DRAM Prices Climb as AI Triggers a Memory Supercycle

Memory is no longer a sideshow in the AI trade. Contract prices for
DRAM, NAND flash and high-bandwidth memory (HBM) have all moved sharply higher
in 2026, and analysts at Morgan Stanley
now describe the imbalance as a structural shortage rather than a passing
cyclical squeeze. The setup is reshaping winners and losers across the
semiconductor supply chain — and beginning to bite the PC, smartphone and
server brands that rely on cheap memory to protect their margins.

What changed: HBM ate DRAM capacity

The mechanism behind the move is simple. High-bandwidth memory — the stacked
DRAM that sits next to NVIDIA’s H200, B200 and AMD’s MI350 accelerators — uses
roughly three times the wafer area of a standard DDR5 die for the same bit
output, according to disclosures by
SK hynix and
Micron. As the big
three memory makers — Samsung Electronics, SK hynix and Micron Technology
(MU)
— have converted leading-edge fabs to HBM3E and HBM4 production to satisfy
hyperscaler order books, the wafer pool left for conventional DDR5 and LPDDR5
has shrunk.

The result, tracked by TrendForce‘s
DRAMeXchange index, is a contract market in which buyers compete for a smaller
non-HBM supply just as AI server bills-of-materials lift memory content per
box. Server DDR5 modules now carry several times the per-GB price they did at
the 2023 cycle trough, and HBM3E stacks remain effectively sold out into 2027.

The price tape

Micron framed the dynamic on its most recent earnings call, telling
investors that DRAM bit demand growth driven by AI servers is outpacing
industry bit supply growth and that data-center pricing has been rising for
consecutive quarters. The company’s filings are available via
SEC EDGAR.

Who’s winning the cycle

The pure-play memory makers are the cleanest beneficiaries. Micron has
re-rated as gross margin guidance moved higher and as the company secured
multi-year HBM supply agreements with key accelerator customers. SK hynix —
the HBM market leader by share — has guided to record annual revenue, and
Samsung’s memory division has flagged improving profitability after a year of
qualifying its own HBM3E with NVIDIA.

Equipment suppliers are downstream beneficiaries. Memory capex has been
revised upward at all three IDMs as they prioritize advanced packaging
(TSV-stacked HBM) and 1c/1d node DRAM. That spending flows to
Applied Materials,
Lam Research
and KLA,
plus packaging specialists like
Besi for hybrid bonding tools.

Who pays the bill

Anyone who buys memory and resells it inside finished hardware is on the
wrong side of this curve. That includes the PC OEMs —
HP Inc.,
Dell,
Lenovo —
the smartphone brands led by
Apple and the Android pack,
and the cloud hyperscalers themselves, who absorb HBM costs inside their AI
server bills-of-materials. Several PC vendors have already telegraphed
component-cost pressure in recent guidance updates, and channel research
houses such as Canalys have flagged consumer notebook price hikes
linked to DRAM.

Even NVIDIA is not insulated. HBM is one of the most expensive line items
in an H200 or B200 module, and any pass-through of HBM scarcity into the
accelerator price stack is now a recurring topic on hyperscaler earnings
calls.

Why Morgan Stanley calls it structural

The structural label rests on three claims, summarized from the bank’s
research and corroborated by company commentary:

• HBM has a multi-year demand runway. Hyperscaler AI capex
  guidance — disclosed in
  Amazon,
  Microsoft,
  Alphabet
  and Meta
  10-Qs — points to triple-digit AI infrastructure budgets in 2026 and 2027.
• Capacity to expand HBM is bottlenecked. Adding TSV
  packaging and high-yield 1b/1c DRAM lines is slower than the spot market
  expects, and there are limited number of suppliers with the IP and yield to
  ship qualified HBM3E and HBM4.
• Non-HBM supply is shrinking by choice. Each percentage
  point of wafers diverted to HBM is a percentage point removed from the DDR5
  and LPDDR5 pool, tightening pricing for the rest of the market.

Risks to the bull case

Two scenarios would interrupt the cycle. The first is an AI capex pause —
either from a hyperscaler digestion period or from regulatory pressure on
specific accelerator exports. The second is a yield breakthrough at Samsung
that floods HBM3E supply faster than demand grows. Memory is also notoriously
cyclical, and the same pricing power that lifts the IDMs today eventually
pulls new capacity into the market.

For now, the price tape is telling a different story. With HBM allocated,
DDR5 contract prices stepping higher each quarter and the memory makers
guiding margins up rather than down, the burden of proof has shifted to the
bears.

Sources

• TrendForce memory price coverage
• Micron Technology SEC filings and earnings releases
• SK hynix investor relations
• SEC EDGAR – Micron Technology
• Morgan Stanley research ideas hub

Disclosure: This article was produced with AI assistance and reviewed before publication. It is for informational purposes only and is not investment advice.