VideoCardz first reported Pua’s comments from a Taiwanese interview, in which he described DRAM and NAND supply as structurally constrained by AI infrastructure demand. He said memory suppliers were seeking as much as three years of prepayment, and forecast major pressure on smaller consumer-electronics makers through the end of 2026. Tom’s Hardware and PC Gamer separately covered the interview and its translated claims, including Pua’s belief that many lower-margin manufacturers could cut product lines or leave the market.
The important correction is that a shortage “until 2030” is not an industry forecast backed by disclosed production data. It is Pua’s worst-case assessment of a supply chain where NAND manufacturers have more pricing power than their customers. Phison’s own recent financial reporting supports the narrower conclusion: NAND supply remains tight, prices are elevated, and enterprise and AI demand are consuming a growing share of the company’s attention. It does not establish that every consumer SSD will remain scarce or expensive for the next four years.
For Windows users, the practical message is less dramatic but more useful: price shopping still makes sense, but waiting for a broad NAND-price collapse is now a risk rather than a strategy.
Phison has a direct interest in the enterprise shift
Phison is not a NAND fab operator like Samsung, Kioxia, Micron, SK hynix, or SanDisk. It is a major controller designer and storage supplier whose products appear inside many retail SSDs, OEM drives, industrial devices, and enterprise systems. That position gives Pua useful visibility into component purchasing and customer demand, but it also gives his company a commercial incentive to prioritize better-margin buyers.
That incentive is no longer subtle. In Phison’s November 2025 results, the company said NAND availability was constrained and cited continued cloud-service-provider adoption of enterprise SSDs. Its May 2026 quarterly announcement again said the NAND market remained tight, with higher NAND prices, while highlighting AI platforms and enterprise SSDs as growth areas.
Independent reporting makes the direction clearer. PC Gamer reported in March that enterprise SSDs accounted for roughly 30% of Phison’s first-quarter 2026 revenue. In a May earnings-call account, Phison management was reported as saying consumer business had fallen to less than 3% of the company’s mix while retail allocation had been cut sharply.
That does not mean Phison is abandoning consumer SSDs, or that every drive using a Phison controller will disappear from shelves. It means the company selling controllers into both markets is making more money where hyperscalers and enterprise buyers can absorb higher component prices. Retail SSD availability will therefore depend not only on how much NAND gets manufactured, but on how suppliers choose to allocate it.
For a home PC builder, this is the uncomfortable part of the story. A 2TB PCIe 4.0 drive purchased for a gaming library or a Windows 11 workstation competes indirectly with storage demanded by data-center deployments that place vastly larger orders and sign longer contracts.
The current shortage is documented; the 2030 endpoint is not
Pua’s comments about multi-year shortages were not an isolated complaint. In late 2025, he said TLC NAND pricing had risen from about $4.80 per 1-terabit die in July 2025 to about $10.70 by November. Tom’s Hardware and PC Gamer both reported that Phison said NAND capacity for 2026 had effectively sold out, while the company’s financial disclosures confirmed tight supply and lengthening delivery conditions.
Those are real warning signs for SSD pricing. NAND is not the only cost inside an NVMe drive — controller, DRAM or host-memory-buffer design, PCB, firmware validation, packaging, distribution, and retailer margins all matter — but flash capacity is the dominant input cost for most higher-capacity drives. When the raw NAND component more than doubles in price, bargain 1TB, 2TB, and 4TB SSDs become much harder to sustain.
Still, Pua’s 2030 prediction should be treated as a scenario, not a calendar commitment. NAND producers can add output through wafer-capacity expansion, denser 3D NAND designs, improved yields, and changes in product allocation. Those responses take time and capital, and manufacturers remember the industry’s recurring problem: heavy expansion during a shortage can produce a supply glut and collapse prices when demand cools.
There is also a crucial distinction between tight supply and permanent retail scarcity. The market can remain constrained overall while consumer drive prices move unevenly. A vendor may substitute different NAND, reduce overprovisioning, change controller configurations, sell lower-end QLC models more aggressively, or trim warranty terms to keep a product near a price target. That is why an SSD buyer should compare exact models and specifications rather than assuming every 2TB NVMe drive has become equivalent.
