On 17 July the Nikkei 225 closed at 64,141, down 2,694 points on the day. That is a fall of 4.03%. As a one-day drop it ranks fifth on record; as a week it was the largest ever, at 4,416 points. Semiconductors and the wider AI complex led the way down: Kioxia Holdings fell 16.1%, Ibiden 9.9%, Tokyo Electron 8.2% and Advantest 7.2%.
The headlines reached for a warning about the AI growth story. Line the day’s moves up in a row, though, and the story that was supposedly under threat is nowhere to be seen. What shows up is duller than that.
Take each stock’s fall and sort it by the stock’s beta: how much it moves, on average, when the market as a whole moves 1%. Beta is not a published figure, so it has to be estimated. Here it comes from regressing each stock’s daily returns on the Nikkei over the 250 trading days to 16 July.
Mitsubishi Heavy, on a beta of 0.69, fell 3.71%. Hitachi, at 0.83, dropped just 2.30%. Tokyo Electron, at 1.49, was down 8.17%; Advantest, at 2.02, 7.20%. And Kioxia, at 2.20, fell 16.10%.
Low betas fell least, high betas fell most, and the match is close to exact. Across the 10 stocks the rank correlation between beta and the size of the fall was -0.88, with less than a 0.1% chance of the ordering arising by accident. Drop the most volatile name, Kioxia, and it still holds at -0.83.
So what happened on 17 July was this. The Nikkei fell 4%. Each AI-related stock fell by its own beta: a beta-two stock by 8%, a beta-one stock by 4%. That is all.
＊
To fall by your beta is to be sold without discrimination. There is a way to test whether that is what happened.
Look at whether each stock fell further than its beta would predict, or held up better. Fall further and the market singled the stock out to punish; hold up and it singled it out to protect. Call the difference the excess return.
On 17 July not one of the 10 stocks had a statistically significant excess return. The furthest out, Kioxia, came in at a t-statistic of -1.50, short of the -2 that marks significance. Across the historical record, a typical day throws up fewer than one significant excess return among these 10 names, 0.65 on average, and the busiest days as many as six. On this day the count was zero. No stock was singled out to be punished, and none to be spared.
Hitachi held up best, at 2.30%, not because the market rewarded its grid backlog but because its beta is 0.83. It is a low-beta conglomerate with large non-AI businesses, and a low-beta stock had a low-beta day. There is nothing more in it than that.
Mitsubishi Heavy’s steadiness at 3.71% is the same story. One could read into it the tie-up with NVIDIA on data-centre work that the company had announced the day before. But its excess return was -0.83%, on a t-statistic of -0.37, well within what beta already explains. On 16 July, when the tie-up was reported, the shares moved 0.13%.
And Advantest was sold on the very day its market widened. On the same 16 July, TSMC raised its 2026 capital budget to $60bn-$64bn, up $8bn across the range from the $52bn-$56bn it set out at the start of the year. Management said in the same briefing that testers were in short supply and that it was putting extra capital into them. For a maker of testers this should have been a tailwind. The stock fell 7.20% anyway. Even so, that is 0.82% shallower than the 8.03% its beta of 2.02 would predict. It was not sold because it was disliked. It was sold by its beta.
＊
Why was nothing singled out?
Because the selling was not picking stocks. There was no single trigger to 17 July. The reported causes run to several: word from the American market the day before that Alphabet’s AI model development had slipped; higher oil on the sharpening attacks on Iran; a rise in US rates; and a decision by South Korea’s government to tighten rules on leveraged single-stock exchange-traded funds, which, with the Korean market shut, pushed pre-emptive selling into Japanese AI names.
The path into semiconductors was more specific. The day before, TSMC’s results beat expectations. But with the good news taken as already in the price, its American depositary receipts fell, and the SOX index of semiconductor stocks dropped more than 4%. Tokyo simply inherited that. Whether an individual company’s news was good or bad did not matter; one index dragged another. Risk-off selling hit the index futures, and the high-beta stocks that make up the index fell in proportion to their weight.
And the selling did not stop at semiconductors. Here is the plainest evidence that no stock was being chosen. IHI makes aero-engines and defence equipment, with almost no exposure to AI infrastructure. Its beta is a low 0.73. That low-beta stock fell 5.07% on the day, deeper than the Nikkei’s 4.03%. A name from outside the AI basket was sold, as a member of the index, harder than its own beta. The reason for the selling was membership of the index, not the content of the business.
There is a second, indirect piece of evidence: the spread of the falls. Had the market been selecting, the gap between the disliked and the protected would have widened, and the spread of returns with it. The opposite holds. On the roughest days the spread between stocks is governed by the market’s own move, and individual stories are buried. Since 2015, on days the Nikkei fell 4% or more, the spread of the nine stocks’ falls averaged 2.0%, against 1.5% on quiet days. The rougher day carries the wider spread, but only because everyone moves hard together, not because selection is at work. The spread on 17 July sits in the very middle of the past decade’s distribution of down days.
＊
That is what happened. What did not happen matters more for judgement.
There were no forced sellers. Nothing in the spread of the falls or in the excess returns shows margin liquidation, forced closes on collateral calls or mechanical index-exclusion selling amplifying the drop. What there was, was risk-off in the index futures and the beta-driven fall that tracked it. No forced selling means this was not the kind of drop that feeds on itself until the floor gives way.
But there were just as few forced buyers.
What has been buying semiconductors and the AI complex, in this phase, is the discretionary investor: momentum money, theme money, money that buys because a thing is going up. The participant who must buy regardless of price, the passive fund tracking an index, does not step up its buying because a stock has fallen. Passive buys a set amount on a set schedule, executed mechanically and with no regard for a dip.
A market without forced buyers has no mechanical floor. Discretionary buyers come in not because the price fell but because they judge it cheap, and when and at what level that judgement lands is not something this analysis can tell you. What it can tell you is that until it does, no automatic force is holding the downside up.
In the event, discretionary buyers did come in that afternoon. The Nikkei fell more than 4,100 points at one point, to the 62,700 area, then clawed back 1,400 into the close. It broke below the 75-day moving average that domestic brokers had flagged as a floor, then closed back above it. Advantest recovered from 10.8% down at the lunch break to 7.2% down at the close. Someone caught the knife. Not because they were forced to, though. Because they judged it cheap.
＊
What can be said about 17 July is modest.
The drop was less the work of some fresh warning about AI than a high-beta cluster, the AI complex, being sold by its beta inside a risk-off session. No stock was selected. The content of the exposure, which company is levered to power, which to NVIDIA’s shipments, which to Taiwan’s capital spending, does not show up in the day’s tape.
Whether that content is ever priced in is a separate question. This piece does not test it. What it can say is that, for the single day of 17 July at least, there is no sign the market looked at the content and chose.
Which is to say that the material for judging whether 17 July was oversold, or under-sold, is not in the day’s price action. The tape followed beta and said nothing more. The material for judgement sits outwith the tape, on the side of structure, in what each company is exposed to. That is set out elsewhere.
Nothing here is investment advice. Prices and falls for 17 July are from press reports and the published figures of the Tokyo Stock Exchange. Betas and excess returns are the author’s estimates from daily returns to 16 July 2026, and will shift with the method used.