The global memory market has flipped from glut to scarcity in less than two years, and the warnings are getting blunter. The chief executive of Taiwanese memory maker Team Group said this week that the shortage heading into 2027 could become severe enough that “even money can’t buy goods.”
That is a supplier talking its book, but the underlying signal is corroborated across the industry. High-bandwidth memory attached to AI accelerators is sold out well into next year. Conventional DRAM used in servers, phones and PCs is being reprioritized toward data center customers.
How a commodity became strategic
Memory spent a decade as a cyclical commodity, punished by overbuilding and rewarded by consolidation. AI changed the demand curve. Each accelerator generation attaches more HBM stacks, and each stack consumes wafer capacity that would otherwise produce standard DRAM.
The result is a market where allocation agreements signed a year in advance determine production plans. Server makers are designing around memory they are contractually guaranteed rather than memory they would prefer.
Knock-on effects are already visible
AMD has introduced memory-management technology that tiers DRAM and NAND flash and predicts application data access patterns, moving data from flash to DRAM before the CPU requests it. The pitch is efficiency. The subtext is that DRAM has become expensive enough to engineer around.
Consumer hardware will feel it next. Device makers typically lock component pricing six to nine months ahead of launch, which means 2027 product lines are being costed against today’s spot market.
Where it ends
Memory capacity responds to capital, not to demand. New fabs take two to three years from announcement to volume output, and the current expansion plans were drawn up after prices had already started climbing. The realistic read is that tightness persists through 2027, with relief dependent on how quickly those lines come online and how much of the new output is claimed by AI accelerators before it reaches anyone else.