SK Hynix Filed the AI Warning, Then Sold $26.5B

Introduction

On June 24, SK Hynix told the SEC that "a slowdown in demand for our products from AI infrastructure investment could adversely affect our results of operations." Fifteen days later, BofA Securities, Citigroup, Goldman Sachs and J.P. Morgan led the syndicate that priced $26.5 billion of that same company's stock at $149.00 an ADS, and the underwriting discount fixed at that moment came to $257,545,830. If you bought SKHY when it opened on Nasdaq on July 10, you paid $149.00 into the risk that filing describes, and the four banks that sold it to you had already fixed what they would make on the deal.

The Risk Factor and the Price Sheet

The paperwork is all in one place. SK Hynix registered the offering on Form F-1 on June 24, 2026 under Registration No. 333-296987, filed Amendment No. 2 on July 6, and priced through a 424(b)(4) final prospectus on July 9. The pricing table is unambiguous: 177,900,000 ADSs at $149.00, gross proceeds of $26,507,100,000, proceeds to the company before expenses of $26,249,554,170. Trading opened the next morning under the ticker SKHY.

The AI-slowdown language is one bullet among roughly twenty in the risk-factor summary, boilerplate of the kind every issuer files, not the headline risk the amendment leads with (that spot goes to industry cyclicality and oversupply). What makes this particular bullet non-boilerplate is that the same document supplies the proof: SK Hynix booked a ₩7.73 trillion operating loss and a ₩9.14 trillion net loss in 2023, disclosed a few sections away from the warning.

Concentration makes the bullet sharper. US customers, the same buyers the risk factor names, accounted for 68.8% of SK Hynix's 2025 revenue, and the filing cites IDC data putting the company at 29.1% of the global DRAM market and 56.4% of global HBM revenue in Q1 2026.

What the Fee Table Says

Underwriting discount and commissions: $1.4477 per ADS. Multiply by 177.9 million and you get $257,545,830, an effective 0.972% of gross proceeds, taken off the top before SK Hynix sees anything. That figure was contracted at pricing on July 9, and no subsequent move in the stock or the index touched it.

Before pricing, Korean financial press (Seoul Economic Daily and Korea Herald, both July 4) estimated the underwriters' take at roughly $130 million, working from a 0.5% fee assumption. The filed figure is nearly double that estimate, and it's been sitting on EDGAR since July 9 for anyone who opened the prospectus.

Two other details in the same document didn't make the coverage either. Three anchor investors (Baillie Gifford Overseas, Coatue Management, and Situational Awareness Partners) indicated interest in up to $7 billion of the offering at the IPO price, more than a quarter of the deal, with underwriters earning the identical discount on those shares. The syndicate runs well past the four names that led it, with eight more underwriters from Mizuho to William Blair named in the prospectus and Morgan Stanley absent from it; KuCoin reported that industry sourcing ties the exclusion to bearish research on Korean semiconductor names from its analyst Shawn Kim.

Nineteen Days to the Earnings Miss

SK Hynix reported Q2 across July 28-29 with the call on the morning of the 29th. Revenue came in at ₩79.3 trillion (about $59.5 billion), operating profit at ₩60.5 trillion, a 76% operating margin, and net profit at ₩93.9 trillion. Every one of those is an all-time record in absolute terms. Consensus had modeled roughly ₩84 trillion in revenue and ₩64 trillion in operating profit, so the record print missed on both lines. Analysts attributed ₩63.3 trillion of that net profit to gains on the company's Kioxia stake rather than to memory operations.

The market treated the miss as confirmation of something already in motion. SK Hynix fell 14.65% on July 28 and another 9.61% on July 29. The KOSPI dropped 10.84% on the 28th, its fourth-largest single-day decline on record and the eighth circuit breaker of 2026, then 5.98% on the 29th, the first back-to-back circuit-breaker days in the index's history. From an intraday peak of 9,385.59 on June 19 to a trough of 5,262.77 on July 29, the index lost 43.9% in 40 calendar days and roughly $2 trillion in market value.

Hyperscalers Raised Guidance Through the Crash

The headline version of this is "the AI bubble burst," which the buyers' own capital budgets don't support. Amazon raised 2026 capital expenditure guidance to $220 billion from $200 billion on July 30, with Andy Jassy tying the increase to higher memory chip costs and warning Amazon still won't have enough capacity for 2026 demand. Alphabet raised to $195-205 billion while posting $44.9 billion of Q2 capex and negative $5.9 billion in quarterly free cash flow. Meta guided $125-145 billion after two raises; Microsoft guided about $190 billion. Combined, the four are near $725 billion for 2026, up 77% from roughly $410 billion in 2025. The growth rate is decelerating, not the spend: UBS models +76% this year, +25% in 2027 and +6% in 2028, with no hyperscaler announcing an absolute cut.

