HBM is sold out through the end of 2026 at all three memory makers

Try to place an order for high-bandwidth memory with delivery inside 2026. SK Hynix will tell you its capacity is committed. Samsung will tell you the same thing, and Micron's HBM output is described, flatly and specifically, as fully sold out through the end of the year. That is the picture across TrendForce and the industry reporting that has accumulated through 2026, and it covers every manufacturer of the component sitting at the center of the AI buildout.

There is no fourth supplier.

A quick word on the part itself, for readers who live in land use rather than silicon. HBM is the stacked memory bonded directly beside the processor inside an AI accelerator, and it exists because modern AI workloads are starved for memory bandwidth long before they run out of raw arithmetic. There is no substitute component. A cluster design that specifies HBM cannot value-engineer its way to something cheaper, which is why the order books at all three makers filled the way they did, and why the buyers holding those commitments treat the allocation itself as the asset.

Sold out is an allocation regime, not a shortage

In a normal supply chain, sold out describes a bad quarter. Demand ran ahead of forecast, capacity catches up, an expediting fee moves you up the list, and the whole thing reads as a hiccup in the next annual report. Every procurement team on earth carries this model in its head, because for most components, most of the time, it is true.

This is not that.

When every maker of a component is committed through the end of the calendar year, and the commitments run to a small number of enormous buyers who negotiated them long in advance, the word shortage stops being useful. What you are looking at is an allocation regime. The output is spoken for before it exists. A new buyer does not join a queue that is moving; a new buyer negotiates for a place in a future allocation round, against incumbents with standing relationships and larger checks.

Here is the detail that tells you the regime is structural. Data Center Dynamics has reported Samsung and SK Hynix moving to scale up memory production capacity in 2026 specifically to meet AI demand. Set that against the sold-out order books and the picture sharpens considerably: the industry is adding capacity as fast as it knows how, and the capacity is absorbed before it comes online. Expansion is not relieving the queue. The queue is eating the expansion.

There is no spot market for the scarcest component in the buildout.

The wafer math says the queue cannot clear quickly

The reason this does not resolve in a quarter or two sits in arithmetic this lane has covered in detail. One gigabyte of HBM consumes roughly four times the wafer capacity of one gigabyte of conventional DRAM, total DRAM capacity grows only 10 to 15 percent a year per TrendForce, and AI is on track to claim about 20 percent of global DRAM wafer capacity in 2026. Every increment of HBM supply is bought with four increments of conventional output, which is why conventional DRAM contract prices rose 93 to 98 percent quarter over quarter in 1Q26 while the HBM lines ran full.

The escape hatch is small and the crowd is large. That is the whole condition, and no expediting fee changes it.

Hardware has joined power on the critical path

Now the field-note part, which is what this means if you build or underwrite the real estate.

For two decades, compute hardware was the one input a data-center project never worried about. Power was the long-lead item and entitlement was the fight; the servers were an afterthought you procured once the shell had a completion date. Issue the purchase order, take delivery in weeks. That assumption is dead for this class of hardware. Be precise about what replaced it. Memory, and therefore the accelerators built around it, now behaves the way interconnection behaves: a committed multi-year allocation negotiated early, unavailable at any price to whoever assumed it would be there on demand.

Anyone modeling a project on the premise that the tenant can simply order more compute when the building is ready has the sequencing backwards. Increasingly, the hardware allocation has to exist before the building does, the same way the megawatts do.

For site selection this stacks a third multi-year queue on top of the two the industry already knows how to fight about, interconnection and entitlement, and the three queues do not move on the same calendar or answer to the same people. A tenant holding a locked 2027 hardware allocation with no powered, permitted shell to receive it is the new expensive mistake. It is the mirror image of the old one, and I suspect the industry will need to make it a few times in public before the sequencing lesson takes.

My bet is that the window moves out rather than in. When allocation reporting for 2027 arrives, a substantial share of next year's HBM output will already be committed before January does, because the buyers who spent 2026 locked out will not risk a second year in the hallway. If instead a genuine spot market in HBM reappears during 2027, with real volume available inside a quarter, then demand is cooling faster than any current data suggests, and this lane's entire premise deserves the rewrite I would owe it.

This analysis is a source-cited research summary drawn from public reporting and market research, not investment advice. It can contain errors and should be verified independently before any investment decision.