A structural leap in AI's ability to retain and apply knowledge is pushing the global memory market into a multi-year supply deficit cycle.

In a September 14 report, Citi analysts highlighted that as AI transitions from pure training and inference into an era of "continual learning," demand for HBM, server DDR5, and enterprise solid-state drives (eSSD) is set to surge simultaneously starting in 2027. Citi projects the DRAM supply-demand ratio will deteriorate to -8.7% and -9.7% in 2027 and 2028, respectively, while the NAND balance will slip to -6.1% and -5.5% over the same period, with the imbalance expected to persist through 2031.

On the supply side, capacity expansion is severely lagging demand due to HBM's heavy occupancy of fab capacity and slower technology migration. Citi estimates 2027 global memory capital expenditure will jump 46.5% year-over-year to $80.4 billion, yet given the long lead times inherent in capacity construction, even aggressive investment will struggle to close the gap.

Continual Learning: The New Paradigm Driving AI Memory Demand

Citi Research identifies continual learning as the defining AI theme over the next five years. Unlike current models that freeze parameters after training, continual learning allows systems to absorb new knowledge during operation while retaining previously learned information.

The core challenge lies in balancing new knowledge acquisition against preventing "catastrophic forgetting." This demands continuous model updates coupled with efficient access to historical data—both of which impose significantly higher requirements on memory systems.

Tracing the evolution of AI workloads, the 2022 to mid-2025 period was largely dominated by AI training, driving strong HBM demand while traditional DRAM and NAND remained tepid. From the second half of 2025, rising inference complexity has spurred rapid growth in server DDR5 and eSSD demand, cooling some enthusiasm for HBM.

Citi believes 2027 will mark the first phase of continual learning, with demand for HBM, server DRAM, SoCAMM2, and eSSD expanding in tandem. By the second half of 2028, the rise of personal AI and Physical AI will further extend the demand landscape into edge-device memory.

Notably, monthly AI token usage is following an almost perfect exponential curve, with a month-over-month compound growth rate of 31% and a year-over-year surge of 2,434% as of August 2026, providing robust underlying support for sustained high growth in overall memory demand.

HBM: De-Specification Reflects Supply Constraints, Not Weak Demand

Citi offers a clear interpretation of the recent HBM "de-specification" trend: it is a resource efficiency optimization by AI chipmakers under constrained HBM supply, not a signal of softening end-demand.

Citi forecasts HBM bit demand will surge 62% year-over-year to 75.157 billion Gb in 2027, followed by a further 69% jump to 126.990 billion Gb in 2028—double its previous projections. Demand drivers extend beyond Nvidia to ASIC manufacturers like 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 grows from 4.702 billion Gb to 14.003 billion Gb over the same period.

On the supply front, Citi expects HBM capacity (measured by TSV packaging) to expand from 420,000 wafers per month in 2026 to 700,000 in 2027. Even so, the supply-demand gap will remain elevated: the HBM supply-demand ratio is projected at -21% in 2027, roughly flat from -22% in 2026, deteriorating further to -36% in 2028 as ASIC shipments accelerate.

Citi argues that de-specification is a pragmatic choice by chipmakers to maximize accelerator shipments under limited HBM supply and should not be misread as a demand reversal.

DRAM: Historic Supply-Demand Gap Looms in 2027

Citi projects 2027 global DRAM demand will grow 30.2% year-over-year, while supply expands just 18.8%, driving the supply-demand ratio from +0.5% in 2026 to -8.7%, worsening to -9.7% in 2028.

On the demand side, server DRAM growth driven by continual learning is particularly pronounced. Citi forecasts server DRAM demand will leap from 226.3 billion units (1Gb equivalent) in 2026 to 341.7 billion units in 2027, up 51% year-over-year, with servers continuing to account for roughly 67% of total DRAM demand. PC demand is expected to remain relatively subdued, while smartphone demand sees a modest recovery.

Supply constraints are equally clear. First, HBM capacity occupancy continues to crowd out general-purpose DRAM wafer capacity, with industry average DRAM wafer capacity projected to grow only about 8% to 2.295 million wafers per month in 2027. Second, slower technology node migration limits bit output growth per wafer. Third, the lengthy cycle from greenfield project initiation to mass production means supply response will lag even as capital spending accelerates.

Citi predicts DRAM blended average selling prices (ASP) will rise 23.1% year-over-year in 2027, extending the upward trend following the extraordinary 242.4% surge in 2026, though the pace will normalize.

NAND: Strong eSSD Demand Meets Restrained Supply, Widening the Gap

The NAND market is also facing substantial tightening. Citi estimates 2027 NAND demand will grow 29.1% year-over-year, while supply expands just 21.2%, driving the supply-demand ratio from -0.8% in 2026 to -6.1% in 2027, holding at -5.5% in 2028.

eSSD is the primary growth engine for NAND demand. Citi forecasts eSSD demand will grow 52.9% year-over-year in 2027 (with overall SSD growth at 45.0%), driven by three core factors. First, reduced HBM content per accelerator forces more KV cache workloads to migrate to external storage. Second, increasingly complex AI inference workloads push high-capacity QLC eSSD penetration into AI servers. Third, Nvidia's KV cache offloading trend and the shift from HDD to SSD in Chinese AI data centers provide additional upside.

Supply constraints deserve equal attention. Citi expects industry average NAND wafer capacity to grow just 3.2% in 2027 to 1.460 million wafers per month, with Samsung Electronics' NAND capacity actually projected to decline 4.7%, while SK Hynix and Micron stay roughly flat. The primary reason is that manufacturers are channeling more resources toward DRAM/HBM greenfield expansion while remaining cautious on incremental NAND investment.

On pricing, Citi expects NAND blended ASP to rise 45.3% year-over-year in 2027, continuing the upward trajectory following the 236.2% surge in 2026, albeit at a moderating pace.

Capital Spending Jumps, But Insufficient to Close the Mid-Term Gap

Citi projects 2027 global memory capital expenditure (DRAM+NAND) will surge 46.5% year-over-year to $80.4 billion, a significant acceleration from $54.9 billion in 2026.

DRAM capital spending is expected to grow 51.6% year-over-year to $58.6 billion, led by Samsung Electronics ($20.6 billion), SK Hynix ($17.5 billion), and Micron ($15.8 billion). However, Citi emphasizes that a substantial portion of incremental DRAM investment will still flow toward HBM capacity and advanced DRAM process nodes, offering limited near-term relief for general-purpose DRAM supply.

NAND capital spending is projected to grow 34.2% year-over-year to $21.8 billion, with Samsung Electronics investing $7.9 billion, Kioxia $3.7 billion, SK Hynix $3.1 billion, Sandisk $2.9 billion, and Micron $2.1 billion.

Citi notes that despite the sharp jump in capital expenditure, the lengthy lead times for greenfield capacity expansion mean higher spending levels will not suffice to close the 2027 supply-demand gap. Based on this analysis, Citi extends its projection of global memory supply-demand imbalance through 2031, arguing that the scaled deployment of continual learning and personal AI/Physical AI will serve as the core pillar sustaining this ultra-long-term structural shortage.