The savings frenzy in memory and storage has come to an end. What started as a rebound following the 2022-2023 crash has turned into a broad and persistent shortage impacting DRAM, NAND flash, and HDDs simultaneously, something unusual in the last decade. The underlying reason is clear: AI data centers are consuming global capacity at a pace that manufacturers and foundries cannot — or will not — match with the same speed, after years of traumatic cycles of overinvestment and plummeting prices.
The result is a bottleneck affecting all segments: from DDR4 modules and consumer SSDs to all-flash enterprise arrays and high-capacity nearline HDDs. hyperscalers (large cloud providers) are securing the priority supply with multi-year contracts, while the rest of the market competes over remaining stock… at rising prices.
From oversupply to rationing: how the market turned around
In 2022 and early 2023, both NAND and DRAM were selling below cost. Manufacturers responded with drastic production cuts, which emptied inventories and — by the second half of 2023 — shot up spot and contract prices. In NAND, for example, 512 Gb TLC doubled in price within just six months, leading to visible effects on retail shelves: a 2 TB SSD that was around $120 at Christmas jumped to $175–180 in weeks.
In DRAM, the shift came with a delay of one quarter, but the pattern was the same: PC DDR4, which seemed clearance stock in 2023, faced dwindling supply as lines started closing. By Q3 2025, some analysts projected +38–43% quarterly growth in PC DDR4 and +28–33% in servers. GDDR prices also compressed: with the upcoming GDDR7 for next-gen GPUs, GDDR6 surged by about 30%. DDR5 has climbed more slowly, but with an upward trend.
HDDs — traditionally an outlet when flash prices rise — have not escaped: manufacturers announced 5–10% increases and, according to tracking firms, high-capacity nearline drives are facing shortages and lead times exceeding a year for certain capacities. When nearline is lacking, some warm data workloads have migrated to flash (QLC), further tightening NAND.
Main reason: the “diet” of AI data centers
Each memory cycle has its trigger. There was a time when smartphones led demand; then, laptop SSDs; and later, cloud. Now it’s AI. Training language models and serving inferences at scale requires hundreds of GB of DRAM and several TB of flash per GPU node, and thousands of nodes per cluster. The appetite for memory and storage is structural, multiplying with every new wave of GPUs and model sizes.
Impact-wise: hyperscalers and AI labs block supply for years ahead. Nodes of HBM planned for 2026 are nearly 100% pre-sold; next-generation V-NAND lines are committed before launch; contracts that once covered just one quarter now span many years. What doesn’t go into HBM is allocated to NAND QLC for enterprise arrays, leaving less investment for standard DRAM and consumer TLC.
The cascading effect reaches consumers: for instance, Raspberry Pi announced in October 2025 a $5–10 price hike on 4 GB/8 GB variants due to memory cost increases, admitting that chip costs are about 120% above last year. It’s the visible tip of a broad rise affecting even the maker community.
Why aren’t manufacturers building more fabs?
They are, but not quickly or cheaply. A memory fab greenfield costs tens of billions and takes years to reach volume. Even expanding existing lines involves months of installation and tool calibration, in a context where ASML, Applied Materials, and other suppliers already have full order books and bottlenecks.
Moreover, manufacturers have learned that overinvestment during demand peaks typically ends in price collapses (2019, 2022). They prefer discipline: sell less at higher margins rather than crack open the taps and cycle back into ruin. Currently, the priority is where margin is: HBM and advanced nodes. Every wafer dedicated to HBM is one less for DRAM; every engineer focused on enterprise QLC leaves fewer resources for consumer TLC.
Add to this geopolitical factors: export controls on lithography and rare earths (like neodymium magnets for HDDs), cross-restrictions among major powers, and talent shortages. The puzzle isn’t solved with a simple “let’s build more.”
Category outlook
DRAM (DDR4, DDR5, HBM)
- DDR4: demand retreating faster than supply drops. Shortages in PCs and servers; quarterly increases of about 30–40%, typically referenced in the market.
- DDR5: rising more gradually, with an upward trend as it becomes mainstream.
- HBM: fully booked across many manufacturers through 2026; main focus of capex.
NAND (TLC, QLC, Enterprise)
- Consumer TLC: rationing; a strong rebound from 2023 lows.
- Enterprise QLC: deployment acceleration to ease nearline demand → extra pressure on supply.
- Next-gen V-NAND: pre-allocated to enterprise/hyperscale clients.
HDDs (nearline and bulk)
- Nearline: lead times expected to exceed 12 months for capacities over 20 TB; 5–10% increases already communicated by manufacturers.
- Substitution with QLC in some workloads: easing one bottleneck… but creating another in NAND.
How long could this last?
The more cautious estimate suggests “several years” of tension; some claim it could stretch to a decade of supply tightness. The truth probably lies somewhere in the middle and depends on three factors:
- The actual cadence of AI: if demand remains exponentially high (larger training runs, inference everywhere), absorption will stay high. If there are pauses after stockpiling or improvements in model/infra efficiency (parameter economy, sparsity, in-memory compute), the pace might moderate.
- Production rollout: how much new volume HBM/DRAM and NAND add in 2026–2027 and whether it arrives on time.
- Geopolitics and supply chain: equipment, materials, and magnets for HDDs; subsidies and licenses; labor.
In any case, 2025–2026 looks like a period of high prices and prioritized enterprise clients. Consumers are at the tail end: fewer offerings and more constrained capacities.
Practical implications (by profile)
Consumers and PC builders
- SSDs: the “bargains” on 1–2 TB drives at throwaway prices won’t return soon. If you need to upgrade, buy when a good deal appears.
- RAM: DDR4 prices will rise more than DDR5; if you’re still on DDR4, consider completing your upgrade quick.
- Strategy: prioritize capacity over marginal speed; compare 2×16 GB versus 2×8 GB, and 2 TB TLC versus 1 TB “premium” models.
Small and medium businesses
- Planning: longer lead times; quarterly forecasts no longer suffice.
- Capex: set aside additional budget for memory and storage; negotiate annual contracts with delivery schedules.
- Architecture: combine nearline with QLC where latency allows; evaluate compression/deduplication and more aggressive tiering.
Hyperscale and large enterprises
- Multi-year contracts with manufacturers (not just distributors).
- Dual supply chains and secondary providers (when available).
- Efficiency: optimize dataset movement, caching, and formats (e.g., parquet, zstd), and strengthen observability of real-world HBM/DRAM usage by workload.
Three ideas to mitigate impact
- Buy wisely: choose TLC with good DWPD and proven controllers for production; QLC for heavy reading and low write cycles. Avoid false savings with low-quality brands or lines.
- Hybrid architectures: not everything needs to reside in NVMe. Keep layers (memory → NVMe → QLC → nearline), with movement policies and data expiration.
- AI optimization: before “fixing” with more hardware, adopt efficiency techniques (quantization, LoRA, sparsity, smart shard loading, inference batching). Reducing memory per request is gold in this cycle.
Will things return to normal?
The memory/storage market is cyclical. New fabs will arrive, some projects will falter, and there will be oscillations. But the price floor could stay higher than 2023 for three reasons:
- Structural demand driven by AI (training + inference + analytics),
- Capex discipline learned from past cycles,
- Rising costs for equipment and labor in new regions.
Until then, the strategy is plan, diversify, and maximize efficiency. Memory and flash have become scarce assets again. And when oil prices rise, everything depending on it feels it: servers, racks, GPU pods, PCs… and yes, those 2 TB SSDs you won’t see at €80 anytime soon.
via: tomshardware