A record quarter that cost shareholders half their gains

On paper this should have been a party. The second-quarter numbers SK Hynix published on July 29 read like a typo: revenue of KRW 79.3187 trillion, operating profit of KRW 60.5426 trillion, and an operating margin of 76%. A year earlier the same quarter produced KRW 9.2129 trillion in operating profit, so profit grew 557% — more than sixfold. Revenue grew 257% from KRW 22.232 trillion. A 76% quarterly operating margin is a number neither Nvidia (65.6%) nor TSMC (60.3%) has ever printed. A memory manufacturer out-marginning the entire logic industry is close to unprecedented in semiconductor history.

Then the stock fell more than 15% intraday and closed down 9.61%. Here's the deal: analysts had penciled in KRW 64 trillion of operating profit and KRW 84 trillion of revenue. A record quarter that misses. That roughly KRW 4 trillion gap turned out to be the first domino.

And it kept falling. By CNBC's tally, more than $1 trillion came off the market caps of the world's most valuable chip stocks this week. SK Hynix lost $176 billion, Samsung Electronics $173 billion, and Micron $113 billion; Nvidia, AMD and TSMC each shed over $100 billion. Korea's KOSPI closed July 29 down 5.99% at 5,663.24, capping a 12.6% two-day drop — the worst in the index's history. The index is down 33.10% for July alone. And yet it is still up 34.44% year to date, which tells you what actually broke here. It wasn't earnings. It was expectations.

One more thing landed in the same week. Two days before the earnings release, reports said a Chinese state-owned company had entered mass production of homegrown immersion DUV lithography machines. The two load-bearing assumptions of the AI memory supercycle — demand keeps compounding, and China can't catch up — got shaken inside the same 48 hours. This piece is an attempt to lay out exactly what numbers came out, what's verified, and what is still guesswork.

Four players, including a Shanghai company nobody had heard of

SK Hynix is the biggest single winner of this cycle. It got to high-bandwidth memory first as Nvidia's primary supplier, and that premium is now printed straight onto the income statement. The company started HBM4 mass shipments in Q2, saying HBM4 "achieved customer-required operating speeds, industry-leading power efficiency, and cost competitiveness." HBM4E samples have already gone out. DRAM accounted for 73% of Q2 revenue. Cash and equivalents sit at KRW 88 trillion, up KRW 33.6 trillion from Q1; against KRW 18.6 trillion of total debt, that's a net cash position of KRW 69.4 trillion. For a company that spent years being described as balance-sheet constrained, that is a different animal entirely.

Samsung Electronics has the more absurd position in this story. Its July 7 preliminary Q2 figures showed roughly KRW 171 trillion of revenue and operating profit of about KRW 89.4 trillion (near $59 billion) — up around 19x year over year and 6.2% above the KRW 84.16 trillion Yonhap Infomax consensus, comfortably clear of the KRW 86 trillion WiseReport number too. Multiple outlets called it the largest quarterly operating profit any technology company has ever reported, ahead of anything Nvidia or Apple has posted. Revenue did land slightly under the KRW 172.18 trillion consensus, and the stock fell about 7% anyway — the explanation at the time being that the earnings were already in the price, while capex intensity and the odds of slower AI infrastructure spending by U.S. hyperscalers weren't. It dropped another 5.23% on July 29. The beat didn't work either — that's the signal worth sitting with. Samsung began the industry's first commercial HBM4 sales for Nvidia's Vera Rubin platform in Q1 and shipped HBM4E samples in Q2. Nothing in its execution story explains the selling.

Micron is the American proxy. Its most recent quarter produced record DRAM revenue of $31.3 billion, up 343% year over year. The stock is down more than 30% from the roughly $1,200 high it touched in June, closing at $776.13 on July 29. CEO Sanjay Mehrotra has been locking in five-year long-term supply agreements, with long-term DRAM contract pricing running from the "low teens to mid-$20s a gigabyte," and argues gross margins at the floor of this cycle will land well beyond previous cycle peaks. That's a strong claim. The market is declining to pay for it in advance.

