NAND prices up 600% as HBM crowds out AI memory supply
Key insights
- NAND contract prices surged over 600% since September 2025, the most severe memory shortage in 15 years per TrendForce.
- Hyperscalers redirected chipmaker cleanroom capacity to HBM production, directly causing 4.9% DRAM and 4.2% NAND global supply deficits.
- No new fab capacity is expected before late 2027, meaning elevated memory costs will persist as a structural AI infrastructure constraint.
Why this matters
AI infrastructure buildouts are now facing a cost floor baked in by the same hyperscaler capex decisions that enabled HBM-fed GPU clusters, meaning teams budgeting for storage and server memory in 2026 and 2027 are operating on fundamentally different unit economics than 18 months ago. Founders building data-intensive AI products face margin compression that cannot be solved by switching vendors, because the shortage is global and supply-side relief is years away. Technical leaders evaluating on-premise versus cloud tradeoffs need to reprice storage-heavy workloads upward and reassess architectural choices that assume cheap NAND, such as large local vector stores or edge inference with flash-heavy caching.
Summary
Memory markets are in the worst shortage in 15 years, and the culprit is a deliberate capacity shift by the world's largest chipmakers. NAND Flash contract prices have risen more than 600% since September 2025, with conventional DRAM up roughly 400%, according to TrendForce data. The mechanism is straightforward: hyperscalers pushed Samsung, SK Hynix, and Micron to convert cleanroom capacity toward High Bandwidth Memory, which carries higher margins and feeds AI accelerator demand. That conversion pulled supply away from standard NAND and DRAM, leaving a 4.9% DRAM deficit and 4.2% NAND deficit globally with no new fab capacity arriving before late 2027.
Essentially: (Samsung, SK Hynix, Micron) optimized for HBM revenue and created a structural shortage in every other memory segment.
- NAND contract prices are up over 600% since September 2025, the steepest rise in roughly 15 years.
- The global DRAM supply deficit stands at 4.9%, NAND at 4.2%, both driven by cleanroom repurposing rather than demand destruction.
- New fab capacity capable of relieving the shortage is not expected online before late 2027, locking in elevated costs for at least 18 months.
For AI infrastructure builders, this means server and storage bills are a structurally higher cost floor, not a temporary spike to wait out.
Potential risks and opportunities
Risks
- Cloud infrastructure startups and mid-market AI companies without long-term memory supply contracts face spot-price exposure that could compress gross margins by double digits through 2027.
- AI hardware OEMs (Supermicro, Wiwynn, Quanta) building HBM-adjacent server lines may face customer pushback or contract renegotiation as NAND and DRAM costs inflate total system prices beyond original bids.
- Enterprises that locked in multi-year AI infrastructure expansion plans at 2024 memory price assumptions may need to revise capex projections upward significantly, potentially delaying or canceling planned GPU cluster expansions before new fab supply arrives in late 2027.
Opportunities
- Memory-efficient AI architecture vendors and model compression tooling providers (Neural Magic, Qualcomm AI Research) gain a stronger procurement-side argument as high memory costs make parameter-efficient models a financial priority.
- Storage optimization and tiered-memory software companies (Vast Data, WekaIO, Lightbits Labs) are positioned to capture budget from enterprises trying to squeeze more utility from expensive NAND allocations.
- Investors and acquirers focused on fab diversification plays gain leverage as the shortage validates the case for non-Samsung, non-Hynix memory supply chains, benefiting emerging NAND producers and DRAM foundry entrants seeking capital.
What we don't know yet
- Which specific hyperscalers (AWS, Google, Microsoft, Meta) drove the largest cleanroom conversion commitments, and when those contracts were signed relative to the September 2025 price inflection.
- Whether any major memory producer has announced plans to reverse HBM capacity allocation toward standard DRAM or NAND ahead of the 2027 fab timeline.
- How spot versus contract price divergence is affecting smaller cloud providers and ODMs who lack long-term supply agreements with Samsung, SK Hynix, or Micron.
Originally reported by techspot.com
Read the original article →Original headline: NAND Contract Prices Surge Over 600% Since September 2025, DRAM Up ~400% as AI Server Demand Overwhelms Memory Supply