Memory squeeze as AI scales
Reports indicate that aggressive infrastructure scale‑ups by frontier AI labs and specialist clouds are tightening global DRAM availability, creating supply pressure that can ripple into system BOMs and lead times. The coverage links recent AI capacity builds to sellouts and memory shortages in consumer and pro device lines (techtimes.com).
The memory chips that feed computers and servers are getting tighter and pricier as artificial intelligence builders soak up more supply. (trendforce.com) Dynamic random-access memory, the short-term working memory inside a machine, is now under pressure from two directions: bigger cloud server orders and a shift toward high-bandwidth memory, the stacked version used beside artificial intelligence accelerators. TrendForce said on March 31 that server demand is still pushing prices higher in the second quarter of 2026, while graphics memory supply remains constrained. (trendforce.com) TrendForce also said earlier that cloud service providers expanded data-center purchases enough to turn a third-quarter 2025 balance into a broader shortage, and that server dynamic random-access memory demand is expected to grow by more than 20 percent in 2026. Its October outlook said conventional dynamic random-access memory prices in late 2025 were revised up to an 18 percent to 23 percent rise. (trendforce.com.tw) Counterpoint Research said in January that memory prices jumped 40 percent to 50 percent in the fourth quarter of 2025 and were set to rise another 40 percent to 50 percent in the first quarter of 2026, with about 20 percent more in the second quarter. The firm tied that pricing power to “insatiable” artificial intelligence and server demand. (counterpointresearch.com) That squeeze reaches beyond data centers because the same manufacturers make both premium artificial intelligence memory and the more ordinary chips used in servers, personal computers, and other devices. S and P Global Market Intelligence said in January that high-bandwidth memory had become one of the most lucrative corners of semiconductors, drawing production away from standard dynamic random-access memory. (spglobal.com) IEEE Spectrum reported in February that Micron said cloud-related and high-bandwidth memory products rose from 17 percent of its dynamic random-access memory revenue in 2023 to nearly 50 percent in 2025. The same report said Counterpoint had tracked dynamic random-access memory prices up 80 percent to 90 percent in that quarter. (spectrum.ieee.org) Suppliers are signaling how booked up the market has become. Samsung said in February that it expects high-bandwidth memory sales to more than triple in 2026 from 2025 and is expanding high-bandwidth memory 4 production capacity. (news.samsung.com) SK hynix told investors in results published January 29 that 2025 revenue hit 97.15 trillion won and operating profit reached 47.21 trillion won, both records, after a year driven by artificial intelligence memory demand. CNBC reported after SK hynix’s October 2025 earnings that the company had already sold out chips for 2026. (skhynix.com; cnbc.com) The consumer side is starting to show visible strain. MacRumors and 9to5Mac reported on April 11 that several higher-memory Mac mini and Mac Studio configurations were marked “currently unavailable” on Apple’s United States online store, though those reports also said an M5 refresh could be another explanation. (macrumors.com; 9to5mac.com) Apple has not publicly tied those stockouts to memory supply, and recent coverage has framed the missing configurations as a mix of supply constraints and possible product-transition timing. What is clear from supplier and market reports is that artificial intelligence infrastructure spending is now strong enough to shape memory prices, lead times, and product availability well outside the data center. (thenextweb.com; trendforce.com)
