SK Hynix Sold Out Through 2029 and Still Made $63 Billion

SK Hynix has just delivered the kind of quarter that makes sceptics go quiet, at least for a day. On Wednesday 28 October 2026, the South Korean memory giant reported an operating profit of 83.3338 trillion won for the third quarter — about $63 billion — on revenue of 104.6155 trillion won, and said its DRAM and high-bandwidth memory volumes are effectively sold out through 2029, counting secured long-term agreements together with demand now under specific negotiation. If you have spent the past month wondering whether the AI boom is a bubble about to pop, this is the print you have to reckon with first.

It is not a clean, uncomplicated triumph, and that is precisely why it matters. The same results that smashed expectations on operating profit also showed net profit falling sharply from the previous quarter. The same company handing back roughly 100 trillion won to shareholders is also warning that shortages could last into the next decade. Hold both of those truths at once and you get a much clearer picture of where the AI hardware trade actually stands: demand is real, supply is constrained, and the money is now so large that the old vocabulary for describing a chip company barely fits.

We checked the arithmetic ourselves: 83.3338 trillion won divided by 104.6155 trillion won gives an operating margin of 79.7%. That is not a typo, and it is not a rounding trick. For every 100 won of sales in the quarter, SK Hynix kept almost 80 won as operating profit before interest and tax. It is a record margin for the company, up from 76.3% in the second quarter, and it puts a memory-chip maker — a business that, for most of its history, has been brutally cyclical — in territory normally reserved for software monopolies.

The print, in plain numbers

Start with the headline figures, because they are genuinely startling. Revenue of 104.6155 trillion won rose 327.9% from a year earlier and 31.9% from the second quarter. It was the first quarter in SK Hynix's history above 100 trillion won in sales, making it only the second Korean company, after Samsung, to cross that line in a single quarter.

Operating profit of 83.3338 trillion won rose 632.1% year on year — roughly 7.3 times the 11.4 trillion won it made in the same quarter last year — and 37.6% from the second quarter's 60.5426 trillion won. This was the company's sixth consecutive quarter of record revenue and record operating profit, a run stretching back to the second quarter of 2025. Across the first nine months of 2026, revenue reached 236.5105 trillion won and operating profit 181.4867 trillion won. Those are not quarterly blips. They are a sustained step-change in the scale of the business.

The result also beat expectations, and not by a whisker. The LSEG SmartEstimate for operating profit was 76.7 trillion won, according to Reuters, which puts the actual print roughly 8 to 9% ahead of the consensus weighted toward the most consistently accurate analysts. In a market that has spent weeks punishing anything short of perfection, beating a high bar by that margin is itself a statement.

The market noticed. SK Hynix shares rallied as much as 9.4% in morning trade in Seoul on 28 October, according to Reuters, with the stock up about 8.65% at 10:50am local time to 1,457,000 won. The wider KOSPI rose sharply too, with a buy-side sidecar triggered as buying surged. Whatever else you think about valuations, the immediate verdict from traders was unambiguous: this was not the quarter the bears had been waiting for.

Underneath the totals, pricing did much of the work. DRAM average selling prices rose by a high-20% range from the second quarter, while NAND prices rose by a mid-20% range, according to Reuters-derived reporting. Server DDR5 shipments rose sharply, and sales of 321-layer TLC enterprise solid-state drives expanded. In other words, SK Hynix did not just sell more chips. It sold them, on average, for a lot more money — a classic sign of a market where buyers are competing for scarce supply rather than haggling over surplus stock.

The net-profit catch nobody should skip

Now the complication, because a serious account of this quarter cannot leave it out. Net profit came in at 39.7756 trillion won. That is up 215.7% from a year earlier — a huge annual increase by any normal standard — but it is down 57.7% from the second quarter, when net profit was 93.9226 trillion won.

That gap between operating profit and net profit is the single most important nuance in the release, and it deserves to be stated plainly rather than buried in a footnote. Operating profit surged to a record. Net profit fell sharply quarter on quarter. Both statements are true at the same time, and any article that gives you only the first one is misleading you.

Part of the explanation lies in the second quarter itself. That quarter's net profit was unusually large — so large, in fact, that it exceeded operating profit, a pattern that points to exceptional items outside the core chip business flattering the bottom line. The third quarter, by contrast, looks like a quarter in which the operating business did the heavy lifting and the net line came back toward something closer to the underlying result. Reuters-derived reporting also noted that net profit of around 40 trillion won came in far below an estimate of around 55 trillion won, and that the company's headline figures did not explain the gap.

