SK Hynix Printed 76% Margins. The Market Sold It Anyway.

The pass-through gap is an accounting problem, not a demand collapse. At 3.3x forward earnings, the market is pricing in a recession the physical data doesn't show.

SK Hynix reported the most profitable quarter in its history on Tuesday night. Revenue hit a record ₩79.3 trillion, operating profit surged 557% year over year to ₩60.5 trillion, and the operating margin reached 76%. It was the kind of quarter that would normally send a semiconductor stock soaring.

Instead, the shares fell another 7.4%.

They are now down 21% from their ADR debut price of $168 on July 10 and 11% below the $149 IPO price. The forward P/E has compressed to roughly 3.3 times estimated earnings. That is a cyclicals-in-a-recession multiple for a company sitting on ₩69 trillion of net cash and ramping HBM4 into the most compute-intensive AI cycle on record.

The selloff is real. But it is pricing the wrong problem.

The Numbers

The per-share figures are stark. The ADR priced at $149 two and a half weeks ago, opened at $170, touched $194 intraday on massive volume — and closed at $132 on earnings night. The $26.5 billion record foreign-company IPO raised no floor under the price when the broader memory trade rotated.

The Pass-Through Problem

This is not a demand collapse. It is an accounting dispute.

TengYan framed it precisely the day before the print: the market was pricing "the wrong failure mode." The bear case from Korea Investment & Securities did not call for demand destruction. It called for slower pass-through. KIS revised its Q2 op profit estimate to ₩60.4 trillion — almost exactly what the company reported — on the assumption that blended DRAM price realization would land near +29% rather than the +40% bulls expected.

The reason is contract structure. HBM pricing is largely fixed through annual agreements, and a substantial portion of conventional DRAM also moves through longer-term contracts. Market prices can rise faster than reported revenue can capture within a single quarter. The gap is deferred revenue, not lost demand.

That is the narrow mechanism behind the miss: incremental revenue-to-profit conversion ran at roughly 86%, below the ~98% Micron recently reported but still exceptional for any manufacturing business. Consensus wanted ₩84 trillion of revenue. The company delivered ₩79.3 trillion. The Street's ₩65 trillion op profit call came in at ₩60.5 trillion.

As we noted in Monday's pre-print piece, the selloff into the number was already a leverage-driven unwind — the options positioning showed short gamma and extreme IV, a mechanical amplifier, not a fundamental verdict. The print confirmed the KIS accounting assumption but gave no evidence of a demand break.

Management's Case

The earnings call was unambiguous. SK Hynix management laid out four pillars that contradict the bear narrative:

Jevons, not peak. "We do not see more efficient AI models reducing infrastructure demand. Rather, they will further broaden the AI base. As models and systems become more efficient, the same infrastructure can support more users and services." They cited explosive demand for recent high-efficiency models as evidence that higher efficiency drives broader adoption, not reduced infrastructure spend. The data backs this: weekly token consumption has surged roughly 131% since tracking began, even as the effective price per token has collapsed.

LTAs are locking in demand, not capping it. The company has concluded multi-year agreements with approximately 10 customers, including key hyperscalers. These contracts include deposits and pricing mechanisms designed to reduce volatility. They are not just volume guarantees: they are co-development agreements for next-generation memory aligned to customer roadmaps. The HBM book is being structured like a utility, not a commodity.

HBM4 is real, and the moat is intact. Management pushed back hard on competitor-catch-up fears: "Accumulated competitiveness in time to market, product performance, mass production yield quality, and customer trust are differentiators that cannot be replicated in a short period." HBM4 is in mass production now, with a full second-half ramp underway. HBM4E samples shipped to a major customer in the first half. The HBM leadership position the stock is built on remains intact.

Physical tightness persists. Morgan Stanley's Joseph Moore noted on July 20 that memory shortage intensity "shows no signs of abating," with Q3 prices up at least 25% on a like-for-like basis. The GPU compute market is tightening: H200 spot pricing has entered "Tight" territory, and the blended Compute Tightness Index has climbed 10 points in 30 days.

The Setup

At $132, the math is unusual:

Three analysts cover the stock post-ADR. The consensus is "buy" with a mean price target of $282 and a range of $160 to $355. The low end of sell-side estimates — $160 — still sits 21% above the current price.

The options market is extreme: ATM implied volatility at 123.7%, indicating the market expects further large moves. The put-to-call open interest ratio of 1.96 is heavily protective. The put wall at $120 and call wall at $200 frame the battleground. Short gamma means dealers amplify moves in either direction.

The NVIDIA revenue-to-capex correlation — NVDA's datacenter revenue leads hyperscaler capex by roughly three quarters, with a peak correlation of +0.37 — suggests hyperscaler spending continues rising through at least early 2027. Memory is the physical bottleneck, and SK Hynix is the bottleneck's gatekeeper.

What Could Still Go Wrong

Two risks demand honesty. First, if the pricing-realization gap widens in Q3 rather than narrowing as contracts reset, the deferred-revenue thesis falls apart. Second, the rotation out of AI chip stocks into less-loved sectors has momentum — MU fell 8.9% in the same session — and that rotation does not need a fundamental trigger to continue extracting capital from the memory complex.

The quarter was a Rorschach test. Bears see a consensus miss and a memory cycle rolling over. Bulls see 76% margins, HBM4 ramping, ₩69 trillion of net cash, LTAs locking in a decade of AI memory demand, and a forward multiple that prices in a recession the physical data does not show.

The market chose the bear reading tonight. The tape will tell us whether the bulls were early or wrong.

More research at bargo.ai/research.