HBM

High Bandwidth Memory (HBM) is a specialized type of DRAM that stacks multiple memory dies vertically and connects them through silicon vias, achieving far greater data transfer speeds and lower power draw than conventional memory chips. Because training and running large AI models requires moving vast amounts of data between memory and compute at extreme speed, HBM has become an essential component of the GPUs and custom accelerators powering modern AI infrastructure.

Three companies—SK Hynix, Samsung, and Micron—supply virtually all of the world's HBM, placing them at the center of the AI hardware supply chain. Demand has risen sharply as hyperscalers and chip designers push for more HBM capacity per accelerator, and all three suppliers have invested heavily in successive product generations offering higher bandwidth and capacity.

HBM's importance has made it a strategic asset in broader technology competition. Major AI chip designers work closely with HBM suppliers to qualify memory for new accelerator platforms, and memory chipmakers have emerged as some of the AI boom's biggest beneficiaries. As AI model complexity grows, HBM bandwidth and capacity remain key constraints and competitive differentiators across the industry.