The perfect storm anticipated in DRAM memory could be mirrored—and last longer—in flash NAND. Pua Khein-Seng, CEO of Phison, has warned that the NAND supply will not be able to keep pace with demand and that the market faces a “supercycle” that could become strained in 2026 and continue for “up to ten years”. This warning, reported by the Taiwanese magazine CommonWealth and cited by Tom’s Hardware, comes the day after news of the mega-agreement between OpenAI and Samsung & SK Hynix to reserve nearly half of global DRAM production for its Stargate AI project.
The combination of these two trends —DRAM hoarding and structural NAND scarcity— sketches a scenario where memory shifts from being a “cyclical commodity” to a strategic resource for AI. Effects would extend from data centers to consumer PCs.
From training to monetization: why inference drives NAND demand
Pua’s thesis is straightforward: major cloud providers have already made massive investments in GPU and HBM for training models during 2022–2023. The next phase —monetization— involves large-scale inference, which is not only taxing on compute but also strains storage.
Specifically, the executive emphasizes “nearline storage”: an intermediate category that’s not as fast (or expensive) as hot storage nor as cheap (or slow) as cold/archival (tape). Nearline stores previews of models, non-daily query datasets, traces, user profiles, query results, and artifacts that need to be handy with reasonable latencies. This tier consumes significant NAND volumes.
“After investing so much capital in AI, how will cloud companies recoup their investment? With inference,” summarizes Pua. And inference “requires storage — and plenty of it.”
A long-term imbalance (and incentive misalignment)
Pua recalls that NAND profitability has been fragile for years: when manufacturers invested aggressively, supply surpassed demand and prices fell, making capex return difficult. This vicious cycle led to slashed spending in 2019–2020. Later, in 2023, Micron and SK Hynix redirected a major part of their capex toward HBM—attractive margins—further depleting NAND investments.
The pendulum now swings the opposite way: after a phase of depressed prices and excess inventory, SanDisk, Micron, and Western Digital have announced freezes and price increases amid a demand recovery. If AI continues to absorb resources at massive scale, the previous underinvestment may lead to insufficient supply to meet the new peak.
Pua’s prognosis: 2026 will be “severe” for NAND, and beyond that, supply will remain constrained “until 2036” unless capex accelerates throughout the entire supply chain (fabs, equipment, materials).
DRAM today, NAND tomorrow: two sides of the same pressure
The alarm in DRAM has been loud: OpenAI’s deal to consume about 900,000 wafers/month until 2029 — in a market predicted to hit 2.25 million monthly by late 2025 — effectively locks in about 40% of the this global capacity. To sustain this, the industry needs approximately $160 billion in new capex, from ASML (lithography) to Applied, Lam, or Tokyo Electron (process equipment).
The NAND sector does not operate in isolation from these flows: capex and engineering resources (tools, white rooms, technicians) compete for resources. Companies adjust their investments toward endeavors with higher margins (HBM, advanced packaging). If industry priorities shift towards HBM and logic chips, NAND gets pushed back, slowing its ramp-up precisely when inference needs it most.
What does a “supercycle” of memory mean?
The term describes periods during which structural demand grows faster than supply, elevating prices and keeping them high for years — not just months — until capex reestablishes equilibrium. For NAND, the ingredients are clear:
- Demand: AI storage consumption (nearline), data growth (video, sensors, logs), cloud and edge expansion.
- Supply: Underinvestment previously, diversion of capex toward HBM/logic, increasing complexity (more 3D layers, advanced processes), and supplier concentration.
- Elasticity: few direct alternatives for nearline, especially when high latency and massive capacity are needed — tapes grow, yes, but serve different needs (archival/legal retention).
If capex in NAND does not escalate parallel to DRAM/HBM, Pua’s supercycle scenario doesn’t sound “catastrophic”: it sounds likely.
Who wins and who suffers if Pua is right
Potential winners
- NAND manufacturers with capacity and vertical integration (chips, controllers, modules).
- Equipment providers (lithography DUV, deposition, etching, metrology) if capex responds.
- Controller and firmware designers (like Phison) if customers prioritize efficiency, endurance, and TCO optimization.
