The Kingston rep said it out loud. Not in a shareholders’ call, not in analyst commentary — in a Tom’s Hardware article, to the people who build their own computers. NAND costs are up 246%. Prices will continue to go up. Don’t wait.
That sentence is a purchasing recommendation. It’s also an admission that something has broken in the deal.
For twenty years, the deal was: wait. RAM gets cheaper. Storage gets cheaper. The components you can’t afford this cycle will be affordable next cycle. Upgrade when the price is right, not when the manufacturers need the revenue. The deal was structural — it emerged from the dynamics of the industry itself, from manufacturers competing for consumer volume, from yields improving every year, from the whole logic of a market where the primary buyer was a person with a PC.
That deal is over. NAND or DRAM does not became scarce in some temporary way, but the primary buyer changed.
Samsung, SK Hynix, and Micron — three companies, 95% of global DRAM production — have been reallocating their fabs toward High Bandwidth Memory since 2024. HBM is the variant of DRAM that AI accelerators require: the chips inside NVIDIA’s H100 and B200 that run the training and inference workloads. It uses approximately twice the wafer area of standard DRAM, with yields between 50% and 60%, which means every HBM chip that ships consumes the capacity that would otherwise have made three or four sticks of consumer RAM. Micron tracked this internally as a 3:1 conversion ratio. By mid-2026, HBM had taken 23% of total DRAM wafer capacity globally.
I want to name this condition precisely before explaining what it does.
The consumer is now a residual buyer.
That’s not the industry’s framing. The industry calls it a “demand surge” and “supply constraints” — both technically accurate and implying symmetry. The consumer isn’t experiencing a demand surge. Consumer PC sales are roughly flat. What happened is that the manufacturer’s best customer changed, and the economics followed. The margin profile of selling HBM to NVIDIA — or to Microsoft, or to the four hyperscalers that own AI infrastructure — is an order of magnitude better than selling DDR5 to someone upgrading their gaming rig. When that margin gap opened up, reallocation was the only rational response.
16GB DDR5 chips, trading at $6.84 in September 2025, cost $27.20 by December. A 298% increase in three months. 32GB DDR4 kits that sold for $60–90 in October crossed $150 by January 2026. SK Hynix, which bet early on HBM and locked in a tight partnership with NVIDIA, captured approximately 60% of the HBM market — enough to end Samsung’s 40-year run at the top.
Micron discontinued the Crucial brand. This matters more than the price charts. Crucial was the consumer-facing identity Micron used to sell RAM and SSDs to hobbyists and system builders. The consumer was no longer worth a brand.
The NAND story is the same story, two years earlier. In 2023, SSDs were historically cheap — a 1TB NVMe briefly touched $45 at retail. That price wasn’t sustainable and wasn’t the future. It was the residue of a pandemic oversupply: manufacturers had overbought capacity anticipating a consumer boom that stopped, then slashed production to stem the losses, then watched spot prices rebound before anyone adjusted retail expectations. It was real and it was a fluke.
What closed it permanently was data centers needing NAND for inference throughput — fast, repeated reads at a scale that HDDs can’t do. Serving LLM requests requires read speeds that magnetic storage can’t sustain. Cloud providers began absorbing NAND supply at scale in 2024. NAND contract prices rose 33–38% quarter-over-quarter in Q4 2025. Q1 2026 came in at +85–90%.
The analyst consensus says this normalizes by late 2027. That forecast requires AI buildout to plateau and manufacturers to add capacity faster than demand grows. The capacity addition is real — fabs are expanding. The plateau is uncertain. HBM4 is already in development, and each HBM generation requires more wafer area than the last. Nobody is building smaller models. If HBM4 takes more than 23% of wafer capacity, the normalization scenario needs a capability ceiling that isn’t visible from here.
This is the part I find difficult to hold. The “prices normalize by 2028” argument might be right. But it requires a set of conditions that are all downstream of decisions being made by four or five companies building infrastructure for one application category, and “AI demand plateaus” isn’t the direction any of them are currently betting.
Smartphone and notebook manufacturers are already adjusting. TrendForce noted in December 2025 that brands were beginning to ship 8GB where they’d previously shipped 16GB, taking the spec downgrade rather than the margin hit. Multiply that across enough product lines and it becomes the new baseline for what a mid-range device ships with. It doesn’t show up anywhere as a reduction. It’s just what the device is now.
The Kingston rep was right to say don’t wait. What he couldn’t say, because it would be unfair to make him say it: the waiting strategy was only ever available when the consumer was the primary customer. It assumed a market structured around you. That market still exists. You’re just not who it was structured around anymore.