Limit-up on Friday, minus 8.8% on Monday

Here's the deal. On Friday July 31, the Korean stock market did something it had never done. The KOSPI closed at 6,595.45, up 1,001.89 points — a 17.91% single-day gain. Both the point move and the percentage move were all-time records. SK Hynix finished limit-up at 1,718,000 won, a 29.95% jump and its first limit-up since 2009, seventeen years earlier. Samsung Electronics rose 26.81% to 262,500 won, reclaiming a spot in the trillion-dollar club at a $1.212 trillion market cap and climbing two places to 12th in the global rankings. Foreign investors bought a record 8.7709 trillion won net. Retail investors sold a record 10.3834 trillion won net. SK Group chairman Chey Tae-won reportedly booked roughly 1.3 billion won in paper gains overnight on 3,620 SK Hynix shares he had picked up the day before.

Three days later, the market opened and immediately went the other way. On Monday August 3 the KOSPI gapped down to 6,358.27 — already 237 points and 3.60% lower before anything happened. It sank to 6,247.64 intraday and closed at 6,257.45, down 338.00 points or 5.12%. Samsung dropped 23,000 won (8.76%) to 239,500. SK Hynix dropped 151,000 won (8.79%) to 1,567,000. Roughly half of Friday's gain evaporated in a single session.

What makes this strange is that the overnight tape from the United States was good. Amazon surged 15.32%. Microsoft rose 3.02%. The Dow added 276.97 points (0.53%) and the Nasdaq added 251.68 points (1.00%). Those are the companies pouring money into AI infrastructure, and the companies collecting that money are Korean memory makers. The tailwind simply did not cross the Pacific. MoneyToday attributed the disconnect to profit-taking colliding with fears of a yen carry trade unwind.

And there's a bigger oddity underneath. Nothing broke operationally. Five days earlier, on July 29, SK Hynix reported a quarterly operating profit of 60 trillion won. The next day, July 30, Samsung reported 89 trillion won. Both were all-time records by a wide margin. The companies are earning more money than they ever have, and the stocks are down roughly half from their peaks. This piece is about where that gap comes from, what "yen carry unwind" actually means in mechanical terms, and what KB Securities is actually leaning on when it calls the current price an excessive discount.

Who is standing on this floor

Start with Samsung and SK Hynix themselves, because they are no longer just large caps. Together they account for roughly half of the KOSPI's total market capitalization. That single fact does most of the explaining. When two names that carry half the index each fall 8%, the index mechanically falls more than 4%. The 5.12% drop on August 3 wasn't a mysterious event — it was close to arithmetic. Buying Korea, whether through an index fund, a pension allocation, or an offshore ETF, has become almost indistinguishable from buying two memory chip stocks.

Second, foreign investors. Foreign ownership sits at 46.70% of Samsung and 51.22% of SK Hynix — meaning overseas holders already own a majority of SK Hynix. On August 3, foreigners net sold 991.4 billion won and institutions net sold 77.9 billion won, while retail investors net bought 1.0420 trillion won. That is the exact mirror image of the previous session, when foreigners bought and retail sold. The entire direction of flow flipped inside two trading days. No fundamental story explains a reversal that fast.

Third, and this one is less familiar outside Korea: single-stock leveraged ETFs. These products first listed domestically on May 27, 2026. Within fifty days their combined market cap ballooned from 4.4 trillion won to 11.9 trillion won and they accounted for roughly 38% of all ETF trading value in the country. Retail money piled into 2x products tracking Samsung and SK Hynix specifically. The structural problem is that leverage amplifies declines — falling prices trigger stop-outs and forced unwinds, which feed back into selling of the underlying shares. On July 16 regulators announced an indefinite halt on new listings and raised the minimum account deposit from 10 million to 30 million won, cash only, then pulled the effective date forward from August to July 31.

