人工智能热潮的最大赢家刚刚索要280亿美元

Key Takeaways

• A quarter that ate a year: SK Hynix booked ₩40.3 trillion in net profit in the first quarter of 2026 alone, nearly matching its entire haul for all of 2025.
• The raise equals the profit: The Nasdaq offering is sized to bring in roughly $28 billion, almost exactly what the company earned in net profit across its strongest year yet.
• Nobody is cashing out: Every share in the deal is newly issued. This is not insiders selling the top; it is the company itself choosing to raise equity at one.
• A pricing trap: Korean securities law restricts how far below its Seoul-listed stock SK Hynix can price the new US shares, tying the deal’s size to a home-market tape that is falling into the debut.

A Record Quarter, and a Tin Cup

Start with the number that makes the rest of this story strange. In the first quarter of 2026, SK Hynix reported net profit of ₩40,346 billion. That is roughly $26 billion in a single three-month stretch, using the ₩1,533.44-per-dollar rate the company itself cites in the filing. For the whole of 2025, its best year on record, it earned ₩42,948 billion. The company made almost as much money in three months as it did in the previous twelve.

A business printing cash at that rate does not obviously need to ask strangers for more. Yet on July 6, 2026, SK Hynix filed the second amendment to its Form F-1, the registration statement for a United States Initial Public Offering (IPO), and the numbers inside describe a raise of about $28 billion. That figure is not a coincidence you should skip past. Converted at the filing’s own exchange rate, the $28 billion it wants from public markets is almost exactly the ₩42,948 billion it earned in profit last year.

$\frac{42{,}948 \text{ billion KRW (FY2025 net profit)}}{1{,}533.44 \text{ KRW per USD}} \approx 28.0 \text{ billion USD} \approx \text{net IPO proceeds}$

The dominant maker of Artificial Intelligence (AI) memory is raising, in one stroke, a sum equal to its strongest year of earnings. The interesting question is not whether the deal will sell. With that profit engine behind it, it will. The question is why a company this flush is holding out a cup at all, and what the person dropping money into it is actually buying.

What the Filing Actually Says

Strip away the headlines and read the terms. SK Hynix is selling 17,790,000 common shares in the form of American Depositary Shares (ADS), the mechanism that lets a foreign company trade on a US exchange. Each ADS represents one-tenth of a common share, so the deal is roughly 178 million receipts. The estimated net proceeds are about $28.0 billion after underwriting fees, and the shares will list on the Nasdaq Global Select Market under the ticker “SKHY.” The existing Korean listing keeps trading in Seoul under code 000660. Reuters reports the offering is scheduled to price on July 9, with the Nasdaq debut the following day.

Two details matter more than the topline. First, the use of proceeds. The company states it intends to use the money “for general corporate purposes, including capital expenditures.” That is boilerplate, but it is honest boilerplate: this is money to build, not money to distribute. Second, and more important, every share in the offering is newly issued by the company. No founder, no fund, no early backer is using this listing to sell into the AI frenzy. When a company insider dumps stock at a peak, that is a signal to worry. This is the opposite pattern. The people who know the business best are diluting themselves to raise cash they will spend.

Three names have already committed to catching a large slice. Baillie Gifford, funds managed by Coatue Management, and an entity called Situational Awareness Partners LP have indicated interest in buying up to $7 billion of the offering as cornerstone investors, though those indications are not binding. That is a quarter of the entire deal spoken for before the roadshow ends.

The Korean-Law Pin

Here is the mechanism much coverage misses. The price of this US listing is not floating freely toward whatever demand supports. It is tethered to a Seoul share price that has been falling.

Under Korean securities regulation, a company issuing new shares faces restrictions on pricing them at a discount to the recent trading price of its existing stock. The filing spells this out: the IPO price will be set through negotiation with the underwriters but is constrained by the last reported trading price of the common shares on the Korean exchange. In a rising market that is a floor. In a falling one it is a trap. Bloomberg reported that SK Hynix shares in Seoul dropped about 17% in the month heading into the deal, dragging the placeholder offering size down from an initially indicated figure closer to $29 billion. The pricing is chained to a tape that is moving the wrong way in the final week.

That is why the exact size keeps shifting and why the debut is genuinely uncertain rather than a formality. The deal is large enough to be the biggest US listing ever by a foreign company, per Bloomberg’s framing, but its final dollar amount is being decided by Seoul’s closing prints in the days before it opens, not by a fixed number the company chose. It would rank alongside the season’s other mega-listing, SpaceX’s public market debut.

