Citi: AI Continuous Learning Drives HBM/Server DRAM/eSSD Supply Shortages Extending Through 2031
nashnova research
Citi's latest report projects that AI's shift to continuous learning will strain HBM, server DRAM, and enterprise SSD supply lines simultaneously, with shortages potentially deepening through 2028 and lasting until 2031 — marking a structural shift from short-cycle restocking to architecture-driven demand.
What is "continuous learning," and why does it consume more memory than training?
Continuous learning — a mode where AI models absorb new data while retaining old knowledge — faces a core challenge: catastrophic forgetting, where learning new information erases what came before.
This means → the model must store both old and new data at the same time, consuming far more compute and memory than one-off training.
Citi sees this as the defining AI theme for the next five years, directly driving demand across HBM, DRAM, and eSSD.
How big is the HBM gap — and what does "down-speccing" really mean?
Citi forecasts 2027 HBM bit demand rising 62% year-on-year to 75.2 billion Gb, then another 69% in 2028 to 127 billion Gb — roughly double its prior estimate.
Demand growth extends beyond Nvidia: Broadcom, Google, and other ASIC chips — custom-designed AI processors — are becoming major drivers.
Supply cannot keep up: the 2027 gap is roughly 21%; by 2028, as ASIC shipments ramp, it could widen to about 36%.
Citi reads the recent HBM "down-speccing" trend as a supply-constrained efficiency measure, not a demand weakness signal. In plain terms = builders aren't choosing lower specs — higher specs simply aren't available, so they adapt.
Server DRAM: why does "offloading" increase demand instead of relieving it?
Citi expects 2027 server DRAM demand to jump roughly 51% year-on-year; servers would then account for about 67% of global DRAM demand.
When workloads like KV cache — memory regions that store intermediate results during AI inference — shift from HBM to server DRAM, they add to DRAM and eSSD demand rather than substituting.
This reflects a key dynamic: AI memory demand is expanding at every tier simultaneously, not shifting between tiers in a zero-sum fashion.
Citi also notes that memory customers are extending long-term supply agreements from three years to five — This means → the market's confidence in multi-year demand visibility is rising.
eSSD: who picks up the overflow from HBM?
Citi projects 2027 enterprise SSD demand growth at 52.9% year-on-year, outpacing the broader SSD market at roughly 45%.
The core driver is KV cache offloading: as AI accelerators reduce HBM allocation, more workloads migrate to external storage, boosting demand for high-capacity QLC enterprise SSDs — solid-state drives using four-layer storage cells for greater density.
Additionally, AI data centers — including those in China — are increasingly considering replacing HDDs with SSDs, potentially expanding the eSSD market further from 2027.
NAND supply is also tightening: a 6.1% gap is expected in 2027, largely because memory makers are prioritizing DRAM and HBM capacity expansion, squeezing NAND investment.
Capex and pricing: where is the money flowing?
Citi expects global DRAM and NAND capex to rise from roughly $54.9 billion in 2026 to $80.4 billion in 2027 — up 46.5% year-on-year.
DRAM capex alone is forecast at $58.6 billion (+51.6% YoY); NAND at $21.8 billion (+34.2% YoY). By 2031, combined capex could reach approximately $322.5 billion.
On pricing, after DRAM blended ASPs surged 242.4% in 2026, Citi still expects a further 23.1% rise in 2027.
This means → the memory profit cycle is likely to run longer than past cycles — because this time the driver is not smartphone or PC restocking, but a structural shift in AI compute architecture.
Which names does Citi favor — and what is the key proof point?
Top memory picks: Samsung Electronics, SK Hynix, Micron Technology, SanDisk, and Kioxia.
In equipment and materials: Montage Technology, Applied Materials, Lam Research, TES, Eugene Technology, and TechWing.
Citi's core thesis is that this memory cycle has shifted from a traditional short cycle to an AI-driven structural demand cycle.
In plain terms = Citi is betting that the AI memory shortage will persist for years — and whether that bet pays off will hinge on actual shipment data in 2027–2028, the cycle's most critical proof point.
市场有风险，内容仅供研究参考，不构成投资建议。