1. What did SK Hynix report for the third quarter?
South Korean memory manufacturer SK Hynix said on Wednesday, October 28, that its third-quarter revenue reached 104.6 trillion won, the first time quarterly sales have passed the 100 trillion won mark. Operating profit came to 83.3 trillion won, an operating margin of roughly 80%. In this SK Hynix record quarter, the company also warned that demand for AI memory will exceed what it can supply through 2030.
The figures represent a steep jump from a year earlier. Revenue climbed 328% compared with the same period in 2025, while operating profit rose 632%. A margin near 80% means the company kept about 80 won of operating profit for every 100 won it sold, a ratio that is almost unheard of for a manufacturer that spends heavily on fabrication plants and equipment.
Analysts surveyed ahead of the release had expected operating profit of 77.5 trillion won, so the actual result landed meaningfully above forecasts. The company framed the quarter less as a one-off peak than as evidence of a structural imbalance between how much memory customers want and how much the industry can make.
To put the scale in context, a single quarter now exceeds the 100 trillion won threshold that analysts had long treated as a distant milestone for the memory industry. Memory makers have traditionally reported modest margins in weak years and strong ones in good years, but a profit margin of about 80% on more than 100 trillion won of sales shows how far pricing power has shifted toward suppliers of AI-oriented chips. The report covers the July-to-September period and was released alongside the company's earnings presentation.
2. Why did net income grow more slowly than operating profit?
Net income for the quarter was 39.8 trillion won. That is 216% higher than a year earlier, but it is also 58% lower than the second quarter, a pattern that looks odd beside an operating profit of 83.3 trillion won. The gap comes from items that sit outside the business of selling chips.
Chief Financial Officer Kim Woo-hyun attributed the difference to 32.6 trillion won of net non-operating expenses. Two components stood out: an 11.6 trillion won loss tied to currency movements and a 22.9 trillion won loss on the valuation of investments. Neither reflects how well memory products are selling, which is why operating profit is the figure most analysts use to judge the underlying business.
Currency losses can appear when a company holds assets or debts denominated in foreign money and the exchange rate shifts against it. Valuation losses arise when the carrying value of investment holdings is marked down. Together they absorbed well over a third of operating profit this quarter, leaving net income far below the operating line even as the core business set records.
For readers comparing the headline numbers, the distinction matters. Operating profit measures earnings from the company's normal activities before financing costs, currency effects and investment gains or losses. Net income includes all of those items. A large swing in non-operating results can therefore push net income sharply away from operating profit in either direction without implying any change in how well the factories are performing or how strongly customers are buying.
3. How long does SK Hynix expect demand to outrun supply?
The company told investors that demand for memory will keep running ahead of supply through 2030. It also said its DRAM and high-bandwidth memory volumes through 2029 are effectively sold out, meaning customers have already claimed the output it expects to produce over the next three years.
DRAM is the standard working memory found in servers, smartphones and personal computers. High-bandwidth memory, usually shortened to HBM, is a stacked form of memory placed right beside AI processors so they can be fed data quickly enough to keep their computing cores busy. Training and running large AI models consumes enormous amounts of both, which is why makers of accelerators compete for every chip SK Hynix can deliver.
Kim Woo-hyun said the shortfall will persist for some time, pointing to two causes. Advanced manufacturing processes are difficult to master, and building a new fabrication plant takes years. Even a company willing to spend heavily cannot add capacity quickly, so the constraint is expected to last through 2030.
A supply gap of this length is unusual. In most past memory cycles, shortages lasted a few quarters before producers added capacity and prices eased. Here the company is describing a tightness that it expects to last roughly four more years, and it is already treating output through 2029 as spoken for. That outlook rests on the company's own assessment, and it is a forecast rather than a certainty, but it explains why long-term purchase agreements have become the main tool both sides use to manage risk.