That last point matters for readers, not just analysts. When a company reports a record operating profit and a net profit that misses consensus, investors are entitled to ask what sits between those two lines — tax, financing, investment gains and losses, and other non-operating items. SK Hynix has not, in the material covered here, closed that question neatly, and this article will not pretend otherwise. The honest summary is this: the chip business has never been more profitable, and the net-profit line is messier than the headline suggests. Do not let either half of that sentence erase the other.

It is also worth keeping the balance sheet in view, because it frames everything that follows. Cash and short-term instruments stood at 114.5 trillion won, up 26.5 trillion won from the end of the second quarter, while borrowings fell to 15.8 trillion won. Net cash was 98.7 trillion won, against 69.4 trillion won at the end of June. A company sitting on that much net cash, in a market this tight, has options that most of its competitors can only envy — and, as we will see, it is choosing to use them.

The guidance — and the qualifier that matters

The guidance, more than the backward-looking numbers, is what will move markets in the weeks ahead. SK Hynix said its DRAM and HBM volumes are effectively sold out through 2029, counting secured long-term agreements together with demand now under specific negotiation. Read that sentence twice, because the second half of it is doing real work.

This is not a claim that every wafer the company will produce through 2029 is already contracted and paid for. It is a statement that combines two things: long-term agreements already secured, and demand that is currently under specific negotiation with customers. Put those together, and the company says, the volumes for the next three years are effectively spoken for. That is still an extraordinary position for a manufacturer to be in. It is not the same as a fully contracted order book, and any reporting that drops the qualifier is overstating the claim.

The structure behind it is concrete. SK Hynix has long-term agreements with 14 major customers, including global cloud providers, and the secured volume under those agreements amounts to more than 50% of total production capacity for 2027 and 2028. Typical terms run to around five years, with extensions under discussion, and pricing reflects market conditions within agreed ranges, with deposits and performance mechanisms included. That is the architecture of a seller's market: multi-year commitments, money down, and prices that move with the market rather than being locked in at yesterday's rates.

Chief executive Kwak Noh-jung put the point in terms of trust rather than tonnage, saying, according to Seoul Economic Daily: "Customer trust built on our technological competitiveness is leading to long-term agreements, improving visibility into medium- and long-term demand and creating conditions for us to secure stable profitability." It is a deliberately unflashy sentence, and that is the point: the company is presenting these agreements as a stabiliser, a way of smoothing out the boom-and-bust cycles that have defined memory for decades.

The demand side, as described on the call, is even starker. Customer requests for HBM amount to about twice what SK Hynix can actually supply, according to Reuters and Seoul Economic Daily reporting of the call. That is a company statement about its own order inquiries, not an independently audited order book, and it should be attributed as such — but if even the direction is right, it explains both the pricing power and the margin. When buyers want twice what you can make, you do not need to discount.

And the company expects the tightness to last. SK Hynix expects supply-shortfall conditions to persist through 2030, citing the difficulty of advanced-process production and the time it takes to build new fabrication capacity. That is a company expectation, not a verified fact about the future, and it should be read as one: a forecast from the party that benefits most from a narrative of scarcity. Even so, it is a forecast with consequences. If shortages really do run through 2030, then the current pricing environment is not a spike to be waited out. It is the medium-term operating environment for everyone building AI data centres.

For 2027, the company expects DRAM demand growth in the mid-20% range and NAND demand growth in the high teens. Capital expenditure is set to rise significantly next year, on advanced-process equipment and on fab infrastructure in Korea and overseas, with the Yongin fab moving to full-scale operations next year. In short, SK Hynix is planning to spend heavily to expand — and still telling customers it may not be enough.

Returning about 100 trillion won to shareholders

Then there is the money going back out the door, which is almost as eye-catching as the money coming in. SK Hynix plans total shareholder returns of about 100 trillion won for 2026, comprising more than 30 trillion won in cash dividends and more than 60 trillion won in buybacks and cancellations, on a principle of returning at least 50% of free cash flow. In dollar terms that is about $75 billion, though the precise dollar figure moves with exchange rates and different outlets have rendered it slightly differently. The won figure is the one to trust.

Within that total, the 28 October announcement added a further 60 trillion won — 30 trillion won in dividends and 30 trillion won in buybacks and cancellations — on top of the 40 trillion won programme announced in August and completed last week. The board approved 14 trillion won in third-quarter cash dividends and a 20 trillion won buyback, of which 10 trillion won is for shareholder returns, to be cancelled, and about 10 trillion won is for employee compensation and is not counted in the return total. The August 40 trillion won buyback, whose shares are to be cancelled shortly, underlines the same intent: this is not a token distribution. It is a structural return of capital.