Fourth, the analysts, who are openly split. Kim Dong-won, head of research at KB Securities, kept buy ratings and top-pick status on both names in an August 3 note. His argument: "As of August 2026, big tech customers' memory and AI substrate demand is only about 60% fulfilled." Unfilled orders roll into the following year, and that rollover cascades like dominoes for three years. He added that "completing a new memory fab and running it at full capacity takes more than three years," which pushes any meaningful relief past 2029. Kyobo Securities took the other side, arguing that what ultimately matters is whether big tech's cloud and advertising businesses generate real profit — and that AI demand ends when investors stop funding it. Same data, opposite conclusions.

Fifth, an institution that never appears by name in the headline but stands behind the phrase "yen carry unwind": the Bank of Japan. The BOJ raised its policy rate from 0.75% to 1.0% in June 2026, the highest level since 1995 — a 31-year high. At its July 31, 2026 meeting it held at 1.0%, but the vote was 8-1. Board member Takata Hajime dissented, arguing the situation had shifted into a new phase requiring a nimble response to upside price risks from overseas demand shocks and to changes in overseas financial conditions. A hold with a hawkish dissent reads, to markets, as a hike moving closer.

What actually happened inside that 8%

You can't read August 3 without reading July first. The KOSPI entered the month at 8,476.48 and fell to 6,023.66 by July 28 — down 28.94% in under a month. That exceeded October 1997 during the Asian financial crisis (-27.25%) and October 2008 during the global financial crisis (-23.13%), making it the largest monthly drawdown since 1990. On July 28 and again on July 29, circuit breakers fired simultaneously on both the main board and KOSDAQ — two consecutive days, the first time that has ever happened. After the July 31 rebound the final monthly figure landed at roughly -22%, still the worst among major markets. This was not a global AI correction that Korea participated in. It was a Korea-specific event.

The July 31 explosion came off that floor, and several things stacked at once. Chip stocks had been strong in New York overnight. Three straight days of collapse had produced enormous bargain-hunting demand. And the leveraged ETF deposit rule took effect that very day, removing one of the mechanical volatility amplifiers — leveraged product turnover reportedly collapsed to about a quarter of its prior level. There was also market chatter that a large US hedge fund had finished liquidating positions that had been weighing on chip supply-demand, though that was never officially confirmed and should be treated as speculation.

Then Monday arrived, and the rally's own physics took over. Three forces made up the 5.12% drop. First, ordinary profit-taking: selling a stock that rose 27-30% on Friday is not panic, it's discipline, and when a buyer who absorbed 8.77 trillion won of stock in a single session sells even a fraction back, the tape moves. Second, the yen carry fear. Third, index concentration — with two names at half the market cap, an 8% move in those names cannot produce anything other than a 5% move in the index.

So what is a yen carry trade, mechanically? Borrow yen, which costs almost nothing, and put the money into assets that yield or appreciate more. Korean chip stocks were one of those assets. The trade needs two conditions to hold: yen funding stays cheap, and the yen doesn't strengthen. A BOJ hike breaks the first, and the resulting yen appreciation breaks the second. When both break, borrowers have to sell foreign assets to repay yen loans — and the selling order is always "whatever went up the most, and whatever is easiest to sell." Large-cap Korean semiconductors satisfy both criteria perfectly. They get sold not because something is wrong with them, but because they are liquid and crowded. That is the first reason price and fundamentals can walk in opposite directions.

Metric 2026-08-03 Previous session (July 31)
KOSPI 6,257.45 (-338.00 pts, -5.12%) 6,595.45 (+1,001.89 pts, +17.91%, all-time record)
KOSPI open / intraday low 6,358.27 (-3.60%) / 6,247.64 —
Samsung Electronics 239,500 won (-23,000, -8.76%) 262,500 won (+26.81%)
SK Hynix 1,567,000 won (-151,000, -8.79%) 1,718,000 won (+29.95%, first limit-up since 2009)
Foreign flow 991.4bn won net sell 8.7709tn won net buy (all-time record)
Retail / institutions Retail +1.0420tn buy / institutions -77.9bn sell Retail -10.3834tn sell (all-time record)
Foreign ownership Samsung 46.70% / SK Hynix 51.22% —
Overnight US session Amazon +15.32%, Microsoft +3.02%, Nasdaq +1.00% —
SK Hynix Q2 2026 Revenue 79.3187tn / operating profit 60.5426tn (76% margin) +256.8% revenue, +557.2% operating profit YoY
Samsung Q2 2026 Revenue 171.4995tn / operating profit 89.4924tn (52.2% margin) Chip division = 99.7% of operating profit
Valuation (KB Securities) Samsung 2027E PER 3.8x / SK Hynix 3.6x Targets 600,000 won / 4,200,000 won, Buy
July KOSPI -22% for the month (-28.94% at the July 28 low) Worst monthly drop since 1990

The row that should stop you is the margin line. SK Hynix ran a 76% operating margin in the second quarter. That is software-company economics inside a capital-intensive fab business, and it is essentially unprecedented in the industry. First-half cumulative revenue reached 131.895 trillion won, crossing 100 trillion for the first time, with 88 trillion won in cash and 69.4 trillion won in net cash on the balance sheet. At Samsung, the Device Solutions division delivered 99.7% of total operating profit — 89.2 trillion won on 127.5 trillion won of revenue. Phones and appliances contributed close to nothing. Samsung has, for accounting purposes, become a memory company.

And yet SK Hynix fell hard on the day it printed those numbers. The reason was shareholder returns. Sitting on 69 trillion won of net cash, the company gave no concrete dividend or buyback plan on the earnings call, and instead emphasized "capex discipline" while pulling forward M15X and preparing the Yongin fab for 2027. Investors read that as: record profits, none of it comes to you, all of it goes back into fabs. In a cyclical industry that message is genuinely frightening, because everyone knows the fab you break ground on today is the oversupply you live through three years from now.

Who gets what out of a crash like this

The most obvious winners, on paper, are the retail investors who bought on August 3 — 1.0420 trillion won net. But you have to read that number alongside the July 31 record: retail net sold 10.3834 trillion won on the day the index gained 17.91%. Retail sold the melt-up and bought the drop. Sequenced that way it looks like the worst possible timing, but it isn't stupidity so much as structure. People who had been trapped for two months finally got an exit on the spike, and a different cohort watching from the sidelines used the pullback as an entry. That pattern repeats in every violent market.

Foreign investors hold the strongest hand here. If you bought 8.7709 trillion won of stock on Friday and sold 991.4 billion won on Monday, you still own the overwhelming majority of what you bought, and you just watched your own selling improve the price at which you could add more. Only that weight class can move a market and manage its own average cost at the same time. Ownership levels of 46.70% and 51.22% mean this group can push the index in either direction essentially whenever it chooses to.

Memory buyers — the hyperscalers and server OEMs — gain a psychological negotiating tool. Contract pricing is always a standoff between "you need us" and "we have nowhere else to sell," and a supplier whose stock has halved looks weaker across the table. Except that if KB's 60% fulfillment figure is anywhere close to right, the tool is blunt. You cannot squeeze pricing on a part you can't get enough of. Whether this correction shows up in actual contract prices over the next two quarters is the real litmus test for whether August 3 was flow or fundamentals.

The clear losers are the leveraged product holders. In a 2x product, an 8.8% move in the underlying is more than 17% gone in a day. Worse, these instruments bleed through volatility decay: when the underlying whipsaws violently and returns to where it started, the product does not. Anyone holding leveraged Samsung or SK Hynix exposure through July 31 and August 3 likely got hurt on both legs. That is exactly why regulators moved, and it is also why they were accused of showing up late.

The bill also lands on the broader Korean economy. Estimates put the market cap destroyed between late June and the July trough at roughly 2,800 trillion won. Losses of that scale flow through consumption, tax revenue, pension returns, and corporate funding costs in sequence. In a market this concentrated, the memory cycle effectively becomes the national household-wealth cycle. The deepest policy lesson from this episode isn't about leverage products at all — it's about the concentration itself.

We've seen versions of this before

The nearest precedent is August 5, 2024. The BOJ hiked on July 31 that year, the yen carry trade unwound violently, and the following Monday the Nikkei 225 fell 12.40% to 31,458.42, its second-worst day since Black Monday in 1987. The KOSPI closed down 8.77% at 2,441.55 — the largest point drop in its history and the fourth-largest percentage drop, with a circuit breaker at 2:14 p.m. What matters is the ending. No company's earnings were damaged by that event, and the indices recovered most of the ground within weeks. Crashes caused purely by capital relocation tend to reverse quickly. The reason "yen carry unwind" appeared in Korean coverage on August 3, 2026 is that the market pulled that exact memory off the shelf.

March 2020 is the second case in the same family. COVID drove the KOSPI to 1,457.64 on March 19, and even though the real-economy shock was genuine, the asset-price drop massively overshot it. Liquidity returned and the index finished that year at 2,873.47 — roughly a double in nine months. The takeaway isn't optimism, it's a condition. Both of those recoveries share one trait: forward earnings never rolled over. Crashes where the profit outlook holds get retraced. Crashes where the profit outlook breaks do not. That is close to the only discriminator that works.

Now the failure case, and it's the one that should worry anyone quoting a low PER right now. The 2018 memory supercycle. Samsung posted a then-record 58.89 trillion won of annual operating profit that year and SK Hynix a record 20.8 trillion won. Samsung's stock peaked on May 23, 2018 — and both companies' operating profits didn't peak until the third quarter. The share price turned before the earnings did. DRAM prices halved in 2019 and profits roughly halved with them. As Newsis laid out, SK Hynix traded at a historically cheap 4.6x PER in 2018 and then fell another 42%. Park Jong-hoon of the Institute for Knowledge Economy calls this "the PER trap": in a cyclical industry, peak earnings are not confirmation of value, they are the starting line for the market's estimate of decline. His summary line is the one to remember — the low PER is a fact, but what matters is whether profits keep growing.

The older failure is the 2000 dot-com bust. Telecom and semiconductor earnings were genuinely strong at the time, and the underlying technology thesis was not wrong — with twenty years of hindsight, if anything it was too conservative. The stocks still collapsed for years. "Is the technology real?" and "is this price right?" are entirely different questions, and each cycle relearns that the hard way. Ask whether AI memory demand is real today and the honest answer is probably yes. That still isn't an argument that 239,500 won is the right price. KB's three-year domino thesis and Kyobo's show-me-the-cloud-profits thesis diverge precisely at that seam.

How the other players counter-punch

Start with Micron. It rides the same cycle as the two Korean firms but trades in a different capital pool, which means it is far less exposed to Korea-specific variables like yen carry positioning or domestic leveraged ETF rules. That makes it a natural control group: if August 3 really was a Korean flow event, Micron should have held up materially better over the same stretch. If it fell just as hard, then what happened was a global re-rating of memory, and KB's "excessive discount" call gets a lot shakier.

China is the slower and more dangerous variable. Share gains by domestic DRAM makers such as CXMT are widely treated as the most realistic downside risk to this cycle. The technology gap in leading-edge products like HBM4 is still wide, but every additional bit of Chinese commodity DRAM supply pushes the price floor lower. That's why the Newsis analysis named CXMT market share, alongside DRAM spot pricing, as the two things to monitor. It moves far more slowly than a quarterly print, but once the direction sets, it doesn't easily reverse.

Hyperscalers have two counters. One is custom silicon. Google's TPUs and Amazon's Trainium don't use less memory — they change the negotiating structure, because buying memory directly instead of through Nvidia rewrites who captures the margin. The other is simply slowing down. One of the triggers of the early-July collapse in Korea was skepticism about AI investment returns, so a single softened capex guide from any of the big four could shake Korean memory names again. Cutting the other way, Amazon jumping more than 15% on the night of August 3 suggests that skepticism has lost its grip at least for this quarter.

Korean regulators have a move left too. The leveraged ETF deposit rule is in force, but the stated intent to close the loophole of buying equivalent products listed in the US or Hong Kong still has to be executed. Politicians across parties have floated a parliamentary inquiry into how these products were approved in the first place. Tighter rules will damp short-term volatility, but they also thin out turnover — and a thin market takes the next shock worse, not better. That is the dilemma sitting on the regulator's desk.

Finally, the Bank of Japan. That July 31 hold was not unanimous: 8-1, with the dissenter arguing for nimbler response to overseas-driven risks. If the next meeting produces a hike, the yen strengthens and the carry unwind pressure stops being a fear and becomes a flow. It sounds absurd that a Korean chip investor should be reading Japanese central bank minutes, but in a market where half the capitalization sits in two stocks and roughly half of those two stocks sits in foreign hands, that is just competent risk management.

What actually changes now

For an ordinary investor, the practical move is to separate what this selloff confirmed from what it didn't. Confirmed: both companies are earning more than at any point in their history, and August 3's drop happened with no earnings miss, no customer loss, no product failure attached to it. Not confirmed: how long those earnings last. KB says three years. Kyobo says it depends on whether big tech's cloud and advertising businesses actually pay for the buildout. Both are forecasts, not facts. So the sentence "record earnings and the stock is halved, therefore it's obviously cheap" is dangerous, and the counterexample is right there in the record — SK Hynix at 4.6x in 2018 fell another 42%. The thing to track is not the multiple, it's the direction of the earnings estimate.

For engineers and operators buying chips and building data centers, the signal points the opposite way from the chart. A 60% fulfillment rate means you order and receive 60% of it. SK Hynix started HBM4 mass-production shipments in Q2 and plans to ramp hard in the second half, with M15X pulled forward and Yongin queued for 2027 — none of which suggests slack. Practically, memory lead times and unit costs are likely to be a binding constraint on project schedules for the next year or two. If you're planning server expansion, lock in DRAM and HBM allocation as aggressively as you'd chase GPUs. When the price chart and the procurement reality disagree, procurement reality usually turns out to be the one describing the present.

For investors, the watchlist compresses to four items. One: DRAM spot pricing, which moves ahead of contract pricing and leads cycle turns. Two: next-quarter capex guidance from the hyperscalers. Three: whether foreign investors keep the 8.77 trillion won they bought on July 31 or distribute it over the following sessions. Four: the BOJ's next decision and the yen. If any one of those four turns, the "it's just flow" reading loses its expiry date. If all four hold, this drop ends up as a long lower wick on a chart nobody remembers by winter.

For policymakers, the homework is structural and unpleasant. A market where two companies are half the index is not safe no matter how well those two companies execute, because the memory cycle becomes the pension return, the retirement account, and the household balance sheet all at once. July producing the worst monthly drawdown since 1990 and the first-ever back-to-back dual circuit breakers was not really a story about two corporations — it was a story about concentration. Leveraged product rules treat a symptom. The actual prescription is broadening the listed base and deepening the market, and that is a multi-year project that does nothing for the current quarter.

And there's something here even for people who own no stocks at all. The market cap that vanished on August 3 passed straight through the National Pension Service and corporate retirement plans, which are wired into most Korean salaried workers' balance sheets whether they follow markets or not. That's why the KOSPI is front-page news rather than a business-section item. Whether a market in which two companies' quarterly results effectively set the slope of national wealth is a healthy market is a question that outlasts this particular crash.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? More than you'd think, even if you own zero shares. National pension and retirement plans hold the KOSPI, and half the KOSPI is these two stocks, so a day like August 3 shows up in your account whether you looked or not. That said, a single session is not a signal — what to watch is whether the earnings outlook rolls over.

— Why did this happen right now, of all moments? Because the previous session was the biggest one-day gain in KOSPI history, up 17.91%. Stocks that rise 27-30% in a day get sold the next morning; that's normal. Layer on renewed expectations of a Bank of Japan hike and the yen carry unwind fear that comes with it, and you get 8.8%. The key detail is that no new bad news — no guidance cut, no lost customer — landed that day.

— Is this the bottom, or is there more downside? Too early to call with any confidence. KB Securities points to a 60% demand fulfillment rate and shortages persisting to 2029 to argue that 3.6-3.8x 2027 earnings is an excessive discount. On the other side sits the 2018 record, when SK Hynix at 4.6x fell another 42%. In cyclicals a low multiple can be evidence of a peak in earnings rather than a floor in price. DRAM spot prices and hyperscaler capex will answer this, not valuation math.

Sources

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