4. What are customers doing to secure chips?
Some customers have changed their HBM specifications, a move SK Hynix reads as a sign of excess demand. Because buyers must build more AI accelerators from a limited pool of memory, they are adjusting requirements to make the available supply go further. On the earnings call, the company said customers' requested volumes are more than double what it is currently able to provide.
To lock in supply, SK Hynix has signed long-term agreements with 14 key customers, including leading global cloud providers. These contracts cover more than half of its 2027 and 2028 production capacity. Such arrangements typically reserve volumes years in advance, can allow prices to move within an agreed range, and often include safeguards such as deposits to discourage buyers from walking away.
For a business that has historically swung between shortages and gluts, multi-year commitments change the risk profile. Memory has long been a cyclical industry, with prices collapsing when supply catches up with demand. Contracts covering more than 50% of capacity for two years reduce, though they do not remove, exposure to that cycle.
The adjustment of specifications deserves attention. When buyers redesign around what they can actually obtain, rather than what they would ideally order, it signals that memory availability, not processor design alone, is setting the pace of AI hardware deployment. SK Hynix interprets this as confirmation that real demand exceeds its capacity to deliver, and it is one reason the company has been willing to commit more than half of two years of output to a small group of customers.
5. Is there an AI bubble behind this demand?
The question surfaced directly on the earnings call. Citigroup analyst Peter Lee asked what justified confidence in AI investment when returns from the technology appear slow to materialize. The query came eight days after Citadel founder Ken Griffin said investors are almost certainly in an AI bubble, a view that has made some market participants wary of chip stocks.
SK Hynix Chief Executive Kwak Noh-Jung answered with customer data. He said major cloud customers reported year-over-year cloud revenue growth of 30% to 80%, and that their cloud order backlogs had expanded two to five times. In his telling, spending is being matched by real revenue and committed orders rather than speculation alone.
Kwak also argued that demand is broadening. Spending on AI is spreading from a handful of large technology companies to non-technology businesses, sovereign AI programs, which are government-backed national computing projects, and on-premises deployments, where companies run AI on their own servers. A slowdown at any single buyer, he implied, would therefore hurt less than it would have when a few customers dominated orders.
The exchange illustrates the tension investors face. On one side are warnings that valuations and spending have run ahead of profits; on the other are suppliers reporting record sales and order books that are already full. Both can be true for a time, because suppliers are paid when equipment is bought, whereas the returns on that equipment may take years to appear. The question for suppliers is whether buyers keep placing orders long enough for the investment to pay off.
6. How are PCs and smartphones affected?
Not every segment benefits. SK Hynix expects weaker demand from makers of personal computers and smartphones, because higher memory prices and tight supply make components more expensive for device builders. When memory costs rise, manufacturers may trim memory capacity in new products, delay launches or pass costs on to buyers.
The company said it expects demand from AI infrastructure to offset that decline. This is a notable reversal from earlier years, when consumer electronics were the largest source of memory demand and data centers a smaller share. Today the order of priority is reversed: the highest-margin products, led by HBM, go to AI customers first.
The split between AI and consumer demand is likely to shape pricing across the electronics industry in the coming quarters, since the same factories that produce memory for servers also produce memory for phones and laptops, and capacity devoted to one product is capacity unavailable to the other.
7. What comes next for SK Hynix?
Investment is set to rise significantly in 2027, according to the company, as the first phase of its Yongin chip complex enters full-scale operation. New capacity is the only lasting remedy for a supply shortage, but the company's own comments suggest that even with Yongin ramping up, the market will remain short of memory for years.
For investors, the main variables to watch are whether long-term agreements keep expanding beyond 2028, whether the non-operating losses that depressed net income recur, and whether the cloud revenue growth cited by Kwak persists. Any softening in AI spending would first show up in order backlogs and in the specifications customers request.
The result also matters beyond Korea. Memory supply is a bottleneck for the entire AI hardware chain, and the pace at which SK Hynix can fill orders influences how quickly accelerators reach data centers. With DRAM and HBM sold out through 2029, the company's output has become one of the scarce inputs that sets the speed of AI expansion.