AMD's Lisa Su flags HBM shortage

AMD CEO Lisa Su said on May 24 that high-bandwidth memory, or HBM, has become the next major bottleneck in AI chip supply, shifting attention from GPUs alone to the memory stacks that sit beside them. TechTimes reported Su’s remarks as memory prices continue to rise and suppliers lock in output far in advance. Micron has already sold its calendar 2026 HBM supply under price and volume agreements, according to the company’s investor materials. TrendForce said conventional DRAM contract prices rose 90% to 95% quarter-on-quarter in the first quarter of 2026. ### Why is HBM suddenly at the center of the AI hardware squeeze? HBM is a specialized form of memory used in AI accelerators from companies including AMD and Nvidia, where bandwidth and power efficiency matter as much as raw compute. TechTimes reported Su said HBM, rather than advanced packaging, is now the binding constraint on AI chip supply. AI servers use HBM because large models need fast access to data during training and inference. That makes HBM a required component, not an optional upgrade, for many high-end accelerator systems. When HBM supply tightens, chipmakers can end up constrained even if they can secure wafer capacity, packaging, or GPU demand. (techtimes.com) ### What numbers show how tight the market has become? TrendForce said on February 2 that conventional DRAM contract prices for the first quarter of 2026 were revised up to a 90% to 95% quarter-on-quarter increase, citing persistent AI and data-center demand and broad supply gaps across buyers. The firm also said PC DRAM prices were expected to at least double quarter-on-quarter in the same period. (techtimes.com) Micron said in investor presentation materials that it had completed agreements on price and volume for its entire calendar 2026 HBM supply, including HBM4. The company also forecast the total addressable market for HBM would grow to about $100 billion in 2028 from about $35 billion in 2025. ### What does Micron’s sold-out 2026 supply actually tell buyers? (trendforce.com) Micron’s fully allocated 2026 HBM output shows that large customers are reserving supply well before chips ship. That matters because HBM is made by a small group of suppliers and requires advanced manufacturing and packaging steps that are difficult to expand quickly. The Micron disclosure also indicates pricing is being negotiated through long-term agreements rather than spot purchases for much of the market. (investors.micron.com) That can make supply harder to access for smaller buyers and for companies without multiyear AI infrastructure plans, according to the structure of the company’s announced agreements. This is an inference based on Micron’s statement that its 2026 supply is fully covered by price and volume agreements. ### Why can’t the industry add relief quickly? HBM capacity cannot be expanded with a single change on a production line. New output depends on wafer capacity, advanced packaging, yield improvement and equipment additions across a concentrated supplier base. TechTimes reported no meaningful new HBM fab capacity is expected before late 2027. (investors.micron.com) Other industry reports have described a similar timeline, though the exact start of relief varies by source. Several reports say meaningful incremental HBM supply is more likely in late 2027 or later, after current expansion projects begin producing in volume. ### Who is most exposed as the shortage deepens? AMD, Nvidia and hyperscale cloud customers are exposed because their AI accelerators depend on HBM availability. (techtimes.com) Memory suppliers including Micron, Samsung Electronics and SK hynix are central because they control most of the supply base for advanced memory used in these systems. TechTimes framed Su’s comment as a sign that the supply constraint has moved deeper into the component stack. (storageswiss.com) The next concrete markers are likely to come from Micron disclosures on HBM bookings, TrendForce pricing updates and supplier capacity announcements through late 2027. Micron’s investor materials already show 2026 HBM is fully allocated, and TrendForce has continued to publish quarterly memory pricing revisions tied to AI demand. (investors.micron.com) (techtimes.com)