SEOUL, 21 AUG 2026 — High-bandwidth memory now costs roughly US$200 a stack for HBM3, US$300 for HBM3E and about US$550 for HBM4. HBM3E contract pricing has risen around 20 per cent quarter on quarter as three suppliers allocate capacity against booked accelerator demand.
Everyone tracks the price of the chip. Almost nobody tracks the price of the memory bolted to it, and that is where the constraint actually sits.
The numbers
The HBM4 figure is an estimate for a 36-gigabyte twelve-high stack. SK Hynix is reported to have begun sampling HBM4 to its lead accelerator customer ahead of schedule.
Why memory decides what an accelerator can do
An AI accelerator’s performance is limited not by arithmetic but by how fast it can be fed data.
Running a large model means moving enormous quantities of weights between memory and compute for every token generated. If memory cannot supply data quickly enough, the processing units wait. An idle accelerator costs just as much as a busy one, which is why memory bandwidth and capacity matter more for serving than raw floating-point throughput.
High-bandwidth memory solves it by stacking memory dies vertically and placing them beside the processor on the same package, with a very wide connection between them. It is expensive, it is difficult to manufacture, and there is currently no alternative that gets close.
We wrote yesterday about Google contracting Marvell for memory interface controllers and near-memory computing. That component list is a buyer telling you where its bottleneck is, and these prices are the same statement from the supply side.
Manufacturing difficulty is what keeps the supplier count at three. Stacking a dozen thin memory dies with through-silicon vias, bonding them, and then attaching the result to a processor package without warping anything is a yield problem at every step, and a defect anywhere in the stack ruins all of it. This compounding yield risk explains why a new memory generation can be demonstrated years before it is economical to produce. The price of a new tier reflects the cost of the many stacks that are thrown away, not just the good ones.
Four hundred dollars a stack is a system-level number
The jump from HBM3E to HBM4 is roughly US$250 a stack, and a single high-end accelerator package carries several stacks.
Multiplied across a rack and then a data hall, the memory bill moves from being a component cost to a line item in the capital plan. For an operator specifying a build eighteen months out, a 20 per cent quarterly movement in contract pricing is a larger source of uncertainty than almost anything else in the bill of materials.
This also explains why accelerator vendors ration memory configurations so carefully. The difference between two variants of the same chip is often just how much HBM is attached, not how fast the logic runs.
This runs against the usual narrative of falling compute costs. Every previous generation of computing hardware got cheaper per unit of capability on a predictable curve, and a great many budgets and business plans still assume it. Memory priced on allocation rather than on cost does not behave that way, and for as long as demand exceeds what three suppliers can make, the curve for the component that determines serving capacity points upward.
Three suppliers is the whole market structure
SK Hynix at 50 to 55 per cent, with Samsung and Micron dividing the rest, is a market with no meaningful fourth option and enormous switching costs.
Qualifying a memory supplier for an accelerator package is an engineering programme, not a simple purchasing decision. It involves the packaging house, thermal design and months of validation, meaning a buyer cannot switch suppliers next quarter to chase a better price.
That is the same condition we described in Samsung raising advanced foundry prices while losing market share. Booked capacity plus high switching costs produces pricing power regardless of competitive position, and it is now visible in both logic and memory at once.
The early HBM4 sampling matters for the same reason. Being first into a lead customer's qualification cycle is worth more than being cheapest, because the qualification is what locks the generation in.
What this means for buyers in this region
Regional cloud operators, national AI programmes and universities buy accelerators rather than memory, and they still pay this bill — inside a price they cannot itemise.
For capacity planning, this means assuming the cost of compute is not falling on the schedule anyone took for granted two years ago. A programme budgeted in 2025 against expected price declines is going to find that the memory content of a modern accelerator has moved in the opposite direction.
The design response is one this desk keeps returning to. Model choice affects memory demand directly: a smaller model that fits in less memory is not merely cheaper to run, it is procurable in a market where the largest configurations are allocated to buyers with more leverage than a regional institution has. Sizing the model to the memory you can actually obtain is a more reliable strategy than waiting for the memory to get cheaper.
The other consequence lands on regional manufacturing. Malaysia and Singapore host substantial assembly, test and advanced packaging operations, and packaging is precisely where HBM is attached to logic. Tight memory supply constrains the packaging step as much as the fab, and the capacity fight is happening in facilities in this region.
What we could not establish
Whether these prices are contract or spot, and for whom. Memory pricing varies enormously between a hyperscaler with a multi-year commitment and a smaller buyer, and a single quoted figure per stack conceals a wide distribution.
Also unestablished: the source and methodology behind the per-stack estimates; how much of the HBM3E increase reflects cost rather than scarcity; the actual allocation split between the three suppliers' largest customers; HBM4 yield rates, which determine whether the price holds; whether Samsung or Micron have matched the increases; and when HBM4 reaches volume production.
What to watch
Watch whether the quarterly increase repeats. One quarter of 20 per cent is allocation; two consecutive quarters is a repricing, and it would change the cost basis of every AI deployment planned for next year.
Then watch HBM4 yields. A new memory generation at US$550 a stack is priced on scarcity as much as capability, and yield improvement is what turns that into a normal component cost — or does not.
Finally, watch whether anyone credible announces a fourth supplier or an alternative architecture. With only three suppliers and no substitute, the HBM market is the tightest structural position in the AI supply chain, yet it receives little discussion relative to its importance.
