AI Memory Shortage: Why DRAM Stays Tight Through 2027 While NAND May Ease

AI infrastructure is turning memory into a central constraint on computing growth. Demand for high-bandwidth memory (HBM) and high-density server DRAM is absorbing manufacturing and packaging capacity that would otherwise support more conventional memory products. The result is a seller-friendly DRAM market, unusually thin inventory buffers at leading suppliers, and a cost problem that reaches from AI accelerators to smartphones. 745

Samsung and SK hynix: strong demand, very little buffer

KB Securities estimated in September that Samsung Electronics and SK hynix each held less than 10 days of memory inventory during the third quarter. That figure is an analyst estimate—not a directly comparable company disclosure—so it is best read as an indicator of exceptionally tight available supply rather than a precise operating metric. 1921

The important distinction is that thin inventories do not mean weak business conditions. They reflect demand and allocation pressure: memory producers are prioritizing high-value HBM and server DRAM while customers seek supply commitments further in advance. HBM capacity constraints have also contributed to the broader memory shortage and rising prices. 2910

Why HBM4 tightens conventional memory too

HBM is stacked DRAM placed alongside AI processors to provide very high bandwidth. Nvidia’s Rubin GPU, for example, is reported to use up to 288GB of HBM4 per chip. 31

That architecture makes HBM a system-level dependency for AI accelerators. But the impact extends beyond accelerator shipments. Building more HBM requires scarce wafer capacity and sophisticated packaging, while manufacturers also direct capital toward server-class memory. As capacity shifts toward those higher-value products, less production is available for general-purpose DRAM used in PCs, phones and other devices. 2945

This is why the market has become more than a normal inventory cycle: AI memory demand can constrain the physical availability of memory, not merely raise its price.

The 2027 forecast: DRAM and NAND diverge

The best-supported forecast is not a single shortage across all memory categories.

• DRAM: TrendForce expects continued HBM allocation, AI-server demand and rising memory content per server to keep DRAM supply constrained and prices on an upward path in 2027. 7
• NAND flash: The same forecast expects NAND to move into a looser supply environment in the second half of 2027, as new capacity arrives while consumer-electronics demand remains weak. That would put downward pressure on NAND pricing. 7

Forecasts differ on the exact size of a 2027 DRAM gap. A JPMorgan-based estimate cited in market reporting projects a roughly 7% supply-demand growth gap, while other forecasts have put the projected deficit lower. The direction is more consistent than the number: DRAM is expected to remain tight in 2027. 3342

For buyers, that means SSD and storage pricing may eventually become less stressed than DRAM pricing. It also means a claim that all memory will stay scarce until 2033 is too broad. The evidence supports persistent pressure in leading-edge AI memory and DRAM, but points to a potentially easier NAND market in late 2027. 7

Price effects: HBM first, then the wider memory market

HBM4’s ramp is already associated with sharp price increases. Korea International Trade Association data reported an average HBM export price of $76.13 per unit at the end of July, up 9.5% from the prior month. 27

Conventional DRAM is affected indirectly: when supply and investment shift toward HBM and server DRAM, mainstream buyers have less flexibility and suppliers gain leverage. TrendForce expects the constrained DRAM balance to sustain a seller’s market in 2027. 7

What it means for smartphones, PCs and GPUs

Smartphones and PCs

Memory is a meaningful part of a device bill of materials, particularly for lower-cost models. Higher LPDDR and NAND procurement costs leave manufacturers with limited options: raise retail prices, offer less memory or storage at a given price point, cut other specifications, or absorb lower margins. Reporting on the handset market indicates that memory price increases and constrained supply are influencing industry scale and procurement decisions. 4551

The pressure is likely to be felt most acutely in budget devices, where there is less margin to absorb component increases.

GPUs and AI servers

For AI hardware, HBM is not simply an input whose price can be passed along. It is an essential, qualified component of the accelerator package. A shortage can therefore limit how many high-end accelerators and AI servers can ship, even when demand for GPUs is strong. 3129

That encourages cloud providers and other large buyers to secure longer-term supply agreements, reducing the amount of uncommitted capacity available to smaller or more price-sensitive customers. 10

How chip design is changing

The shortage reinforces a shift toward memory-first system design. AI performance increasingly depends on moving data efficiently between compute and memory, not just on adding more arithmetic capability. Designers are therefore motivated to prioritize:

• higher memory bandwidth per watt;
• HBM-based and tightly integrated logic-memory packages;
• advanced packaging and larger on-package memory pools; and
• software and model techniques that reduce memory movement and memory-capacity requirements.

This is an engineering implication of HBM becoming a capacity bottleneck, rather than a quantified market forecast. The underlying trend is clear: memory availability and bandwidth are becoming design constraints for AI systems. 317

Why capacity expansion will not create instant relief

New memory capacity takes time to become usable output. Fab construction is only the first step; manufacturers must install tools, ramp yields, qualify processes, expand advanced-packaging capability and complete customer validation. IDC identifies AI-server demand growing faster than supply can respond, while technology restrictions limit the effective contribution from Chinese producers at advanced nodes. 8

Chinese companies, including CXMT in DRAM and YMTC in NAND, are expanding and could become increasingly important suppliers. Still, reporting on their 2026–27 additions suggests much of that output may be absorbed by AI-server demand, import substitution, advanced-wafer limits and lengthy customer certification. It is not an immediate replacement for qualified leading-edge HBM supply. 411

The practical conclusion is nuanced: meaningful additions can help over time, especially in NAND and mainstream memory, but they do not quickly resolve the HBM and server-DRAM bottleneck. Until supply, yields, packaging and qualification catch up, AI infrastructure demand is likely to keep DRAM unusually tight through 2027. 78