Shanghai Aishengna Electronic Technology Group is the name almost nobody knew a week ago. It's a state-owned entity incorporated in August 2023 with RMB 7 billion of registered capital, and it absorbed teams from lithography startup Yuliangsheng — an affiliate of Huawei-backed equipment maker SiCarrier — and from Shanghai Micro Electronics Equipment (SMEE). The plan reported by The Information and confirmed to Reuters by a source: about five machines this year and roughly 20 in 2027, delivered to SMIC, Hua Hong Semiconductor and CXMT. ASML, SMIC, Hua Hong, CXMT, Aishengna, its shareholders, SMEE and Yuliangsheng all declined to comment or did not respond.

Then there's the offstage character: SoftBank. Treated as a levered proxy for AI hardware through its Arm stake, it fell 6.95% in this week's rout. The shares hit ¥4,595, the lowest since April and roughly half the year's high, with Arm down about 46% from its own peak. When an AI trade unwinds, the proxies tear first.

The numbers, pulled apart

Start with the figure that gets misread most. The headline this quarter isn't operating profit — it's net profit of KRW 93.9226 trillion, a 118% net margin. Net income larger than revenue is not a normal thing. The cause is KRW 62.17 trillion of non-operating gains: KRW 63.27 trillion in investment-asset gains plus KRW 1.15 trillion of foreign exchange gains, and most of the investment line is roughly KRW 60 trillion from the sale and revaluation of the company's stake in Japan's Kioxia. The original position cost about KRW 3.9 trillion back in 2018. Pretax profit came to KRW 122.7083 trillion, a record for any Korean company.

That's a spectacular trade, and it also muddies the quality of the print. The chip business generated KRW 60.54 trillion; the other ~KRW 33 trillion is an eight-year-old bet finally settling. It does not repeat next quarter. Part of why "record net profit" drew no cheers is right there.

Guidance was ambiguous in a way that mattered. Q3 DRAM bit growth is guided to roughly +10% quarter over quarter, NAND to a low single-digit percentage. Capex for 2026 is targeted at the high KRW 40 trillion range. The company is pulling forward mass production at Cheongju M15X, opening the Yongin Phase 1 cleanroom in early 2027, and lining up the P&T7 advanced packaging line, an M17 NAND base and a new semiconductor cluster. Its 321-layer NAND already carries the largest production share and is slated to reach about 50% of domestic capacity by year-end. Song Hyeon-jong, president of SK Hynix's Corporate Center, put it plainly: "To meet growing chip demand, we will move up the mass production schedule at Cheongju M15X and enhance investment efficiency."

Here's where readings split. Management framed the pull-forward as confidence in demand. A chunk of investors read it as building capacity into a peak — in memory, rising capex has historically been a late-cycle tell. On the call, management pushed back directly on the slowdown thesis, saying it views more efficient AI models "not as a sign of AI investment slowdown, but more as a transition towards higher utilization," and stressed that HBM pricing isn't set off conventional DRAM spot rates but negotiated on technical complexity, opportunity cost and delivered value. 2027 HBM price negotiations are live right now.

Item Figure
2Q26 revenue KRW 79.3187T (+51% QoQ, +257% YoY)
2Q26 operating profit KRW 60.5426T (+557% YoY)
Operating margin 76%
Net profit KRW 93.9226T (118% net margin)
Pretax profit KRW 122.7083T
Investment-asset gains KRW 63.27T (Kioxia ~KRW 60T)
Consensus KRW 64T operating profit / KRW 84T revenue
H1 2026 revenue KRW 131.895T
Net cash KRW 69.4T (KRW 88T cash − KRW 18.6T debt)
2026 capex plan High KRW 40T range
DRAM share of revenue 73%
3Q26 bit growth guide DRAM ~+10% QoQ, NAND low single digit
Long-term agreements ~10 customers, typically 5 years
Stock on results day −9.61% (−15%+ intraday)
KOSPI, July 29 −5.99% to 5,663.24 (−12.6% over two days)

Who cashed in, and the question nobody's asking

The clear winner is SK Hynix's balance sheet. KRW 69.4 trillion of net cash and KRW 122.7 trillion of pretax profit means the company can fund a high-KRW-40-trillion capex year without borrowing a won. The SK Hynix that had to cut spending in 2019 and the one reporting today are financially different organisms. Whatever the stock does, the company bought itself downcycle survivability this quarter.

Customers won too. The long-term agreement structure — about 10 customers, typically five-year terms, with price-volatility smoothing mechanisms and deposits attached — locks in five years of memory allocation for large buyers including Nvidia while giving the supplier demand visibility. Micron is doing the same thing from the other direction. Both sides are essentially buying the ability to plan regardless of when prices roll over.