SK Hynix said as much itself, on the morning of the second drop. Management told the July 29 call that "we believe the industry is moving into a stage of improving utilization and accelerating the monetization of the infrastructure that has already been built, rather than entering a phase of reduced AI investment," and announced a record $31 billion 2026 capex plan on the same call, raising the figure instead of pulling back.

So what actually broke? Margin debt, mostly domestic. South Korea's first single-stock 2x leveraged ETFs, tied to Samsung and SK Hynix, launched on May 27, and by late June retail investors had net-bought ₩14 trillion of them. Outstanding margin loans hit a record ₩38.63 trillion on June 24, with ₩9.1 trillion tied to Samsung and SK Hynix alone by June 19, up from ₩2.53 trillion at the end of 2025. When the unwind came, the KODEX SK Hynix leverage ETF fell more than 80% from its June 23 peak, ProfG Media put retail losses on these products near $39 billion with more than 3% of South Korea's adult population margin-called, and ABC News Australia reported, citing Korean media, that more than 60% of liquidated investors were under 30.

Who Benefits

The four Global Coordinators get money, through the cleanest mechanism in the story: the underwriting discount is negotiated before the shares change hands and deducted from gross proceeds at closing. The discount was earned on July 9, before the 43.9% index crash and the circuit breakers that followed it. These are also the banks that marketed the deal with the risk disclosure in front of them, because they helped assemble the document that contains it.

SK Hynix gets money and position. It cleared $26,249,554,170 before expenses at a price set during the rally, capital that Korean trade press (BigGo Finance, citing the filing) puts toward the Yongin Y1 fab, the Cheongju P&T7 packaging plant, and about ₩12 trillion of EUV lithography equipment. That buildout defends a 56.4% HBM revenue share against Samsung's catch-up effort, and the raise settled before the miss printed.

For Baillie Gifford, Coatue and Situational Awareness Partners, the benefit is access: up to $7 billion of allocation at the offer price with no bidding risk, a preferential arrangement disclosed exactly where the law requires and covered almost nowhere. And the buyer who paid the offer price on July 10 gets what's left: an open position in a company whose own filing describes the downcycle and whose 2023 income statement shows what it cost last time.

The Only Number That Stopped Moving

Nobody hid anything here. The warning is in the file, and so are the fee table, the 2023 loss, and the customer-concentration figure behind both. What disclosure doesn't do is redistribute the risk it describes, and the 424(b)(4) makes that unusually visible: one line in the transaction settles permanently at pricing, and every other line stays live.

Which is why the "$149 buyer got wiped out" version is wrong. SKHY traded at $155.37 on August 24, still above the offer price, after a 4.92% drop tied to US pressure on Korean chipmakers to build memory fabs stateside. The heaviest losses landed on retail investors in Seoul holding 2x leveraged products that launched on May 27, two months before the crash. The asymmetry is narrower than "buyer lost, banks won": the buyer who held on is currently sitting above the offer price, not underwater.

That's also what the "AI bubble bursting" headline buys. It converts a market-structure failure and a fee-timing question into weather, something that happened rather than something that was priced. Sequoia's David Cahn puts the annual gap between AI infrastructure spending and AI ecosystem revenue at roughly $600 billion, and the share of hyperscaler capex funded by incremental debt has climbed from 9% in FY24 to 32% by mid-2026. Spending is rising and the funding behind it is getting more fragile at the same time, which is a slower problem than a bubble popping in 40 days.

The Bottom Line

The 76% operating margin is the tell in this whole sequence. SK Hynix printed the best quarter in its history and the stock fell 9.61%, because the price had already been set against expectations the company had told the SEC in writing might not hold. Buyers absorbed that gap; the underwriters had stepped out of it nineteen days earlier, for $257,545,830.

The open question is what happens on the next one. UBS has hyperscaler capex growth falling to +6% by 2028, and SK Hynix's proceeds are committed to fabs and EUV tools that take years to come online. If the deceleration arrives on schedule, the risk-factor bullet stops being boilerplate and becomes a description of the operating environment. So who prices the next $26 billion memory offering into that window, and does the fee table on that one get quoted before the deal closes or after?