The third-quarter dividend itself is 20,000 won per share, including a 10,000 won special dividend funded from gains on the company's Kioxia investment. That detail is worth pausing on. A special dividend tied to an investment gain is, by definition, not a recurring operating payout — it is a windfall being shared. It does not diminish the scale of the return, but it does mean the third-quarter dividend should not simply be annualised as if it were the new normal.

Why return this much, this fast? The generous reading is discipline: a company that knows memory cycles turn vicious does not hoard cash at the top of the cycle and then watch it evaporate. The sceptical reading is that SK Hynix is signalling it cannot profitably reinvest all of this windfall at acceptable returns, even with a major capex increase planned. Both readings can be true. A business can be both extraordinarily profitable and genuinely constrained in how fast it can build new capacity — and in that situation, returning cash is the rational move, not an admission of defeat.

What it certainly does is raise the stakes for the next downturn. Shareholders who receive roughly 100 trillion won in a single year will expect the machine to keep producing. If demand softens, the question will not be whether SK Hynix is profitable — it plainly is — but whether a company that has just distributed at this scale can sustain the pace without cutting into the investment its future depends on. That tension, between returning cash now and building capacity for later, is the real strategic story inside an otherwise triumphant release.

HBM4 takes the baton

At the technology level, the quarter marked a handover. Shipments of HBM4, the sixth generation of high-bandwidth memory, began in earnest in the third quarter, including for Nvidia's Vera Rubin platform according to Seoul Economic Daily, and HBM4 is expected to overtake HBM3E in the fourth quarter in both shipment share and revenue. Twelve-high HBM4 is shipping on agreed customer schedules, and the M15X fab in Cheongju is ramping toward full capacity by year-end.

That transition matters commercially as well as technically. Chief financial officer Kim Woo-hyun said: "We have finished negotiating transaction terms with major customers, including next year's HBM volumes and prices, and faithfully reflected the higher product value of HBM4." In plain language, the company has already locked in next year's HBM volumes and prices, and it has priced in the premium for the newer, more capable chip. For a supplier, that is the best kind of certainty: not just demand, but demand at a price that recognises the product's value.

SK Hynix is a key supplier of HBM to Nvidia, and Reuters, citing Nomura data, describes it as the global HBM leader with almost half the market. That position is the foundation of everything else in this story — the margin, the long-term agreements, the pricing power, the shareholder returns. High-bandwidth memory is the bottleneck component in AI accelerators, and the company that leads in HBM leads, in a very real sense, in the infrastructure of the AI era.

None of this means the technology race is settled. Competitors are investing furiously, customers are determined not to depend on a single supplier, and each generation of HBM raises the manufacturing bar. But for now, the baton pass from HBM3E to HBM4 looks orderly rather than disruptive, and it is happening at a moment when customers are asking for twice what the company can supply. That is about as favourable a backdrop for a product transition as a chipmaker could ask for.

What this print does — and does not — say about the AI bubble

Here is where this article takes a position, because the bubble question is the one readers actually care about, and hedging it into meaninglessness would be a disservice. Does SK Hynix's quarter prove there is no AI bubble? No. Does it seriously weaken the claim that AI demand is a mirage? Yes — and that distinction is the honest one.

A bubble, in the sense that matters, is not just high prices. It is spending without underlying use, investment without revenue, and capacity built for customers who do not exist. What SK Hynix has reported is the opposite pattern at the infrastructure layer: customers committing to multi-year agreements, paying deposits, requesting twice the available supply, and absorbing price increases in the high-20% range for DRAM in a single quarter. That is not speculative froth. That is rationing.

None of this settles whether individual AI companies are overvalued, whether some data-centre projects will prove uneconomic, or whether the current pace of investment can continue indefinitely. Those are real questions, and this quarter does not answer them. It is entirely possible for the picks-and-shovels business to boom while some of the miners go bust — indeed, that is how most technology booms have played out. The memory business can be genuinely short of supply while particular AI bets still fail.

But the burden of proof has shifted, and that shift is the story. The sceptical case now has to explain away not just enthusiasm, but contracted volumes, deposits, and a 79.7% operating margin earned by selling a physical product that customers cannot get enough of. That is a much harder argument than pointing at a share price and calling it a bubble. The bubble fears that wiped $480 billion off Chinese AI stocks were real, and the sell-off that followed was painful — but SK Hynix's print is evidence that the Chinese crash and the global AI hardware boom are different phenomena, not two chapters of the same story.

Chief financial officer Kim Woo-hyun framed the demand outlook in similarly grounded terms, saying, according to Reuters: "We expect AI investment momentum to continue, as customers' AI spending is translating into broader adoption of AI services and tangible business results, while the range of companies investing in AI is also becoming more diversified." Note what he is actually claiming: not that AI will change everything forever, but that spending is translating into adoption and tangible business results, and that the investor base is broadening. That is a testable claim about the present, not a prophecy about the future — and it is the kind of claim that makes the bubble debate more serious, not less.

