Direct:
The memory shortage has hardened from a 2026 price spike into a multi-year structural problem. The buyers driving it are the hyperscalers, signing two-year supply contracts to feed Nvidia (NVDA) hardware they have not even installed yet.
The mechanism is long-term agreements. Industry sourcing has nearly all of 2026 memory capacity pre-booked, with cloud providers signing two-year LTAs that stretch commitments into 2028. Normally a price move this large pulls in fresh capacity and supply catches up within a few quarters. The LTAs short-circuit that, because much of the future output is committed before the fabs that make it are fully ramped.
The pricing shows how tight it already is. TrendForce put 1Q26 conventional DRAM contract prices up 90 to 95% quarter over quarter, with NAND up 55 to 60%. Its 2Q26 projection called for another 58 to 63% on DRAM and 70 to 75% on NAND. DRAM inventories had already fallen roughly 80% year over year, to something like three to eight weeks of supply by late 2025. Inventory that thin leaves no buffer for a demand surprise.
HBM is the biggest single driver on the DRAM side. Samsung (005930.KS), SK Hynix (000660.KS), and Micron (MU) have sold out HBM through 2026, and HBM now absorbs around 23% of total DRAM wafer output. It consumes far more wafer area and die per usable bit than commodity DRAM, so capacity routed to stacked memory for AI accelerators pulls directly from conventional DDR5 and LPDDR. Micron has called the shortage unprecedented and put it past 2026. NAND is tight for a related but separate reason: years of underinvestment colliding with enterprise SSD demand for AI storage.
Consumer SSD and DIMM buyers feel this downstream. Conventional DRAM and NAND sit behind AI memory and enterprise contracts in the priority queue, and that ordering holds as long as AI demand and the LTAs do. Expect elevated DIMM and SSD pricing to stick around rather than mean-revert on the old cadence.
The honest hedge is that contract forecasts are not deliveries. If AI capex disappoints or new fabs ramp faster than expected, the back half of this softens. Current reporting leans the other way.
Drafted with AI assistance against parallel reporting.
Sources:
- DigiTimes, “Nvidia's AI ramp deepens memory squeeze as cloud providers lock up supply through 2028” (June 10, 2026)
- DigiTimes, “Exclusive: CSPs lock in memory capacity with two-year LTAs through 2028” (November 28, 2025)
- TrendForce, 1Q26 memory price outlook, DRAM +90-95% / NAND +55-60% QoQ (February 2, 2026)
- TrendForce, 2Q26 memory contract price projection, DRAM +58-63% / NAND +70-75% QoQ (March 31, 2026)
- Micron / industry coverage, “unprecedented” memory shortage and HBM sold out through 2026 (late 2025 to 2026)
Reddit: https://ift.tt/znKCeRp