The three memory makers that supply high-bandwidth memory have now committed essentially all of their HBM4 output through the end of 2027, according to supply agreements described by four accelerator buyers and confirmed in earnings commentary from two of the makers this week. The constraint is no longer wafer starts. It is stacking, testing, and the number of known-good dies that survive a twelve-high stack.
What sold out actually means
Every generation of HBM sells out early. What is different this cycle is the duration and the money. Buyers are signing three-year agreements with prepayments of 15% to 30%, and at least two large customers have agreed to take yield risk — paying for stacks that fail final test above an agreed threshold. That is a transfer of manufacturing risk from the memory maker to the accelerator vendor, and it is unusual outside of shortage peaks.
The numbers behind it: a twelve-high HBM4 stack carries 36GB at roughly 2TB/s, and a top-end accelerator uses eight of them. One accelerator, then, consumes 288GB of memory that has to be stacked, bonded and tested. Multiply by every announced cluster and the demand curve stops looking like a cycle.
Why the squeeze is structural
Three things are happening at once. Capacity converted from DDR5 to HBM does not come back quickly, because the conversion is a one-way decision for a fab line. The packaging capacity — TSV, microbumping, hybrid bonding — is the tighter constraint, and it is measured in months of equipment lead time, not weeks. And qualification for HBM4 at 2TB/s takes longer than HBM3E, because signal-integrity problems show up at final test rather than at wafer probe.
The result is that even a company willing to pay cannot get volume before late 2027 unless it was early. That is what a structural shortage means: price does not clear the market, allocation does.
The bill moves to the accelerator
Memory is now 30% to 40% of the bill of materials on a high-end accelerator, depending on stack count and contract. When memory prices rise 40% and memory is a third of the cost, the accelerator rises about 13% before anyone decides to protect margin. Vendors have three options: raise prices, cut stack counts, or absorb it.
Cutting stack count is the interesting one. A part with six stacks instead of eight loses a quarter of its memory capacity and some bandwidth, which changes what fits in a single node. Teams planning million-token context workloads should assume the memory-per-dollar curve flattens in 2027 instead of continuing its historical slope.
Who gets squeezed out
Small buyers and research groups. Allocation follows volume commitments, strategic relationships and the ability to prepay. A university buying forty accelerators is not in that conversation. The realistic path for them is rented capacity, which is exactly what the new remote-access rules have made more complicated for some jurisdictions.
If you are buying
Three things, in order of impact. Price memory separately from compute in your contracts, so a memory increase does not silently reprice the whole node. Ask for the yield-risk clause in writing — if you are paying for failed stacks, you should know the threshold. And model a six-stack part as a real option; it may be the only configuration you can actually get in 2027, and for many inference workloads the capacity loss is cheaper than the wait.