The KOSPI lost a fifth of its value in four sessions. In the same four sessions, the companies at the centre of the panic printed the best numbers they have ever reported.
On 24 July the KOSPI fell 5.7%. On the 28th it fell 10.8% — the worst single session in its history, with 36 trillion won of selling. On the 29th, another 6%. Tokyo moved with it: Kioxia dropped 13.9% in a day and sits 60% below its peak, and Tokyo Electron lost 10.6% in one session. Prediction markets put the odds of an AI bubble bursting this year above 30%.
So which is wrong — the tape, or the earnings?
The bearish case is not imaginary, and the data supports parts of it. Korean semiconductors really did fall 13.9% on the 28th. Memory names really are the most US-coupled sector we track: Micron's move explains about 16% of SK Hynix's next session, and that coupling has tightened, not loosened — the 250-day sensitivity now runs 0.73 against 0.47 across the full sample. When New York sells semiconductors, Seoul follows harder than it used to.
But the crash and the earnings are not describing the same layer of the market — and only one of those layers is confirmed.
The crash had structure
We score how far each Korean sector moves from what its own six-year relationship with the index would predict. On 28 July, against a sector beta near 1.2, a −10.8% index day implied a 13.2% fall in semiconductors. The actual move undershot that by 0.7 points. The largest crash in KOSPI history moved its most-watched sector almost exactly as much as the arithmetic required — mechanical de-risking, not a reassessment of the memory business.
The reassessment showed up elsewhere. Healthcare beat its expected move by 5.2 points on the 24th and 4.9 points on the 28th: two crashes, two defences, the only sector that decoupled both times. Whatever the market was selling, it was not selling that.
Meanwhile the equipment makers answered a different question. Advantest shipped 48.6% of the quarter to China and 18.9% to Taiwan — two-thirds to Greater China, against 18.3% for the Americas. Tokyo Electron's split says it in yen: China 190.7bn, Korea 173.1bn, Taiwan 156.5bn, against 56.7bn for North America. Greater China runs six times the Americas, and Korea — the market everyone treats as a proxy for the US cycle — buys three times what North America does.
Then, yesterday morning, Japan's customs office published the June detail. The first half of 2026 did not merely repeat the pattern; it set a high. Greater China took 66.4% of Japan's chip exports — Taiwan 30.8%, China 21.9%, Hong Kong 13.7%. The United States took roughly 2%, as it has every year since 2021.
This morning the KOSPI reversed hard upward as Korean holders sold dollar bonds and converted back to won. Tokyo Electron opened 9.2% higher, two sessions after collapsing 10.6%. And at 13:15 the Bank of Japan held its policy rate at 1.0% — the highest since September 1995 — while raising its fiscal 2026 growth forecast from 0.5% to 0.6%, citing strong AI-related demand. Three board members voted for a further hike.
Markets priced an AI bubble bursting. The central bank raised growth because of AI demand. The equipment makers raised guidance into the selloff.
What we publish: the measured relationships, the deviation scores, and the failure condition attached to each. What we do not publish: the estimation parameters, the source combinations behind them, or any position. If the next leg down arrives with guidance cuts rather than multiple compression, the divergence reading here is wrong and the tape was early. Kioxia reports 6 August; Japan's next customs release lands 28 August. Both dates are in the calendar we serve, and both will be scored here.
Data: Trade Statistics of Japan (Ministry of Finance / e-Stat), Government of Japan Standard Terms of Use; Korea Customs Service and Korea public data portal; company filings (Advantest, Tokyo Electron, SK Hynix); Bank of Japan. Findings may be cited with attribution to Asia Pulse (asia-pulse.com).