China's equipment ecosystem got something free: perception. Five machines of early production moved half a trillion dollars of semiconductor market cap in a day. Benzinga's tally for the July 28 China-fear session: Nvidia roughly $300 billion, SK Hynix ~$95 billion, Micron ~$82 billion, SanDisk ~$26 billion, Western Digital ~$21 billion, Seagate ~$17 billion — about $541 billion across six names.

Now the uncomfortable part. Is this selloff actually about China, or about a consensus miss? The timing doesn't support either cleanly. Memory names started breaking down over seven trading sessions between June 25 and July 8, a stretch in which the semiconductor sector lost roughly $1.5 trillion, with Micron alone shedding nearly $350 billion. Twenty-five semiconductor names fell at least 20%, and Micron, Samsung, SK Hynix and the Roundhill Memory ETF all entered technical bear markets. But over that same window, memory stocks were still up a median of about 60% since late March, having added roughly $5 trillion in value. So what's happening is $1 trillion to $1.5 trillion coming out of a $5 trillion build. Forrester VP and principal analyst Charlie Dai described the move as "less about weakening AI demand and more about a repricing of expectations after an exceptionally strong rally." A collapse narrative doesn't fit those numbers. But "nothing to see here" doesn't fit either, because the market sold Samsung's beat and Hynix's miss with equal enthusiasm — and no earnings-based story explains that.

2019 happened. So did January 27, 2025.

Memory is a cycle business, and this scene has two precedents pointing in opposite directions: one where the bears were right, one where they were badly wrong.

The failure case is 2018 into 2019. SK Hynix's operating margin was 52% in 2018 — lower than today's 76%, but unheard of at the time, and the same "memory is structurally different now" arguments were in circulation. By 2019 that margin was 10%. Second-quarter 2019 operating profit fell 89% year over year, the third consecutive quarterly decline. The 2022–2023 cycle went further, producing the company's first operating loss in a decade and forcing spending cuts. What makes this precedent scary is that demand never vanished overnight. Inventory built up, customers slowed purchasing, and prices broke first. In memory, price tracks the spare margin between supply and demand, not demand itself.

The success case is January 27, 2025. After DeepSeek's R1 showed Western-class performance at a fraction of the cost, Nvidia fell 17% in a day and lost roughly $589 billion in market cap — the largest single-day value destruction in U.S. market history, more than double the prior record of $279 billion set on September 3, 2024. The conclusion at the time was that cheap Chinese models meant weaker AI hardware demand. What actually happened over the following eighteen months was an acceleration of AI infrastructure spending. SK Hynix's KRW 60 trillion quarter is the invoice for that acceleration. So the inference "China tech headline, therefore hardware demand collapse" has already been wrong once, recently, and expensively.

The way to tell the two scenarios apart is to watch different variables. A 2019 repeat requires prices and inventory to turn first. A 2025 repeat requires capex and contracted volume to keep rising. Right now the data leans toward the second: SK Hynix is raising capex and locking five-year LTAs, and Micron has fixed long-term pricing. But there's a signal on the other side too — CoreWeave, one of the largest AI-cloud memory buyers, is reportedly exploring financial hedges against a future decline in memory costs. When sellers fix prices for five years and buyers shop for downside hedges, both sides think we're near a top.

Every rival is running a different calculation

Samsung's counterplay is already in motion. To flip the narrative that it lost HBM4, it went at the problem through base dies built on its 4nm logic process, and started the first commercial HBM4 sales for Nvidia's Vera Rubin platform in Q1. Plenty of analysts attribute the 19x profit jump largely to that ramp. Samsung's structural weapon is owning memory, foundry and packaging under one roof — an advantage that grows as logic dies become the differentiator in HBM4 and beyond. But the fact that a 6% beat still produced a 7% drop says Samsung's problem isn't execution, it's the valuation regime. You can't fix that by shipping better parts.

Micron's answer is to contract its way out of cycle risk, arguing that trough margins this time will exceed prior peaks. That thesis needs one real downcycle to be validated, and the market is not waiting — it took the stock down over 30% from June. Micron also sits closest to China's line of fire: if CXMT ramps commodity DRAM faster than expected, price pressure lands disproportionately on the conventional bit business, where Micron's mix is relatively heavier than a pure HBM story.

