Memory prices surge amid supply squeeze
Memory suppliers report persistent shortages that have driven average selling prices sharply higher—some module‑makers warn ASPs could rise several‑fold. The supply tightness is occurring even as consumer demand softens, meaning AI and datacentre buyers are competing for constrained inventory. That makes memory a direct cost and delivery risk for any team planning large inference or training purchases. (digitimes.com)
A memory chip is the short-term workspace inside a server, like the kitchen counter where ingredients sit while the meal is being cooked. In April 2026, that counter got much more expensive: TrendForce said conventional dynamic random-access memory contract prices are set to jump 58% to 63% in the second quarter, and flash storage prices 70% to 75%. (trendforce.com) The surprise is that this is happening while parts of the consumer electronics market are still weak. TrendForce said supply is tight anyway because suppliers are moving production toward high-bandwidth memory and server products, where artificial intelligence buyers will pay more. (trendforce.com) That shift starts with three companies: Samsung, SK hynix, and Micron control most of the global dynamic random-access memory market. When those companies steer wafer capacity toward premium server chips, fewer ordinary chips are left for personal computers, phones, and memory modules. (trendforce.com) (news.skhynix.com) (news.samsung.com) High-bandwidth memory is a special kind of memory built for artificial intelligence accelerators, and it works by stacking chips vertically so data travels a shorter distance. Micron said in March 2026 that its high-bandwidth memory output for calendar 2026 was already sold out, which tells you how hard the biggest cloud buyers are pulling on the supply chain. (fool.com) The shortage is not only in the exotic products used next to graphics processors. TrendForce said suppliers’ dynamic random-access memory inventories are at floor levels, and hyperscale cloud companies are still locking in supply with long-term agreements because they expect artificial intelligence server demand to stay high. (trendforce.com 1) (trendforce.com 2) That is why even older memory suddenly behaves like scarce real estate. DigiTimes reported that spot prices for 16-gigabit double data rate 4 chips had surged more than 2,200% over the prior year before finally easing in early April, which is an extraordinary move for a part that many people treat like a commodity. (digitimes.com 1) (digitimes.com 2) The companies assembling memory modules are already showing the effect in revenue. DigiTimes reported that Transcend posted March 2026 revenue of 5.7 billion New Taiwan dollars, up 78.6% from February and 3.8 times higher than a year earlier, after benefiting from the jump in memory prices. (digitimes.com) For anyone buying artificial intelligence servers, memory is no longer a rounding error next to the graphics processor. TrendForce said suppliers are prioritizing server dynamic random-access memory and enterprise solid-state drives, so a training cluster or inference build can now be delayed by the parts that sit around the processor rather than the processor itself. (trendforce.com) The awkward part is that more memory is not optional in newer systems. TrendForce said demand for high-capacity registered dual in-line memory modules in servers is one of the main reasons prices are rising, because larger models and inference loads need more data close to the chip at all times. (trendforce.com) So the 2026 bottleneck is not just “can you get enough graphics processors.” It is also “can you get enough dynamic random-access memory, high-bandwidth memory, and enterprise flash at the same time,” because the artificial intelligence buildout is turning memory from a cheap component into a schedule risk. (trendforce.com) (fool.com)
