SEOUL, 31 AUG 2026 — SK Hynix chief executive Kwak Noh-Jung says he sees no signal of memory oversupply and expects demand to outstrip supply until the end of 2030, with 2027 the worst supply year the industry has seen. He said the same thing in July, which makes this a restatement rather than news.
The claim, and the argument behind it
The forecast has two parts. The near term: 2027 will be the worst year for supply the memory industry has ever recorded. The long term: demand exceeds supply through the end of 2030, with no signal of a glut.
The reasoning behind the forecast is more specific, and it is the part worth examining. Kwak's argument is that memory is no longer a commodity in the AI era, that AI memory is moving from fully standardised product toward partly and fully customised product, and that because capacity is being built against specific customer demand, any eventual oversupply should be far less severe than in past cycles.
Who is making the forecast
This is not an accusation of dishonesty. It is a statement about whose view this is, and it belongs in the first paragraph of any coverage.
The chief executive of a memory manufacturer forecasting a five-year shortage is describing conditions under which his company's pricing power is high and its capacity is fully absorbed. Customers negotiating long-term supply agreements hear a reason to commit early at today's prices. That is a legitimate thing for a chief executive to say and it is not an independent assessment.
The forecast to compare this with would come from the buyers. Nobody quoted in this coverage is a hyperscaler procurement lead or an equipment maker with visibility into fab build schedules, and those are the parties whose forecasts would carry the opposite incentive.
Every memory peak has sounded like this
Memory is the most cyclical business in semiconductors, and the cycle has a rhetorical signature.
The pattern runs: demand outpaces supply, prices rise sharply, producers announce capacity, the capacity arrives eighteen to twenty-four months later, demand growth moderates, and prices collapse. The 2018-19 and 2022-23 downturns both followed periods of confident shortage forecasting from producers, and in both cases the capacity that broke the price was ordered during the shortage.
The phrase this time is different is not automatically wrong, and in the memory business it is a claim that requires evidence.
The testable version of the argument
Kwak's case does have a falsifiable core, which is more than most cycle-defying arguments offer.
If high-bandwidth memory is truly customised per customer, with different stack configurations, interfaces and qualification for a specific accelerator, then capacity built for one buyer cannot be redirected. A demand shortfall at one customer would not flood the general market. Supply and demand stop clearing at a single price, and the classic glut mechanism weakens.
If instead the customisation is largely packaging and qualification on top of substantially standardised dies, then the dies can be redirected, and the cycle works as it always has with an extra step.
This is not publicly resolvable. The degree of per-customer specificity in HBM is not public, and it is the single fact that determines whether the argument holds.
Customisation cuts the other way too
The argument is presented as pure downside protection, and it carries a risk that is not mentioned.
Capacity built to one customer's specification is concentrated exposure to that customer. If AI accelerator demand moderates at a small number of very large buyers, a producer with customised lines has stranded capacity rather than sellable inventory, which is a worse position than a commodity glut, not a better one.
The same property that prevents a general price collapse converts a demand shortfall into an asset-utilisation problem. Whether that is preferable depends on which failure you are more exposed to, and a memory maker with long customer relationships will reasonably prefer it. A buyer should notice that the argument for stability is also an argument for concentration.
What this means for buyers in this region
Memory pricing feeds through the whole regional electronics assembly base, and the practical implication is about timing rather than about who is right.
Contract manufacturers in Malaysia, Vietnam and Thailand buy memory as an input and pass the price through with a lag, which compresses margins on the way up. A device maker planning 2027 volumes has to decide now whether to lock supply at elevated prices or take the risk that the forecast is wrong, and that decision cannot wait for the argument to be settled.
The defensible move is to treat a producer's five-year shortage forecast as one input, weighted for its source, and to watch for changes elsewhere in the chain. We reported that Samsung raised foundry prices by up to 15 per cent while losing market share, which is what pricing power looks like when it is real. Prices moving without volume following is the signal that matters, and it is observable now rather than in 2030.
View author profile → · Editorial policy