Silicon Motion, a Phison competitor in the SSD controller business, offered a less sensational but still concerning outlook in a recent Tom’s Hardware interview. Its senior vice president Nelson Duann said data centers could receive 70% to 80% of memory allocation, leaving 20% to 30% for smartphones, PCs, automotive products, and other markets, and suggested pressure could worsen in 2027. That independently supports the immediate risk to consumer supply without validating a shortage running unchanged through 2030.
The Vera Rubin example mixes a real demand issue with an uncertain calculation
The most striking part of Pua’s warning was the assertion that tens of millions of Nvidia Vera Rubin systems, each with more than 20TB of SSD, could absorb around 20% of the previous year’s global NAND production. The scale is plausible enough to draw attention: AI infrastructure requires enormous data pipelines, checkpoints, model files, datasets, and fast local storage.
But neither Nvidia’s public Vera Rubin product material nor its published NVL72 specification establishes a standard 20TB SSD requirement per Rubin unit. Nvidia lists 20.7TB of HBM4 GPU memory for the Vera Rubin NVL72 rack-scale system. HBM is extremely fast memory attached to accelerators; it is not NAND flash and it is not an SSD.
Nvidia has also described the larger Vera Rubin NVL144 CPX platform as having 100TB of “fast memory,” while earlier GTC material referenced 75TB of fast memory for Vera Rubin NVL144. Those figures show why AI hardware can create enormous memory demand, but they do not specify NAND storage installed per GPU, per node, or per rack.
Pua may be referring to a storage design assumption used in a particular AI deployment, not a universal Rubin configuration. The distinction matters. Counting “tens of millions” of GPUs or systems while assigning each more than 20TB of SSD can generate a massive NAND figure, but the total changes radically depending on whether the unit is a GPU, a server, a rack, or a complete cluster.
Windows Central reached a far lower estimate earlier this year, calculating that projected Vera Rubin server shipments would consume roughly 2.8% of expected 2026 global NAND demand, based on its stated shipment and storage assumptions. Neither estimate is a vendor-issued deployment plan, and Nvidia has not published enough storage configuration detail to settle the discrepancy.
The defensible conclusion is that Rubin-scale AI infrastructure will increase flash demand. The claim that it will consume one-fifth of global NAND output is a conditional calculation, not an established market fact.
What Windows PC owners should do now
For buyers who need storage in the next few months, the best response is to purchase based on capacity requirements and verified drive quality, rather than trying to guess the exact bottom of the NAND market. A full system drive, a developer workstation with virtual machines, and a game-library expansion each have different failure and performance consequences.
A few practical rules follow from the supply situation:
- Buy the capacity you genuinely expect to use over the next two to three years, especially if a known-good 2TB or 4TB model is on sale. Replacing an undersized 1TB boot drive later can erase any money saved by waiting.
- Check the exact SSD revision, NAND type, endurance rating, and warranty before buying. Under price pressure, manufacturers can change NAND or controllers within a retail product line, and a familiar model name does not guarantee identical hardware.
- Keep a proper backup before migrating Windows or installing a new drive. Tight supply and higher pricing are reasons to protect existing storage, not reasons to postpone backup hygiene.
- Avoid paying a large premium solely for PCIe 5.0 throughput unless the workload benefits from it. For most Windows gaming systems, a reputable PCIe 4.0 NVMe SSD remains fast enough, and the money saved may buy more capacity.
- Treat extremely cheap no-name drives with caution. A shortage creates incentives for weak firmware support, low endurance, undocumented component substitutions, and poor warranty service at the bottom of the market.
The strongest evidence says NAND supply is tight now, enterprise buyers are receiving priority, and retail SSD pricing may remain volatile into 2027. Phison’s 2030 deadline is a warning from a supplier with a stake in that high-margin transition, not a fixed expiration date for affordable storage.
