SK Hynix Just Raised $26.5 Billion. The Thing It Sells Is the Real AI Bottleneck.

On July 10, 2026, SK Hynix sold 177.9 million American Depositary Receipts (ADRs, a US-listed certificate that stands in for shares of a foreign company) at $149 each and raised $26.5 billion. Every ten ADRs represent one Korean common share. The stock opened at $170 and closed its first session at $168.01, up about 13%.^[2] That is the largest US listing ever completed by a foreign company, past the $25 billion Alibaba raised in 2014.^[1] The order book was reportedly covered seven times over before pricing.^[1] Worth noting the deal actually priced below its own hype: the pre-deal reporting had SK Hynix targeting up to $29 billion.^[4]

Most of the coverage filed it under “Korea discount ends, AI trade continues.” That is the boring read. The interesting thing about SK Hynix is that it sells the one component in the AI stack that almost nobody outside the industry can explain, and that component, not the GPU, is what actually determines how fast a model runs.

The company controls roughly 60% of the market for high-bandwidth memory. Counterpoint data put SK Hynix at 62% of HBM in the second quarter of 2026, with Micron at 21% and Samsung at 17%.^[9] IDC had it at 56.4% by revenue in Q1, and the company itself cited about 57% on its Q1 earnings call.^[5] Call it 60%. In an industry with three suppliers and no fourth, that number is closer to a toll booth than a market share.

The chip is not the bottleneck. Feeding the chip is.

Here is the thing people get wrong. When you hear that an Nvidia GPU does some enormous number of FLOPS (floating-point operations per second, the raw arithmetic throughput of a chip), you assume the arithmetic is the hard part. It is not. The hard part is getting the numbers to the arithmetic units fast enough that they are not sitting idle.

Every time a large language model generates a single token, it has to stream its weights (the model's learned parameters, the actual numbers that make up the model) out of memory and into the compute units. Not some of them. Effectively all of them, every token. A model with 70 billion parameters in half precision is roughly 140 gigabytes of weights that have to move, per token, if you want it to answer at conversational speed. That is a bandwidth problem, and it has nothing to do with how many FLOPS the chip can theoretically do.

This gap has a name in computer architecture. It is the memory wall. Compute throughput has been scaling far faster than memory bandwidth for decades, so the ratio between “how fast can I do math” and “how fast can I get data to the math” keeps getting worse.^[10] Everything about the shape of modern AI hardware is a response to that one divergence.

HBM is the response. Instead of putting memory chips on the far side of the board and talking to them over a narrow, fast serial link, you stack DRAM dies vertically, drill thousands of vertical wires straight through the silicon (through-silicon vias, or TSVs), and park the whole stack millimeters from the GPU die on a shared substrate.^[10] You trade a narrow fast road for an absurdly wide slow one. Thousands of parallel lanes beats a few fast ones, every time, when what you need is throughput.

Why there are only three suppliers, and why there will not be a fourth

People assume three suppliers means a cozy cartel that some new entrant will eventually break. That gets the causality backwards. There are three because the business ate everyone else.

DRAM used to have more than a dozen serious players. Qimonda, Infineon's memory spinout, filed for insolvency in January 2009 after DRAM spot prices collapsed and Samsung kept producing flat out while Qimonda sold every wafer at a loss.^[16]Elpida, which was itself the merged memory divisions of Hitachi, NEC, and Mitsubishi and the world's number three DRAM maker, went bankrupt in February 2012. Micron bought the remains for roughly $2.5 billion.^[16] What is left is Samsung, SK Hynix, and Micron. The oligopoly is a graveyard, not a country club.

HBM then adds a second moat on top of that. Stacking is brutally hard. A single defective die in a twelve-layer stack ruins the entire package, so yields sit well below what these companies are used to on conventional DRAM wafers.^[12] And the capacity math is vicious: each gigabyte of HBM consumes roughly three times the wafer capacity of DDR5.^[11] Every HBM bit a fab makes is three bits of normal memory it did not make. That is why the memory shortage bled into everything else. DRAM contract prices rose something like 83% to 95% quarter over quarter in Q1 2026.^[6] Your laptop RAM got expensive because Nvidia needed waiters.

