AI buyers seek long DRAM deals
Big AI customers are reportedly negotiating five‑year DRAM supply deals with Samsung and SK Hynix as memory shortages worsen, a move meant to secure capacity for large training and inference workloads. If firms lock long contracts, smaller cloud providers and enterprise buyers could be pushed further down the supply queue, widening infrastructure scarcity beyond GPUs. (en.bloomingbit.io)
Microsoft and Google are negotiating multi-year dynamic random-access memory supply deals with Samsung Electronics and SK hynix, extending contracts that usually run a year or less to as long as five years. (koreajoongangdaily.joins.com) Samsung co-Chief Executive Jun Young-hyun told shareholders on March 18 that the company was working with major customers to shift memory contracts to three to five years from quarterly or annual terms. Reuters reported the change is tied to expectations of strong artificial intelligence data-center demand in 2026. (usnews.com) South Korea’s Korea JoongAng Daily reported on April 14 that Microsoft and Google are the buyers in the current talks, and that no memory supply contracts had previously stretched beyond one year. The same report said both Samsung and SK hynix had signaled the shift at shareholder meetings last month. (koreajoongangdaily.joins.com) Dynamic random-access memory is the working memory that keeps data close to processors while servers train and run artificial intelligence models. TrendForce said on March 31 that cloud service providers were securing supply through long-term agreements as artificial-intelligence server demand pushed memory contract prices higher for the second quarter of 2026. (trendforce.com) The squeeze is not limited to the most advanced chips. TrendForce said suppliers were reallocating advanced manufacturing capacity toward server memory and high-bandwidth memory, tightening supply elsewhere and driving conventional dynamic random-access memory contract prices up about 55% to 60% quarter over quarter in the first quarter of 2026. (prnewswire.com) High-bandwidth memory is a stacked form of dynamic random-access memory that sits next to artificial intelligence accelerators and feeds them data faster than standard server memory. Samsung and SK hynix have been expanding output for that market, with Data Center Dynamics reporting in January that Samsung planned to raise high-bandwidth memory capacity by about 50% in 2026 while SK hynix was increasing infrastructure investment. (datacenterdynamics.com) SK hynix has also been building cash for that expansion. At its March 25 annual meeting, the company said it aimed to raise net cash from 12.7 trillion won to more than 100 trillion won over time to fund infrastructure including the Yongin semiconductor cluster. (news.skhynix.com) The broader market is already moving toward locked-in allocations. Micron said in late February that its entire 2026 high-bandwidth memory output was already sold out, showing how far ahead large artificial intelligence buyers are reserving supply. (finance.yahoo.com) If Samsung and SK hynix finish these longer deals in the first half of 2026, the memory market will look more like the graphics processing unit market that artificial intelligence customers have been navigating for the past two years: capacity spoken for early, and smaller buyers left to chase what is left. (trendforce.com)
