AI accelerator designers are starting to ask for less memory per chip, and the reason is arithmetic.
TrendForce raised its 2027 high-bandwidth memory price outlook on Sept. 29, projecting a blended average selling price increase of 121% year over year. The Taipei research firm also reported that GPU and ASIC makers are evaluating a shift from 12-die HBM stacks to 8-die stacks to make the same DRAM output cover more chips.
That design concession matters because it moves the memory shortage out of the procurement department and into the product roadmap. Samsung, SK hynix, and Micron have each sold out 2027 capacity, TechPowerUp reported in August, so a buyer that missed the long-term agreement cycle has no volume to bid for and no leverage in a negotiation. Cloud providers now face a capital budget where memory is the largest single line item, and the only remaining tool is to specify less of it.
The 8-Hi tradeoff
An HBM stack is a vertical column of DRAM dies wired together with copper through-silicon vias, then bonded to the processor through a silicon interposer. CSIS analysts Sujai Shivakumar, Charles Wessner, and Thomas Howell described stacks reaching 16 layers deep in a Sept. 17 report on the shortage. Fewer dies in a stack means a cheaper bill of materials for the accelerator and more total GPUs shipped from the same wafer supply.
The savings are not as clean as they look. Every stack needs one base die regardless of height, and that fixed cost is spread across less capacity in an 8-Hi part. TrendForce estimates 8-Hi products will carry a 10 to 20% price premium per gigabit over 12-Hi in 2027.
The migration lowers the sticker price on an accelerator while raising the effective cost of every bit of memory inside it. Buyers are trading one pain for another, and the trade only makes sense if you care more about shipping units this year than about cost per gigabit in 2028.
Price momentum through the fourth quarter
Conventional DRAM contract prices still rose 10 to 15% quarter over quarter in 4Q26, according to TrendForce research published Sept. 30, with NAND flash up 15 to 20%. The pace has slowed from earlier in the year. Server DRAM contract prices had climbed a cumulative 64% in the second half of 2025, then jumped roughly 270% across 2026.
The moderation is a rate change, not a reversal. Suppliers keep steering advanced-process wafer capacity toward server parts, and TrendForce expects server DRAM to remain undersupplied through the quarter. Component supply, packaging and test capacity, and front-end DRAM flexibility all limit how fast output can follow demand.
Enterprise SSDs are the one category where growth accelerated. Bit demand for enterprise SSD is tracking above 80% year-over-year growth in 2026 as cloud providers expand AI inference infrastructure, and TrendForce calls it the only memory product line to post a bigger quarterly increase in 4Q26 than in the prior quarter. Demand is shifting from large language model training toward actual deployment, with agentic AI driving exponential growth in data volumes for real-time retrieval and caching.
Consumers absorb the difference
Phone and laptop makers have no such cushion. Notebook inventories are carrying higher component costs into the fourth quarter, and TrendForce notes that is weighing on end-market sales. PC brands are cutting SSD capacities in mainstream models to hold down bills of material, and some brands are reconsidering GDDR6 graphics memory for the same reason.
Smartphone brands are buying the minimum volumes needed to keep supplier relationships intact, the research firm said, with negotiations collapsing into price confirmation on small orders. TVs, set-top boxes, and wearables have even less room to absorb increases.
The memory makers hold the pricing power in this cycle, and they are not pretending otherwise. Anyone expecting the shortage to resolve itself through ordinary market corrections should read the CSIS report, which sees tight supply running through 2027 and into 2028 by some industry estimates. TrendForce projects that DRAM and NAND flash combined will take 47% of cloud provider capital expenditure in 2026 and 68% in 2027, which is a budget line that did not exist in that proportion three years ago.
Where the capacity actually lands
Micron is the only one of the Big Three with meaningful U.S. fabrication investment, and CSIS flags that as a policy problem. Commercial-scale HBM production stays concentrated in Asia because advanced DRAM fabrication, stacking, packaging, and customer qualification all sit in the same geography. You cannot ship a fab's output into an HBM market without the packaging and qualification ecosystem sitting next to it.
As the Big Three divert more wafers to AI memory, conventional DRAM and NAND supply tightens further and Chinese producers have room to fill the gap. CSIS warns that additional Chinese capacity could ease shortages while deepening dependence on suppliers Washington treats as a security risk. That is the policy version of the same arithmetic.
TSMC's experience shows how slowly capacity arrives in practice. CFO Wendell Huang said in July that the company is raising its Arizona commitment to $265 billion across 12 fabrication and packaging facilities plus a research center, and that its first Arizona fab is hitting yields "as good as" its flagship plant in Taiwan.
Physical constraints bind anyway. "There are physical constraints, the number of construction workers available, the infrastructure available," Huang told Reuters, adding that the company will "work closely with the government to solve these issues." Land is scarce in Taiwan too, he said, and the most leading-edge ramps have to stay home because they need R&D and operations sitting side by side.
What to watch next
TrendForce expects HBM and RDIMM together to take 51% of total DRAM bit supply in 2026, then projects a 27% rise in combined server DRAM and HBM bit supply in 2027 as new fab capacity ramps in the second half of the year. That ramp is the most plausible thing to loosen the market, and it lands in the second half of 2027 rather than early.
It's not clear the timing holds. New DRAM capacity announced today takes two to three years to become billable output, and the same cost discipline that produced this shortage also gives suppliers a reason to add less capacity than headline demand implies. A 27% supply increase against a 121% HBM price rise is not a correction, it is a softer landing.
The 8-Hi migration will show up in product announcements before it shows up in pricing. If 2027 accelerators ship with less memory per chip than the generation before them, the bill of materials worked, and the memory market never really loosened. More coverage of the supply chain is at our semiconductors desk.