TrendForce revised its 2027 HBM price outlook upward on 2026-09-29. The headline number: blended HBM average selling price is now projected to jump 121% year over year. The driver is not a demand spike — it is yields.
TrendForce states that next-generation products' yields and output "will take time to ramp, keeping memory supply tight through 2027." Suppliers have "taken a firmer stance on raising HBM prices in 2027." Three things feed the revised forecast: persistent supply constraints, a growing share of higher-priced HBM4, and the gradual ramp-up of HBM4e in the second half of 2027.
GPU Vendors Are Considering Less HBM Per Device
The practical consequence is already visible upstream. TrendForce reports that GPU and ASIC manufacturers are now considering reducing HBM capacity per device. The reason is explicit in the source: limited HBM supply and "quickly increasing system costs" — not waning end-market demand.
The mechanism: compared to earlier plans for 12-Hi HBM modules, 8-Hi configurations use fewer DRAM dies. That lowers both the cost per stack and the overall GPU bill of materials, and it lets the same HBM bit output support more GPUs. With supply tight, TrendForce expects many GPU and ASIC vendors to prioritise evaluating 8-Hi in 2027.
Why 8-Hi Costs More Per Gigabit
This is the counterintuitive part, and the one spec negotiators need to internalise.
Every HBM stack requires one base die, and that cost does not decline proportionally with the number of DRAM layers. An 8-Hi stack has less DRAM capacity over which to spread the base-die cost than a 12-Hi stack. Result: 8-Hi carries a roughly 10–20% price premium per Gb over 12-Hi in 2027.
| Configuration | DRAM dies per stack | Base-die cost spread over | 2027 cost per Gb vs 12-Hi |
|---|---|---|---|
| 12-Hi | more | more capacity | baseline |
| 8-Hi | fewer | less capacity | +10–20% (TrendForce) |
So cutting HBM capacity per device does not cut your price per gigabit. It shifts where the cost sits.
What This Means If You Buy GPU Time or Silicon
TrendForce's own caveat is the important one: adjustments to HBM capacity "may help alleviate some GPU cost pressure in 2027," but they are "unlikely to reverse the broader trend of substantial HBM price increases."
For anyone speccing HBM-backed capacity — training clusters, high-throughput inference boxes, or the unified-memory workstations I've written about, such as the AMD Threadripper Halo Station with 576GB of HBM3E — the takeaway is that the 2027 answer to "how much memory per accelerator" gets worse, not better, per gigabit. HBM4's larger share of the mix is a price headwind, not a relief, and HBM4e arriving in 2H27 adds another higher-priced tier on top.
The 8-Hi shift also changes comparison work. When NVIDIA's custom HBM work made bandwidth the constraint, the question was how many GB/s per device. In 2027 the question becomes how many bits per dollar at a given stack height — and an 8-Hi part loses that comparison against 12-Hi even though it costs less to build into a board.
The Tradeoff to Watch
The source gives no absolute dollar figures per stack, no vendor-specific 2027 price lists, and no dates beyond HBM4e in 2H27. Those are the numbers that will actually settle purchasing decisions; TrendForce's 121% and 10–20% figures are forecast-level. If your procurement cycle runs through 2027, treat stack-height negotiation as the lever — it is now a cost variable, not just a capacity one.
Sources: