Whether AI demand is real is a question the market has already answered, and it did not need to wait for SK Hynix’s results. What broke in Seoul on 13 July was not a verdict on that proposition. It was a positioning accident.

The Korea Composite Stock Price Index (KOSPI) closed that Monday down 8.95% at 6,806.93, having opened at 7,412 and touched 6,783. At 1:28pm the Korea Exchange tripped its circuit breaker for the seventh time this year and the thirteenth in its history. SK Hynix fell more than 15% to ₩1.845m; Samsung Electronics lost 10.7% to ₩254,000. The flow split cleanly. Foreign and institutional accounts sold a combined ₩3.9tn (about $2.6bn); retail bought ₩3.88tn, almost the mirror image. Individuals were averaging down.

The accident did not end in a day. The index rebounded 0.73% on the 14th and 6.24% on the 15th, then fell 6.37% again on the 16th to 6,820.60: two circuit-breaker-scale declines in one week. Measured from the closing peak of 9,114 on 22 June, the 13 July low sits 25% lower. None of that was earnings, or rates, or macro. It was five sessions of price doing this to itself.

The question a forced-flow read asks is a narrow one. Who has to act right now and in which direction? And has that changed? Count the buyers on the 13th and the answer is plain. Retail averaging down; foreigners round-tripping intraday. Nobody was compelled to buy. Every buyer was discretionary. So a good number could be sold two days later, and a broken market found no one reaching for the knife.

The one participant that moves mechanically is the single-stock leverage ETF. Listed on 27 May, these products were permitted only two underlyings, Samsung and SK Hynix (the only names clearing the thresholds of 10% of market value and 5% of turnover), and sixteen of them now crowd those two stocks. Their combined assets swelled from ₩4.4tn at listing to ₩14.9tn by 26 June, then shrank to ₩11.9tn (about $8bn) by 15 July after the selloff. Retail net buying since listing runs to ₩13.8tn (about $9bn), roughly 18% of all retail buying on the KOSPI over the period. Money left the junior KOSDAQ market, ₩2.4tn of it, and moved into leverage on two names. Those two now make up 52% of the KOSPI, against 49% in late May. Half the index moves on two memory stocks.

The product structure is what manufactures the forced flow. To hold twice the underlying’s daily return, the manager and its liquidity providers must buy into strength and sell into weakness, near the close, every day. Korea’s sell side has a name for it: long-short gamma. It amplifies in one direction only, and it sits on top of two stocks that are half the index. The more the market moves, the more the ETFs magnify that move into the close.

One popular explanation deserves to be dropped. Much of the commentary blamed forced margin liquidation. The primary data disagree. Settlement-shortfall reverse trades on the 13th came to ₩26.2bn (about $17m), concentrated in the pre-open auction. The larger forced sales had come earlier, before the crash: ₩142.2bn on the 9th alone, not on the 8.95% day itself. It was too small to drive the intraday move, and mistimed for it besides. The forced flow is the ETF’s daily rebalancing, not the margin desk. The retail habit of round-tripping between the leverage and inverse versions, half of the ten most-turned-over names and changing hands as often as two dozen times a day, widens the range but forces no one. It is discretionary noise.

So the map reads like this. What is forced is the ETF rebalancing, and it amplifies procyclically. One more forced event carries a date. From 5 August the base deposit rises to ₩30m (about $20,000) in cash; from 19 August, pledged securities no longer count. A retail account can hold any amount of stocks and bonds and still be unable to add without ₩30m in cash. The marginal buyer who has been averaging down meets a wall, and anyone who cannot post is deleveraged. Everything else is discretionary. With no forced buyer, no mechanical floor arrives.

The watershed falls on 29 July, when SK Hynix reports the second quarter. Consensus is around ₩64.8tn of operating profit (about $43bn) at a 76% margin; reach ₩70tn and the margin nears 83%, above the first quarter’s ₩37.6tn and 72%. The number, though, is half the problem. Earnings out to 2027 are already in the price, and Korea’s own sell side is braced for a good print to sell the news. What matters more than the figure is the third-quarter guidance: high-bandwidth memory (HBM) or commodity DRAM. SK Hynix carries the frontier (HBM) and the canary (commodity memory) in a single name. If the canary sings, it will sing from the guidance.

The volatility did not stay in Seoul. Korea’s Financial Services Commission extended its margin rules beyond domestic ETFs to overseas single-stock leverage products, on the reasoning that the world’s semiconductor names are moving increasingly together. The table it published makes the point. Annualise daily moves from 26 May to 10 July and you get SanDisk 131%, Micron 123%, Kioxia 118%, SK Hynix 113%, Samsung 96%. Japan’s Kioxia sits between Micron and SK Hynix. A Korean regulator drew a line to protect its own investors, and a Tokyo memory stock falls inside it. Kioxia no longer trades as a single Japanese name; it trades as a member of one semiconductor-volatility cohort.

The numbers point the same way. Over the life of the single-stock leverage products, Kioxia’s daily returns tracked the KOSPI at 0.82 and the American semiconductor ETF (SOXX) at only 0.51. Kioxia and the KOSPI topped on the same day, 22 June. The KOSPI works as a memory proxy because half of it is those two names. Japanese memory kept step with Korean memory more tightly than with US semiconductors at large, and Nihon Keizai Shimbun notes that the Nikkei is moving ever more closely with the KOSPI.

The honest caveat is about how that link works. On 17 July the Nikkei 225 closed down 4.03% at 64,141, and its weekly fall of 4,416 points was the largest on record; it was down as much as 4,130 points intraday. The direct trigger was overnight, in US semiconductors, with the SOXX off 4.46% on the 16th, alongside a reported jury award against Kioxia that took the stock limit-down. Not a same-session mechanical spillover from Korea. What crossed to Tokyo was the volatility regime, not the tape itself.

Here is the core of the forced-flow question for a Japanese reader. It is tempting to conclude the amplifier is Korea’s alone, since there is no single-stock 2x ETF on Kioxia or Advantest of the kind Korea listed on 27 May. That reads the mechanism too narrowly. Japan built the same procyclical leverage, not on the stock but on the index.

The Nikkei 225 is price-weighted, so the higher-priced a share, the more each yen of it moves the index. On the end-June fact sheet, technology is 58.85% of the average; Advantest at 11.14% and Tokyo Electron at 11.07% are 22.21% between them, and adding SoftBank Group and Kioxia reaches 32%. Include Shin-Etsu, TDK and Ibiden and it is close to two-fifths. That the price-weighted Nikkei rides on a handful of AI-semiconductor names is a point a companion piece here has made before. On top of this index sit leveraged ETFs, the 2x-daily Nikkei products, which rebalance through 225 futures into strength and weakness at the close. The same long-short gamma, worked at the index level. Since 60% of the index is technology and its top two names are semiconductors, index gamma still lands on AI chips. When Advantest moves, the index moves, and futures and arbitrage flows are parcelled back to the same handful of names; on the 17th, Advantest alone took 772 points off the Nikkei before lunch.

The second layer is margin debt. Outstanding margin buying on the Tokyo market, standardised and negotiable together, reached ¥7.017tn on 26 June (about $43bn), the first time above ¥7tn in records going back to 1994. Of that week’s ¥541bn increase, Kioxia alone was 60%. Its margin-buy ratio ran near 22 to one, an extreme long, and the average price of those purchases was the highest on record: retail had bought high, on borrowed money, which Nihon Keizai Shimbun flagged as fragility to a fall. Bloomberg warned on 3 July that if the price weakened, the positions could unwind at once. It did not wait for the crash. The balance had already fallen for two straight weeks to ¥6.732tn (about $42bn) by 10 July before Korea on the 13th and Kioxia on the 16th and 17th arrived. Kioxia dropped 15% on the 16th and went limit-down on the 17th, more than halving from its June peak. On the same 17th, Bain Capital was reported to have sold its entire Kioxia holding. Informed money stepped off; leveraged retail stayed on. Unlike Korea’s continuous, two-way ETF gamma, this margin layer is latent and one-way. It forces nothing on the way up, where the buying is discretionary leverage, and forces selling only on the way down, through margin calls into liquidation, with the exchange’s own escalation from daily disclosure to a higher deposit requirement able to bolt on a second layer of compulsion.

So the difference between Seoul and Tokyo is not that one has the amplifier and the other does not. Both built a procyclical leverage machine on the same memory cohort. Korea aimed it at the stocks, with a single-stock 2x ETF; Japan aimed it at the index, with a 2x index ETF. But when 60% of the index is technology and a third is four names, betting the index is very nearly betting the stocks. Korea did it explicitly, Japan by implication. There is an asymmetry inside the asymmetry, too. A margin overhang sits on both markets. Korea carries ₩34.8tn ($23bn) of margin loans; Japan’s ¥7tn ($43bn) balance is larger still, and 60% of its recent build was a single name bought at the highest price on record. The latent selling pressure is larger in Tokyo, and more concentrated.

The answer for a Japanese investor follows. There is no forced buyer anywhere; retail averaging down and retail day-trading are both discretionary, so no mechanical floor arrives, as with the recent AI-infrastructure names. But a latent forced seller waits. The Kioxia margin pile gives way to margin calls if the price falls further. Not a floor; a trapdoor. Two dates frame it. On 29 July SK Hynix says whether HBM pricing still holds and, in its third-quarter guidance – the canary – how firmly. On 31 July Kioxia reports in turn. The results will answer that; they will not tell you where the floor is, because there is none. There is a trapdoor.

The direction, then. Volatility from Korean memory is likely to keep reaching Japan’s AI and semiconductor names as correlated movement. But Japan owns its own leverage, index gamma and the margin pile, so the imported volatility can be amplified at home rather than merely absorbed. The forced flow is no longer Seoul’s alone. Sector direction is as far as this piece goes; selection and sizing are the reader’s own. None of this is investment advice.

Figures are drawn from the Korea Exchange, the Financial Services Commission, Japan Exchange Group and the Tokyo Stock Exchange, the Nikkei 225 monthly fact sheet (end-June), Nihon Keizai Shimbun, Kabutan, Bloomberg and FnGuide, with daily closes compiled by the author. SK Hynix’s second-quarter figures are FnGuide consensus. Won and yen are converted at roughly ₩1,500 and ¥160 to the dollar. The Bloomberg article is paywalled. Quarterly figures will be confirmed by SK Hynix on 29 July and Kioxia on 31 July.

— Gyokuro