2027's Memory Is Already Sold Out

It is September 2026. 2027 is nearly four months away, and according to a DigiTimes report citing industry insiders, Samsung, SK hynix, and Micron have sold their entire 2027 DRAM and HBM output, with no additional supply planned.

First, what this report is: DigiTimes citing industry insiders, not an official statement from any of the three companies. None has publicly confirmed it. The distinction matters — supply-chain reporting in memory has a decent track record, but it is not a filing or a press release.

There is, however, one official statement pointing the same way.

SK hynix’s CEO said it himself: 2027 is the worst year on record

SK hynix CEO Kwak Noh-jung told Reuters that 2027 will likely be “the worst year in the industry’s history from the supply perspective”, and projected that customer demand will exceed supply capacity “even beyond 2030”.

The CEO of a major memory maker publicly stating that what he sells will be short for more than four years carries more information than any supply-chain rumour. It implies two things at once: they do not regard this demand as a short-lived bubble, and they do not intend to chase it with large-scale expansion.

Why not just build more? This is memory’s old wound

The obvious outsider question: if it is sold out, why not build new fabs?

Because this industry has been taught the lesson too many times. Memory is a textbook cyclical business: demand rises → everyone expands → new capacity arrives simultaneously → severe oversupply → prices collapse → industry-wide losses → capex is cut → supply contracts → demand returns → shortage again.

The cycle is so violent because memory is a highly standardised commodity — same-spec DRAM from different vendors is essentially interchangeable. There is no brand premium to cushion anything, so price is set almost purely by supply and demand: slight oversupply can halve it, slight shortage can double it.

And the supply side adjusts with enormous lag. An advanced memory fab is a ten-billion-dollar-plus commitment that takes years from decision to volume production. That delay creates a structural trap: manufacturers must commit to expansion when demand is strongest and the information is most optimistic, and the new capacity tends to arrive exactly as demand rolls over. The cycle has repeated several times over twenty years, each round washing somebody out, and the survivors’ lesson is: do not expand at the top.

Which is precisely what the three are doing now: raise prices rather than add capacity. Shareholders call that discipline. For buyers it is a structural guarantee of higher prices — because even the hope of “just wait for new capacity” has been removed.

What is different about this cycle

There is one reason this time might genuinely differ, and it also explains why the three feel safe not expanding.

Previous demand peaks were driven by consumers: PC refresh cycles, smartphone penetration, console generations. That demand shares one trait — it is price-sensitive. Push memory prices far enough and handset makers drop the base configuration from 8GB back to 6GB, buyers defer upgrades, demand falls, and prices soften. That was the self-correcting brake in every previous cycle.

This cycle’s demand is not consumers. It is a handful of hyperscalers building AI infrastructure, and their purchasing logic is different: memory is a significant share of an AI server’s cost, but against the scale of the overall data-centre investment and the market position they are racing for, a 20-30% price rise does not cancel orders. This is price-insensitive demand — and it is locking supply into multi-year contracts.

That is the confidence behind SK hynix’s “beyond 2030” line. When buyers do not walk away over price, the self-correcting brake stops working.

This reasoning holds only while AI infrastructure investment stays intense, of course. If capital markets ever run out of patience for AI capex, the demand turn will come faster than anyone expects — which is how every memory crash has begun.

HBM drains the whole pool

The other factor is HBM (high-bandwidth memory) — the memory bolted onto AI accelerators.

HBM is built by stacking DRAM dies vertically and connecting them with through-silicon vias. The crucial point: the things being stacked are ordinary DRAM dies — the same product, from the same wafer capacity, as the modules in your PC.

It also consumes more capacity than its nominal size suggests, for two reasons:

• Yields multiply. Stack eight or twelve dies and a defect in any layer, or a failed bond at any step, scraps the entire stack — including every good die inside it. For the same wafer output, effective yield as HBM is lower than as plain DRAM.
• The margin gap makes allocation obvious. HBM sells far above commodity DRAM. When AI data centres will pay well over consumer-market rates to reserve capacity, the fab’s allocation decision requires no deliberation.

DigiTimes also reports that DRAM for PCs, laptops, and smartphones is expected to be significantly reduced in 2027 versus 2026. The capacity did not disappear; it was allocated elsewhere.

What it means for you

This reads like inside-industry news, but it will reach you in concrete ways:

• If you plan to upgrade RAM or an SSD, sooner beats later. Not a reason to panic-buy, but the supply direction is fairly clear: the report also notes NAND flash capacity is expected to be fully booked by the end of August 2026. With the big three’s output locked into long-term contracts, neither availability nor pricing in retail channels improves from here.
• Laptops and phones get pricier or smaller. Facing higher memory costs, vendors typically either raise prices or hold the price and quietly cut the base configuration. The second is harder to notice — worth checking the spec sheet before buying.
• Memory-heavy cloud instances may cost more. This passes through slowly, but the same cost pressure ends up in cloud pricing.
• Used and refurbished markets heat up. When new supply tightens, residual values on previous-generation DDR4/DDR5 modules and older SSDs tend to hold longer than usual. Idle hardware may be worth more than you think over the next year or two.

The caveat worth keeping: all of this rests on the insider report that 2027 capacity is sold out being accurate. If actual supply proves better than reported, or if AI data-centre build-out slows enough to loosen demand, prices could move quite differently. Memory forecasting is famously unreliable — which is exactly why that cycle exists.

A coincidence worth noting

In the same week, DeepSeek cut its new model’s output price to $0.60 per million tokens and Anthropic cut cache reads by 75%. The price of model inference is falling fast.

Meanwhile, one layer down, the memory those models run on is being locked into contracts stretching into 2027 and beyond, by companies publicly stating that tightness will persist past 2030.

A price war upstairs, capacity being locked downstairs. How long both can hold at once is one of the more interesting questions of the next two years.