NewsStocksSK Hynix Posts Record Quarter as AI Memory Demand Outstrips Supply

SK Hynix Posts Record Quarter as AI Memory Demand Outstrips Supply

Author: Decrypt·

Key Takeaways

  • •SK Hynix's quarterly revenue exceeded 100 trillion won for the first time in its history, and its operating profit of 83.3 trillion won surpassed the 77.5 trillion won expected by analysts surveyed by Bloomberg.
  • •Net income of 39.8 trillion won rose 216% from a year earlier but dropped 58% from the second quarter, driven by 32.6 trillion won in net non-operating expenses, including an 11.6 trillion won currency loss and a 22.9 trillion won investment valuation loss.
  • •The company said its DRAM and HBM volumes through 2029 are effectively sold out, with customer demand more than double current supply and the shortfall expected to persist through 2030.
  • •Long-term agreements with 14 key customers, including leading global cloud providers, cover more than 50% of SK Hynix's 2027 and 2028 production capacity, typically run five years, and include safeguards such as deposits.
  • •CEO Kwak Noh-Jung countered AI bubble concerns by citing 30% to 80% year-over-year cloud revenue growth among major customers, order backlogs up two- to five-fold, and broadening AI spending beyond big tech to non-tech firms, sovereign projects, and on-premises setups.
SK Hynix Posts Record Quarter as AI Memory Demand Outstrips Supply

South Korean memory chipmaker SK Hynix posted record third-quarter results on Wednesday, reporting revenue of 104.6 trillion won (almost $78 billion) and operating profit of 83.3 trillion won ($65.5 billion), according to its earnings release. It was the first time in the company's history that quarterly sales topped 100 trillion won ($74.5 billion).

Revenue climbed 328% from a year earlier, while operating profit surged 632%, producing an operating margin of 80% — meaning the company kept about 80 won of profit from its core business for every 100 won of sales. The result topped the operating profit of 77.5 trillion won that analysts surveyed by Bloomberg had expected. Margins like these mark a sharp break from memory's past: the industry has long been defined by boom-and-bust cycles in which oversupply repeatedly dragged prices — and profits — down.

Net income totaled 39.8 trillion won, up 216% from a year ago but down 58% from the second quarter. CFO Kim Woo-hyun attributed the gap to net non-operating expenses of 32.6 trillion won — charges outside chip sales — led by an 11.6 trillion won currency loss and a 22.9 trillion won valuation loss on investments.

Sold Out Through 2029

SK Hynix told investors that demand for memory will keep outrunning supply through 2030, and that its DRAM and HBM volumes through 2029 are effectively sold out. DRAM is the standard memory used in servers, phones and PCs, while high-bandwidth memory (HBM) is the stacked variety that sits beside AI processors — such as Nvidia's market-leading GPUs — and feeds them data. DRAM production is dominated by three manufacturers — SK Hynix, Samsung Electronics and Micron Technology — a concentrated structure that magnifies the impact of any shortfall.

“Given the complexities involved in implementing leading-edge processes and the lead times required to construct new production facilities,” Kim Woo-hyun said, “a situation where supply cannot keep up with demand is expected to persist for some time, a constraint projected to continue through 2030.”

The company said some customers have recently adjusted their HBM specifications, which it reads as proof that demand is overwhelming supply, since the changes mean they must build more AI accelerators from a limited quantity of chips. “Demand remains so strong that customers' volume requirement is more than double that of what we can supply today,” the company said on the call.

Locking In Supply Years Ahead

To secure that demand, SK Hynix has signed long-term agreements — contracts that reserve chip volumes years in advance — with 14 key customers, including leading global cloud providers. The agreements cover more than 50% of the company's 2027 and 2028 production capacity, typically run five years, let prices move within an agreed range and include safeguards such as deposits. Deals of that length are a departure for a commodity-style market that has historically traded on shorter-term pricing, giving suppliers multiyear revenue visibility while guaranteeing buyers access to scarce capacity.

The company also expects its investment to rise significantly in 2027, when the first phase of its Yongin chip complex — a cluster of fabrication plants under construction south of Seoul — enters full-scale operation.

Doubts Over AI Spending

Citigroup analyst Peter Lee opened the Q&A by asking what supports confidence in AI investment when monetization looks slow — a doubt that Citadel founder Ken Griffin had raised eight days earlier when he said investors are “almost certainly” in an AI bubble. CEO Kwak Noh-Jung answered that major cloud customers reported year-over-year cloud revenue growth of 30% to 80% and cloud order backlogs up two-fold to five-fold.

Kwak added that AI spending, once concentrated among a few big tech companies, is broadening to non-tech firms, sovereign AI projects — government-backed national computing initiatives — and on-premises setups, in which companies run AI on their own servers. That broadening, he said, means a slowdown by any single buyer would hurt less.

Financing is tightening for AI infrastructure players, though. Elon Musk wrote on Oct. 25 that “debt markets have almost entirely shut off access to money” for firms without large non-AI revenue.

Looking ahead, SK Hynix expects demand from PC and smartphone makers to weaken as higher memory prices and tight supply weigh on them, but says AI infrastructure demand should offset the decline. How that balance holds — and whether new capacity like the Yongin complex arrives in time to ease the squeeze — will shape how long the current sellers' market lasts.