SK hynix breaks 100 trillion won in quarterly revenue with an 80% operating margin, says HBM4 becomes mainstream in Q4, and has over half of 2027-2028 capacity under long-term contracts

The source. The primary material is SK hynix's own newsroom release, SK hynix Announces 3Q26 Financial Results, published on October 28, 2026, together with CEO Kwak Noh-Jung's remarks at the earnings presentation and the board resolutions disclosed alongside it. Every figure, percentage, capacity ratio and product name below is taken from that release. No third-party estimates are used, and nothing the company did not state has been filled in: the customer list, the agreed price ranges inside the long-term contracts, the specific size of 2027 capital expenditure and any single-customer concentration are all undisclosed.

The numbers first. Revenue came in at 104.6155 trillion won, the first quarter in company history above 100 trillion. Operating profit was 83.3338 trillion won, an operating margin of 80%, and net profit was 39.7756 trillion won, a net margin of 38%. Quarter on quarter, revenue rose 32% and operating profit 38%, against second-quarter figures of 79.3187 trillion won in revenue and 60.5426 trillion won in operating profit. The company frames the quarter as another all-time high in both revenue and operating profit, achieved on the competitiveness of high-value-added products in a market where AI memory demand has pushed key product prices up.

What an 80% operating margin actually means. In manufacturing, that number barely exists. It does not reflect a surge in volume but pricing power: the same wafer sold as HBM4 earns several times what it earns as ordinary DDR5. Within the product mix, DRAM was driven by a significant increase in server DDR5 sales plus the start of meaningful HBM4 shipments, while NAND saw 321-layer TLC eSSD sales expand and high-value-added data-centre products grow across the board. The margin therefore does not rest on one hit product but on the entire line shifting toward the high-price tier. The upside of that structure is stability; the downside is that it depends on someone continuing to pay.

The balance sheet matters more than the income statement. Cash and equivalents reached 114.5 trillion won at quarter end, up 26.5 trillion from the previous quarter end; borrowings fell 2.8 trillion to 15.8 trillion won; and net cash therefore rose to 98.7 trillion won. Read those three lines together and the message is blunt: this company does not need to borrow to expand. That is the key to reading the industry's next two years. A supplier holding close to 100 trillion won of net cash, whose downstream capacity is already locked by long-term contracts, has almost no reason to concede on price. For anyone buying memory, whether a cloud provider or a team assembling its own inference boxes, the practical consequence is that next year's procurement budget goes up, not down.

On supply, the company offers a very long timeline. Its demand case is a virtuous cycle: AI-driven productivity gains and new revenue in software development, customer service and search drive more AI adoption, which drives more infrastructure investment, while the investors broaden from a handful of companies in the early phase to general enterprises and nation-level initiatives, so the global AI data-centre infrastructure market keeps growing steadily. On the supply side, the technical difficulty of leading-edge processes and the long lead time to build new fabs mean supply is not keeping pace. The conclusion is that demand will exceed supply through 2030. That is the company's own projection, not a third-party estimate, and it should be discounted for the fact that it comes from the supplier. What gives it weight is that the company acted on it, which is the next paragraph.

The long-term agreements are the part of this release worth copying into your notes. The company says it has signed long-term agreements with 14 major customers, including leading global cloud service providers, that typically run five years, and is actively discussing extensions with several of them. The contracted volume already covers more than 50% of total production capacity for 2027 and 2028. The pricing mechanism matters even more: within a price range agreed with each customer, the contract reflects market conditions, a structure the company says secures stable profitability while still allowing flexible responses to market shifts. In plain terms, there is a ceiling on price increases, but the floor is locked in too. Memory has historically swung on spot prices; this structure puts the cycle inside a signed band. If it holds in practice, the industry's long-standing boom-and-bust character weakens, which is bad for buyers and good for the supplier.

Product cadence and capacity. HBM4 began shipping in earnest this quarter, and the company expects it to surpass HBM3E in both shipment volume and revenue in the fourth quarter to become the mainstream product. To support that switch, M15X will be ramped to maximum capacity by year end. On DRAM, the company will rapidly raise the share of its 10nm-class sixth-generation (1c) process and expand sales of DDR5 and LPDDR5 SOCAMM for AI servers. On NAND, it will raise the 321-layer process above half of domestic production capacity in the fourth quarter to increase supply of high-performance, high-capacity eSSD. Over the medium and long term it intends to develop NAND into a core solution for AI service throughput and cost efficiency, built on HBF and system optimisation. On investment, spending is expected to increase significantly in 2027 as Yongin Phase 1 enters full operation, though the company says it will flex the scale and timing based on customer demand, product competitiveness, profitability and investment efficiency. That last sentence keeps an exit open: it is a promise to spend conditionally, not a commitment to spend.

Shareholder returns and Solidigm. Total returns for fiscal 2026 come to roughly 100 trillion won: at least 30 trillion in cash dividends plus at least 60 trillion in buybacks and cancellations, including the 40 trillion announced in August, which was completed last week with the shares to be cancelled shortly. The board resolved on a third-quarter cash dividend of 14 trillion won and a buyback of 20 trillion won; 10 trillion of the repurchased shares earmarked for shareholder returns will be cancelled in full, while about 10 trillion will go to employee compensation and is excluded from the total. The quarterly dividend is 20,000 won per share, made up of 10,000 won reflecting improved earnings and a 10,000 won special dividend sharing the Kioxia investment gains. The CEO also explained dropping the Solidigm IPO review: the listing was meant to raise external capital for Solidigm's capacity expansion, lighten the company's own investment burden and broaden room for shareholder returns, but on weighing current financial market conditions the company concluded that proceeding now would not deliver the originally anticipated outcome.

How to read this release. First, every figure is company-reported under K-IFRS with no audited quarterly breakdown to check against, and the customer list, price ranges and single-customer concentration inside the long-term agreements are all undisclosed, so the 50%-plus capacity lock can be believed but not verified. Second, the shortage-through-2030 view is the supplier's own forecast and points in a direction that suits it; what to watch is whether capital spending actually follows, and this release gives only a commitment to flex, not a number. Third and most easily missed: an 80% operating margin cannot persist, because margins like that invite capacity expansion and substitution, and the company is itself using five-year contracts to stretch the number out as long as possible. For anyone building AI products downstream, the conclusion is plain: memory will not get cheaper inside next year's inference costs, and if your unit economics assumes hardware price declines, that assumption needs changing now.