CXMT is the real long-duration variable here. It is building toward capacity of more than 600,000 wafers per month, and analysts think it could surpass Micron's capacity around 2030. It is also roughly two generations behind in HBM, and U.S. export controls still restrict access to leading-edge lithography. Counterpoint Research's MS Hwang made the key distinction: Chinese progress will "have a greater impact on commodity DRAM and NAND" than on HBM, where incumbents hold a technology lead. That maps onto this week's tape, where storage names like SanDisk (−9%) took proportionally harder hits than the HBM leader did.

ASML looks like the most directly threatened party and, on the numbers, isn't — at least not yet. China's share of ASML revenue has fallen from 41% in 2024 to 33% in 2025 and 16% in the first half of 2026. Break out Q2 net system sales by destination and it reads South Korea 43%, Taiwan 30%, China 14%, the U.S. 9% — China is already the third market, and that decoupling was well underway before any mass-production report. The volume gap is still enormous: ASML is on track for roughly 130 units in 2026 against China's five, and even if China hits 20 in 2027 against ASML's ~169, the gap stays around eight to nine times. The Chinese tool reportedly handles 28nm-class single exposure and can be pushed to 7nm-class through multi-patterning, but industry assessments consistently call it a generation behind on throughput, overlay accuracy and long-term reliability — and no yield or defect data has been published. ASML's order book is sold out through 2027.

Nvidia occupies the strangest seat. Rising memory prices raise its bill of materials; falling memory prices get read as an AI demand warning. Neither direction is good for the stock, which is part of why its absolute dollar losses were the largest this week.

What actually changes for you

If you build things, not much shifts this quarter. HBM4 ramping and DRAM bits growing about 10% sequentially points toward more memory bandwidth per training and inference dollar from 2027 onward. In the near term, though, server DRAM and high-capacity SSDs are still near historic highs, so a self-hosted GPU cluster quote will hurt. SK Hynix said enterprise SSD revenue doubled quarter over quarter and high-capacity SSDs tripled — meaning the market, not the vendor, is absorbing every bit of output.

If you invest, this week was a valuation lesson delivered at speed. Samsung beat by 6% and fell 7%; SK Hynix missed by 6% and fell 9.6%. Both sold the same direction, which means the market isn't pricing individual results right now. Without your own framework for why you own something, a tape like this punishes you through volatility rather than direction. Worth knowing too: Samsung and SK Hynix together account for more than half the KOSPI's weighting, so the index has effectively become a direct proxy for global AI hardware sentiment. Diversifying across Korean large caps may just be the same bet, repeated.

If you run procurement, revisit your plan. The fact that suppliers are attaching deposits and price-smoothing mechanisms to five-year agreements means large buyers get supply certainty while smaller ones ride the spot market. The inverse risk is real: if memory prices genuinely roll over after 2027, signing long-term contracts at today's levels is the worst available choice. That's exactly the dilemma CoreWeave's reported hedging implies.

If you're just a consumer, the effect arrives late and indirectly. Today's memory prices are already baked into PC, phone and SSD pricing, and this week's equity selloff won't lower a single retail price. If anything, with SK Hynix converting roughly half its domestic NAND capacity to 321-layer product and prioritizing DRAM allocation toward high-value AI server parts, consumer volumes stay at the back of the queue. Meanwhile, a 33% monthly drop in the KOSPI quietly shows up in Korean pension and retirement accounts belonging to people who never bought a single share.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? Not much directly. But if you hold a Korean index product or a pension exposed to it, a 33% monthly KOSPI decline is already in your account. If you're buying AI servers or high-capacity storage, get fresh quotes before you commit.

— If the results were a record, why did everyone sell? The consensus miss was the trigger, but probably not the whole story. Around KRW 33 trillion of the KRW 94 trillion net profit came from the Kioxia disposal, and raising capex into the high KRW 40 trillion range read to some as a late-cycle move. Then again, Samsung beat and got sold too — so earnings alone don't explain it.

— If China can build DUV tools now, is the game over? Too early to call. The plan is five units this year and about 20 in 2027 against ASML's 130 to 169 per year, and the consensus read is that the Chinese tool is a generation behind on throughput, overlay and reliability, with no yield data disclosed. The more defensible view is that it becomes real pressure on commodity DRAM and NAND pricing within a few years, not on HBM tomorrow.

References

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!