2026, although the pace of increase will slow down.

Capital expenditures surged, but failed to bridge the medium-term gap.

Citigroup predicts that global memory capital expenditure (DRAM+NAND) will increase by 46.5% year-on-year to US$80.4 billion in 2027, a significant acceleration from US$54.9 billion in 2026.

DRAM capital expenditure is projected to increase by 51.6% year-on-year to $58.6 billion, led by Samsung Electronics ($20.6 billion), SK Hynix ($17.5 billion), and Micron ($15.8 billion). However, Citigroup emphasizes that a significant proportion of incremental DRAM investment will still flow to HBM capacity and advanced DRAM processes, with limited actual relief to the supply of general-purpose DRAM.

NAND capital expenditures are expected to increase by 34.2% year-on-year to $21.8 billion, of which Samsung Electronics is expected to invest $7.9 billion, Kioxia $3.7 billion, SK Hynix $3.1 billion, Sandisk $2.9 billion, and Micron $2.1 billion.

Citigroup points out that despite a significant surge in capital expenditures, the higher spending levels are insufficient to bridge the supply-demand gap in 2027, given the lengthy lead times for greenfield investment in capacity expansion. Based on this, Citigroup extends its assessment of the global memory supply-demand imbalance to 2031, believing that continuous learning and the large-scale deployment of personal/physical AI will be the core support for this ultra-long-term structural supply shortage.

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.