Why Raise Money You Just Printed

If the profit is real, why the raise? The thing SK Hynix sells is about to get much more expensive to make, and the spending comes before the revenue.

The company is the dominant supplier of High-Bandwidth Memory (HBM), the stacked memory that sits next to every AI accelerator and feeds it data. It is the same product that drove the 2025 AI memory shortage. By its own citation of research firm International Data Corporation (IDC), SK Hynix held a 56.4% share of the HBM market by revenue in the first quarter of 2026, and it was the first company to mass-produce the HBM3 and HBM3E generations. Holding that lead through the next node is a capital problem, not a demand problem. Reuters reported that the company committed to spending 100 trillion won, about $64 billion, to build new chip plants including a NAND flash facility. A single leading-edge fab can cost more than the entire proceeds of this offering.

So the boring, correct reading of the raise is the one the filing gives you. Selling equity while your stock is up several hundred percent on the year is the cheapest capital a company will ever get. You are trading a small slice of ownership for cash at a rich price, and you are locking in that price before the cycle turns. On the terms the company controls, this is a clean, rational decision. It funds the fab plans without loading the balance sheet with debt, and it does it at a valuation the company may never see again.

That last clause is the whole risk.

The Cycle in Its Own Numbers

How violent is the cycle the buyer is stepping into? The filing answers with a figure that has nothing to do with AI hype. Two years before this boom, in a single bad year, SK Hynix posted a loss in the billions of dollars. The whole arc sits in the registration statement, four lines long.

Won is converted at ₩1,533.44 per US$1, the rate the filing cites; at that rate a trillion won is only about two-thirds of a billion dollars, which is why trillion-won figures land as billion-dollar sums. Every figure in that table comes from the registration statement. Read it top to bottom. In 2023, after memory prices collapsed on oversupply starting in late 2022, SK Hynix lost ₩9,138 billion for the year. Two years later it earned ₩42,948 billion. That is the swing you are buying into: a company that went from a $6 billion annual loss to a $28 billion annual profit in 24 months, in a market structurally capable of running the movie in reverse.

The bull case says this time is different: AI accelerators need HBM in volumes the old PC and phone cycles never produced. That may well be true. But “the cycle is dead” is a famously expensive belief in commodity investing, and the buyer at Thursday’s price is paying peak-cycle earnings for a company whose own five-year history contains a full-year loss.

The Froth Tells

Three smaller facts, read together, describe the market this deal is landing in.

First, the leverage is arriving before the stock does. ProShares filed a summary prospectus, dated July 5, 2026, for an “Ultra SK hynix” exchange-traded fund designed to deliver two times the daily performance of the SKHY receipts. A 2x leveraged product exists on paper for a security that has not yet traded a single share. Second, the arbitrage desks are already at work: UBS advised clients to buy the new US receipts and sell the Seoul-listed stock, betting the ADS would trade at a premium, according to Bloomberg. Third, the customer base is narrow. The filing discloses that a single customer accounted for 23.9% of total revenue in 2025, and the two largest together made up 27.2% of revenue in the first quarter of 2026. The filing does not name them. The concentration is the point: this is a company whose fortunes ride on a handful of AI buyers continuing to place enormous orders.

None of those three facts is damning on its own. A leveraged ETF is a product decision, arbitrage is what desks do, and customer concentration is normal for HBM. Together they sketch a market pricing in permanence, wiring up the plumbing for a stock everyone already assumes only goes up. It is the same concentration that let two chip stocks swing all of Korea’s market in June. That is the mood at every cycle peak, and it is never wrong until it is.

What to Watch Thursday

The consensus version of this story is that AI’s leading memory winner is cashing in on the boom. The filing tells a more precise and more interesting story. A company that just earned a year’s worth of profit in a single quarter is raising another year’s worth in fresh equity, spending it on fabs, and pinning the price of that raise to a home-market stock that is sliding into the open. For SK Hynix, this is a smart trade. For the person buying SKHY on day one, it is a bet that a famously cyclical business has stopped being cyclical.

That bet might pay. HBM demand is real, the market share is real, and the profit is very real. But the case for caution is not an analyst’s opinion. It is on page after page of the company’s own registration statement, in the same document that reports the banner quarter. Both things are true at once, and the filing does not hide either.

Before you decide which SK Hynix you are buying, the boom-quarter machine or the deep-loss machine, read the F-1 itself. It is free, it is on the SEC’s website, and it tells you exactly what you are inheriting. The company was honest enough to print both halves, and the only real mistake a buyer can make this week is to read just one of them.