Memory Prices Aren't Coming Back: Lenovo Sees a Five-Year Structural Shift

What Happened

At ISC 2026, Lenovo drew a hard line on the memory market: DRAM and NAND prices will not return to old levels for at least five years, and even after new fabs come online around 2028, the baseline will settle far above its historical floor. The old memory cycle — prices fall, supply refills, prices recover — has broken.

The cause is a change in where supply points. Samsung, SK Hynix, and Micron are pouring capacity into AI-server high-bandwidth memory (HBM) and high-end server DRAM, starving the wafers that would otherwise make commodity DRAM and NAND. HBM eats roughly three times the wafer area of standard DDR5 for the same capacity, so every line the makers redirect toward the profitable product structurally shrinks the supply of ordinary memory. IDC now sees 2026 DRAM supply growing just 16% and NAND 17% — well below the 20–30% that defined the market since 2018. Server DDR5 contract prices are climbing 60–70% per quarter, and some DRAM has more than doubled year over year, with meaningful relief unlikely before 2028–2029. System design is shifting with it: away from “stuff in as much memory as possible” and toward GPU-accelerated compute that leans less on memory. In an era where a 16-channel server demands 1TB or more, that 1TB just got structurally expensive.

What This Means for Founders

Memory has moved from a variable line item to a fixed structural cost. Hardware startups feel it first. Any product with DRAM or NAND in its bill of materials — devices, edge boxes, robots — sees margins permanently compressed. The bet “prices will dip next quarter, just hold on” is now dangerous; rebuild your cost model with no dip as the default.

It hits AI infrastructure too. Memory’s share of inference and training cost rises, so architectures designed to burn memory carry a structurally heavier cloud bill. The teams that do the same work with less memory win on cost. Model quantization, streaming and on-demand loading, and memory-efficient serving move from nice-to-have to survival condition. Lenovo’s “shift to GPU-accelerated compute” is, in plain terms, a signal to trade memory-bound design for compute.

There is a sourcing angle, too. Hyperscalers are locking long-term supply contracts to guarantee their memory, which pushes smaller buyers to the back of the line at the worst prices. If your product ships physical units, securing memory on annual contracts rather than quarter-by-quarter is no longer a procurement nicety — it is a margin decision.

What You Can Do Now

Find the memory line in your cost model and rerun your scenarios at 1.5–2x today’s price. Confirm the margin still survives before anything else. If you ship an AI product, make “can this feature run on compute instead of memory?” a standing question in architecture review. If high memory prices are structural, a design that runs cheap on top of that structure is the moat.