The memory business has spent forty years being the most reliably cyclical industry in technology. On Wednesday, SK Hynix reported what happens at the top of that cycle when demand is not consumer electronics but AI data centres.

Third-quarter revenue was 104.6155 trillion won — the first time the company has crossed 100 trillion, and 327.9% higher than a year earlier. Operating profit was 83.3338 trillion won, up 632.1%. That is an operating margin of roughly 80%: up from 76.3% in the second quarter, some 21 percentage points above TSMC's 58.7%, and the fourth consecutive quarter SK Hynix has out-earned the world's largest foundry on that measure.

What moved

HBM4. The sixth-generation high-bandwidth memory began shipping to major customers in the second quarter of last year, and its revenue contribution grew in the third as AI data-centre spending expanded. Commodity memory followed: server DRAM demand pushed prices up, and enterprise SSD demand pulled NAND with it. Meritz Securities estimates average selling prices rose 22% quarter on quarter for DRAM and 15% for NAND.

The forward book is the more striking number. Executives said the company has sold out of both conventional and high-bandwidth memory through 2029, and that it will invest heavily in capacity. That is the opposite of what memory investors had been pricing in: Nvidia's Jensen Huang spent last week pushing back on the idea of a 2029 supply glut, after memory stocks sold off on exactly that fear.

The payout

SK Hynix will return roughly 100 trillion won — around $75 billion at current rates — to shareholders for the 2026 fiscal year: more than 30 trillion in cash dividends and more than 60 trillion in buybacks and share cancellations, including the 40 trillion buyback resolved in August. A third-quarter dividend of 20,000 won per common share was declared, 10,000 won of it a special payout from investment gains on Japan's Kioxia; the total is 14.1 trillion won, a 1.56% yield at the market price. A further 20 trillion won buyback was resolved, split evenly between shareholder returns and employee compensation.

Handing cash back rather than promising it forward is a choice with a rationale. Capacity takes years to build, the shortage is projected to persist, and a buyback lifts earnings per share now. IDC expects DRAM supply to grow only 16% in 2026 against demand rising far faster.

The cycle has not been repealed

None of this means the memory cycle is dead — only that its timing has moved. Every previous peak in this industry arrived with confident statements about structural demand, followed by capacity landing all at once. Three manufacturers booked out to 2029 is a strong signal about the next three years, not about 2031.

What it does mean is immediate and concrete. HBM is the toll booth on the AI build-out: every accelerator shipped needs stacks of it, and the price is set by whoever holds the capacity. An 80% margin at the supplier is a cost at every hyperscaler, and it is why desktop RAM and SSD prices have been climbing for consumers who have nothing to do with AI. The memory shortage is no longer an abstraction in a quarterly report; it is the number on the spec sheet of the next laptop you buy.

Sources