Which is why a 60% share here is worth more than a 60% share almost anywhere else in the AI stack. UBS reportedly expects SK Hynix to take roughly 70% of HBM4 supply for Nvidia's Rubin platform in 2026.^[13] HBM4 has been running around $500 per 48GB stack,^[15] and Korean reporting suggests HBM4 pricing could roughly double next year, from around $2 per gigabit in the second half of 2026 to $4 to $5 or more.^[14] Supply is spoken for. Micron has said its HBM capacity is sold out through calendar 2026.^[9]

The numbers are, frankly, absurd

SK Hynix's first quarter of 2026 was one of the best quarters any hardware company has ever had. Revenue of 52.6 trillion won, up 198% year over year and 60% quarter over quarter. Operating profit of 37.6 trillion won, an all-time-high 72% operating margin.^[6]

Sit with that. A 72% operating margin. On memory. This is a business that was, for forty years, the canonical example of a commodity with no pricing power. Software companies do not routinely post 72% operating margins. SK Hynix is currently earning software margins on a physical good it ships in trucks.

That left column is the part the bulls keep skipping. In 2023 this company lost 7.73 trillion won for the year on 32.77 trillion won of revenue, a negative 24% operating margin, with a net loss of 9.14 trillion won.^[8] It posted four consecutive quarterly losses. The Q1 2023 operating margin was negative 67%.^[7] It was the company's first operating loss in a decade.^[7]

Thirty months. That is the entire distance between a negative 24% annual margin and a 72% quarterly one. Nothing structural about the company changed in between. The price of the thing it sells changed.

The question the listing is actually asking

Strip away the ticker and the confetti and this offering is a single proposition put to US investors: memory has stopped being cyclical, so you should stop valuing it like a cyclical.

Chairman Chey Tae-won said it more or less directly. He told CNBC the company is confident demand for memory has permanently changed from past boom-bust cycles, and described telling customers SK Hynix would double capacity within five years, only to hear back that it was not enough.^[2]

I believe him about the demand. I do not believe the conclusion he is drawing from it.

“Customers want more than we can make” is not evidence against cyclicality. It is the exact sentence executives say at the top of every cycle, in every capital-intensive commodity business, in every century. It was said about shipping. It was said about oil rigs. It was said about DRAM in 2017 and 2021. The reason it gets said at the top is that the top is definitionally the moment when demand most exceeds supply. That is what makes it the top.

And then the third paragraph of the same story tells you what happens next: SK Hynix is taking $26.5 billion of new capital and spending it on capacity and ASML EUV lithography scanners.^[1] That is the mechanism. Shortage produces price, price produces margin, margin produces capex, capex produces supply, supply kills price. SK Hynix just raised the largest foreign listing in US history explicitly to fund the supply side of that loop. Samsung and Micron are doing the same thing. Nobody is going to under-build here.

So where do I actually come out

Two things are true and people keep insisting on only one.

The structural claim is real. HBM is a genuinely better business than commodity DRAM, and it will stay that way. Three suppliers, punishing yields, multi-year Nvidia qualification cycles, and a product that is sold under long-dated agreements rather than on the spot market. The 3x wafer penalty per gigabyte^[11] means HBM structurally soaks up capacity that would otherwise depress commodity DRAM prices, which is a real and durable support for the whole industry. SK Hynix being the technology leader in a three-horse race with a 60% share is one of the strongest competitive positions anywhere in the AI stack. Better, honestly, than most of the application-layer companies trading at twenty times revenue.

The cyclical claim is also real, and it is the one being priced away. The margin, not the moat, is the cyclical part. A 72% operating margin is not a moat. It is a shortage. Moats persist through supply expansion; shortage premiums do not. Even in a supply-constrained oligopoly, all three players are pouring capital into new capacity right now, and the AI capex that funds the demand side is itself financed by hyperscalers whose spending is not a law of nature.

The market already gave a preview of what that repricing feels like. Three days after the record debut, on July 13, SK Hynix's Seoul-listed shares fell 15.4%, the biggest one-day drop in the stock's history, on nothing more dramatic than a brokerage note questioning whether it would hit quarterly profit estimates.^[3] A single sell-side note took a fifth of the market cap off a company whose fundamentals had not changed in seventy-two hours. That is not the volatility profile of a secular compounder. That is a cyclical wearing a secular costume, and the costume slipped in under a week.

None of that makes it a bad investment. It makes it a cyclical investment, and cyclicals are bought when the margin is bad and the headline is grim, not when a company prints the best quarter in its history and rings the opening bell at the Nasdaq MarketSite in Times Square to a seven-times-oversubscribed book.^[1]

SK Hynix sells the scarcest input in the most important buildout of our lifetime, and I would rather own that than most things. I just would not pay peak-margin multiples for peak-margin earnings and call it secular. The memory industry has buried more companies than most sectors have ever had. Qimonda and Elpida did not fail because they were bad at making memory. They failed because they were fine at making memory in the wrong part of the cycle.