Potential losers
- Hyperscalers and SaaS with nearline fleets sized or lacking long-term supply contracts.
- PC and mobile OEMs if NAND costs/volumes compete directly with data centers.
- Startups relying on cheap flash storage access to scale their services.
Is the industry ready for a “tight” decade?
The answer does not just depend on funds. It requires:
- Sustained capex in fabs and equipment (two to three years per fab).
- Engineering to grow 3D layers without compromising performance/endurance (yields).
- Ecosystem: chemicals, masks, reticles, supply of gases and materials.
- Skilled labor and permits (geopolitics, public incentives).
Furthermore, the mix of HBM vs NAND influences priorities. HBM requires advanced packaging (CoWoS, FO-PLP) and interposers — a visible bottleneck today. If suppliers focus resources there, NAND could lag behind, slowing ramp-up just when inference demands it most.
What can different players do now?
Hyperscalers and large SaaS providers
- Secure multi-year supply contracts with NAND vendors (capacity, prices, SLAs).
- Redesign tiers: shift from “hot to nearline” faster, compress, and deduplicate without sacrificing performance.
- Evaluate alternatives (high-density QLC, hybrid tape systems for part of nearline).
OEMs / PC / Mobile
- Plan Bill of Materials scenarios with rising prices and potential bottlenecks in modules.
- Diversify supply by region/manufacturer to mitigate shocks.
Equipment suppliers
- Align roadmaps with NAND in addition to HBM/logic; prepare capacity and after-sales support for faster ramp-ups.
Regulators and governments
- Encourage capex in NAND (not only HBM/logic), facilitate permits, and strengthen critical supply chains.
Open questions (and realistic ones)
- Will the “severe 2026” scenario materialize for Pua? Clear stress appears likely if capex does not respond immediately.
- Will it last ten years? It depends on how much and how the investments are made, the efficiency of each node/layer, and whether large-scale inference sustains its adoption curve.
- Will it displace consumer markets? The PC/mobile markets will face second-order impacts, but if nearline competes on price/volume, we will see tension in client SSD.
- Will architecture change? More tiering, more compression/deduplication, and more smart control in controllers to extract maximum per TB.
Conclusion: Memory is no longer just an “background” — it’s set to drive AI’s pace
Phison’s warning aligns with the broader picture: AI has transformed memory—DRAM and NAND—from a cyclical commodity into a strategic lever. Yesterday was the DRAM blockade for Stargate; today, NAND is the bottleneck for inference for years. Meanwhile, the industry must fund, build, and operate fabs at an relentless pace amid geopolitical tensions and profit considerations.
If Pua is correct, 2026 will mark the first real-world test: flash memory will no longer be just the cheap and abundant piece taken for granted. The success of AI will depend equally on FLOPS and on well-provisioned gigabits.
Frequently Asked Questions
Why does Phison’s CEO talk about a NAND shortage “lasting up to ten years”?
According to Pua Khein-Seng, the market has been suffering from underinvestment in NAND (with capex restrained in 2019–2020 and diverted toward HBM in 2023), even as demand grows from AI inference (nearline) and the exponential generation of data. This supply/demand shock could extend beyond usual 12–18 months cycles.
What is the difference between DRAM and NAND issues in this context?
DRAM is under immediate pressure (e.g., the OpenAI deal for Stargate), tied to training/serving with HBM/DDR5. NAND comes into play mainly in inference and massive low-latency storage. Both compete over capex and industry resources; if priorities shift to HBM, NAND risk falling behind, slowing its ramp when inference most demands it.
How will this affect SSD prices (consumer and data center) in 2026–2027?
If NAND capex does not accelerate, a price cycle of increases and longer timelines seems probable, impacting data center SSDs (nearline/QoS) and, by extension, consumer SSDs. The extent will depend on demand mix, new factory capacity, and 3D NAND layer efficiencies.
What can companies relying on flash storage do to mitigate risks?
Secure multi-year supply agreements, diversify suppliers, optimize tiers (hot/nearline/cold), implement compression/deduplication, and explore hybrid solutions like tape for latency-tolerant data. In design, adopting high-density QLC and firmware prioritizing endurance and QoS can improve TCO under supply stress.