HBM memory, the critical memory component for proper AI development and application, has experienced an existential demand crisis in 2025. Analysts estimate the peak of the crisis won’t be seen until the end of next year at the earliest, and will run through late 2027 at least.
HBM supplies are tight because each AI GPU needs multiple HBM stacks, and many GPUs are already ordered on backlog. HBM is also more silicon-intensive than most chips: it’s built by stacking dies, which reduces how many finished units you can get from a single wafer. On top of that, it requires advanced manufacturing and complex packaging, so expanding output takes time.
