Micron's entire high-bandwidth memory (HBM) output for calendar year 2026 is sold out.1 The supply constraint reflects a structural shift: AI inference workloads are now consuming memory at a pace that production cannot match.
Micron's Q3 FY2026 guided revenue of approximately $33.5B would exceed its revenue for any full prior fiscal year.1 The company has also begun shipping HBM for Nvidia's Vera Rubin platform, signaling the next generation of AI infrastructure is already ramping.1
The demand catalyst is inference. Sandisk's CEO identified inference and reasoning workloads as the fastest-growing segment of data center demand.1 Unlike training, inference runs continuously at scale — every query to a deployed model requires memory bandwidth. As model deployments multiply, so does the sustained draw on memory supply chains.
Sandisk shares have surged over 700% year-to-date in 2026.1 Micron crossed a $1 trillion market capitalization.1 Both companies trade at forward price-to-earnings multiples of approximately 10-11x despite those gains — an unusual combination that suggests earnings growth is outrunning stock price appreciation rather than the reverse.1
The supply picture is the critical variable. When memory capacity is sold out before the year begins, pricing power shifts entirely to the supplier. Elevated HBM average selling prices flow directly to gross margins, and if supply remains constrained through 2027, that margin expansion could persist across multiple fiscal years.
The key test is whether Micron announces additional HBM capacity expansions and whether hyperscaler supply contracts are renewed or expanded at current price levels. Multi-year agreements at elevated ASPs would lock in the supercycle thesis. Quarterly gross margin trends over the next four quarters will be the clearest signal.
For data center operators, the implication is straightforward: memory availability is now a deployment bottleneck, not a procurement line item. Infrastructure teams planning AI inference capacity must account for constrained HBM supply as a scheduling variable alongside compute allocation.
The memory and storage layer, long a commodity input, has become the rate-limiting factor in AI infrastructure scaling.