Memory and GPU supply squeeze
An industry release reports AI demand is redirecting memory allocations and increasing lead times across manufacturers, while market analysis flags surging GPU prices. The combination is creating downstream cost and availability pressure for systems that rely on external memory and discrete accelerators. (globenewswire.com, investing.com)
Artificial intelligence servers are eating up the fastest memory chips and pushing up graphics processor prices, leaving other buyers facing longer waits and higher bills. (globenewswire.com) Memory is the short-term workspace a processor uses while it runs, and the newest artificial intelligence chips need stacked “high-bandwidth memory” that moves data faster than standard server memory. Samsung says its 12-layer HBM3E parts are built for artificial intelligence systems, while SK hynix says its 12-layer HBM3E reached mass production in September 2024. (semiconductor.samsung.com, news.skhynix.com) Those memory demands have climbed with each new accelerator generation. Nvidia says Blackwell Ultra supports up to 288 gigabytes of HBM3e per graphics processor, and Advanced Micro Devices says its Instinct MI355X also carries 288 gigabytes of HBM3E with 8 terabytes per second of bandwidth. (developer.nvidia.com, amd.com) Micron told investors on March 18 that it had completed price and volume agreements for its entire 2026 HBM supply, including HBM4, and raised planned fiscal 2026 capital spending to about $20 billion from $18 billion. A year earlier, Micron had already said high-bandwidth memory was tightening “non-HBM DRAM supply.” (investors.micron.com, investors.micron.com) TrendForce said on February 4 that major manufacturers had depleted inventory and were shifting capacity toward high-margin artificial intelligence products, tightening conventional dynamic random-access memory supply. The firm said spot prices were staying high and contract price growth was expected to accelerate in the first half of 2026. (trendforce.com) Graphics processors are the other pinch point. Investing.com reported in early April that graphics processor prices were surging as artificial intelligence demand kept growing, adding pressure on buyers that need discrete accelerators rather than integrated chips. (investing.com) That combination hits systems that need both parts at once: a discrete accelerator to do the math and external memory to keep models and data close at hand. Nvidia’s DGX B200 system, for example, uses eight B200 graphics processors and includes 1,440 gigabytes of graphics memory across the box. (docs.nvidia.com) The squeeze is not limited to one supplier. Samsung, SK hynix, and Micron are all racing to expand high-bandwidth memory output, but each new stack uses more advanced packaging, more layers, and tighter thermal controls than ordinary memory chips. (semiconductor.samsung.com, product.skhynix.com, investors.micron.com) For cloud providers building large artificial intelligence clusters, higher prices can be absorbed into multibillion-dollar budgets. For server makers, enterprise buyers, and smaller cloud firms, the same shortages mean tougher trade-offs on system design, delivery dates, and how much memory they can afford to attach to each machine. (globenewswire.com, trendforce.com) The near-term question is not whether demand is strong; suppliers and market trackers have already answered that. The question is how quickly new memory capacity and graphics processor output can catch up before the next wave of artificial intelligence hardware arrives. (investors.micron.com, investing.com)