There is a parallel anxiety worth naming, because it ran through markets in the same week. The same AI-spending anxiety behind Anthropic's $63 billion chip financing — the fear that AI companies are committing to enormous hardware bills before their revenues can support them — has not gone away because one supplier had a great quarter. If anything, SK Hynix's pricing power sharpens that anxiety for buyers: the more expensive the memory, the heavier the bill for everyone building AI systems. A great quarter for the supplier is, by definition, a costly quarter for the customer. Both sides of that ledger deserve to be kept in view.

So the reporter-stance verdict is this: SK Hynix's third quarter does not end the bubble debate, but it does reframe it. The question is no longer whether anyone actually wants AI hardware — plainly, they want more of it than the world's leading supplier can make. The question is whether the end users of that hardware can turn it into enough revenue, fast enough, to justify what they are paying. That is a harder, more interesting question than "bubble or not", and it is the one the next few quarters will actually answer.

The Solidigm footnote

Five days earlier, SK Hynix had scrapped Solidigm's planned US listing, halting review of pre-IPO funding and an IPO on 23 October amid market volatility, share-price trends, strong shareholder opposition and regulatory scrutiny over minority-shareholder and dilution concerns, with Reuters reporting that the sale could have raised $15 billion and valued Solidigm at $150 billion. Our 24 October post on that halt used the earlier Bloomberg figures of about $10 billion and up to $100 billion, and the two sets of reported figures should not be blended. Chief executive Kwak addressed the halt directly on the 28 October call, and the company confirmed it will still build a new US fab for enterprise solid-state drives.

What to watch next

Three things will tell us whether this quarter was a peak or a plateau. First, whether the long-term agreements keep converting into shipments at the prices now agreed — the gap between a signed agreement and a delivered, paid-for chip is where many cyclical stories come apart. Second, whether net profit normalises in a way the company explains clearly, or whether the gap between operating and net profit remains an open question that analysts keep returning to. And third, whether the capacity expansion now planned — higher capex, Yongin at full scale, Cheongju ramping — arrives fast enough to ease the shortage without arriving so fast that it creates the next glut.

Memory has a long history of punishing exactly that last mistake. Companies build at the top, supply floods back, prices collapse, and the same firms that looked invincible start writing down inventory. SK Hynix's long-term agreements are, in part, an attempt to break that cycle by locking in demand before the capacity arrives. Whether five-year agreements and deposits are enough to hold the line when the cycle turns is the great unanswered question of this story — and it is a question no single quarter, however spectacular, can settle.

For now, though, the numbers speak with unusual clarity. Revenue above 100 trillion won for the first time. Operating profit of 83.3338 trillion won, up 632.1% on the year, at a 79.7% margin. Volumes effectively sold out through 2029, counting secured long-term agreements together with demand now under specific negotiation. Roughly 100 trillion won heading back to shareholders. A new generation of memory taking over from the old. Whatever happens next, the third quarter of 2026 will be remembered as the quarter when the AI hardware boom stopped being a narrative and became, for one company at least, an almost absurdly profitable fact.

SK Hynix's quarter, answered

How big was the quarter? Record Q3 2026 operating profit of 83.3338 trillion won, up 632.1% year on year, on revenue of 104.6155 trillion won, up 327.9% — a 79.7% operating margin, not over 80%.

Why did net profit fall while operating profit soared? Net profit was 39.7756 trillion won, up 215.7% year on year but down 57.7% quarter on quarter — the post gives that catch its own section; do not quote the operating figure without it.

Is SK Hynix really sold out through 2029? Only with the qualifier: HBM and DRAM volumes are effectively sold out counting secured long-term agreements plus demand under specific negotiation, with HBM requests running at about twice supply, per the company's call.

What is it returning to shareholders? About 100 trillion won in 2026 shareholder returns — over 30 trillion won in dividends (including a 20,000 won per share Q3 dividend with a 10,000 won special from Kioxia gains) plus over 60 trillion won in buybacks.

What changed in memory itself? HBM4 is shipping and is set to overtake HBM3E in the fourth quarter — the generational handover, not just a bigger quarter.

Sources: Reuters via Business Times, Nikkei Asia, Yonhap, Korea Herald, Seoul Economic Daily, Chosun Biz, MoneyToday/Newspim, and SK Hynix results via Morningstar/PR Newswire.

By Daniel Rao, TechPulse USA
Daniel covers AI infrastructure and the money behind it.

Search This Blog

TechPulse USA