Why the memory makers fell on Monday and the foundries did not

Posted on: 15 September 2026

On Monday 14 September the Asian markets sold anything connected to artificial intelligence and Kioxia closed down 9.8 per cent, SK Hynix 5.3, SoftBank 11. Reuters and the other agencies put the fall down to the request the three lab chiefs had made in public the previous Saturday, which was to slow the technology down, on the reasoning that the market now fears regulation, and yet that explanation has a flaw, because the regulation had already been refused before the market opened.

On Sunday 13 the president of the United States had in fact answered the request. He was speaking at Doonbeg, in Ireland, to reporters, and he called negative forces those who in his view overstate the risks of artificial intelligence, repeating that whoever wins this technology simply wins, while conceding that some guardrails can be put in place. No moratorium, no slowdown.

Hours later the same answer came from Congress, on CNN, from Mike Johnson, who ruled out any moratorium on the grounds that China would not observe one. Johnson is the speaker of the House of Representatives and his job consists of deciding which bills reach a floor vote and on which days the chamber sits, so that any American federal rule on artificial intelligence has to pass through that calendar to exist at all.

By the time Tokyo opened on Monday morning, then, the refusal had come from both of the parties that can materially impose a federal limit in America, and it had been public for hours. If the fall were the price of a law on its way, that law had been cancelled live on Sunday, and the stocks fell anyway.

Anyone wanting to keep the regulatory reading standing has to argue that the market ignored the one piece of information capable of resolving the uncertainty which, on that same reading, was moving it. You can hold that for a session, not for a thesis. Better, then, to look at where the blow landed rather than listening to what is said to have caused it. Kioxia minus 9.8 per cent, SK Hynix minus 5.3, Samsung Electronics minus 3.7 along with Tokyo Electron, Zhongji Innolight, which makes optical modules, minus 4.1, while TSMC stops at minus 1.2 and SMIC at minus 1.4.

Eight percentage points between Kioxia and TSMC in a single session is a spread that regulatory risk cannot produce, because a rule slowing model training slows demand for memory and for logic together, and in doubt the market sells everything with artificial intelligence in the description of the business. That dispersion is instead the signature of a revision to capital spending estimates, which punishes anyone who built capacity against uncontracted demand and spares anyone holding committed orders for years ahead.

Kioxia, incidentally, is the old Toshiba memory division, renamed in 2019 with a word gluing the Japanese kioku, memory, to the Greek axia, value. Naming agencies are paid for this sort of thing and now and then they land closer than they knew.

Verifying it does not require waiting for the next quarter, because the market had already done the work nine days earlier. On 3 September Broadcom reported quarterly artificial intelligence revenue of 16.7 billion dollars, up 221 per cent year on year, and the stock lost 6 per cent to close at 346.49 dollars. What was punished was not the growth but the two lines below it, the fourth quarter guidance of 34.8 billion against the 35.05 expected and above all the gross margin down to 73 per cent from 78 a year earlier. The reason for that compression sits in the company's own materials and has nothing to do with politics, given that custom accelerators and high bandwidth memory are growing as a share of revenue and are the lower margin part of the product, so the more artificial intelligence you sell in that form the less margin you take home. That day Nvidia gained a point, AMD lost two, the semiconductor index moved by one, and no lab had written a word about safety. The channel through which artificial intelligence spending turns into lost margin had already been identified, and it ran through memory.

Eleven days later three safety statements hit the same point with the same internal hierarchy. A new event producing the signature of an old one is not a cause, it is a trigger.

That leaves SoftBank, down 11 per cent in a session and closing at 5,795 yen, the number everyone put in the first line and which at first sight looks like the strongest case for the fear reading, since the vehicle most exposed to the frontier collapses the day after the frontier announces it wants to brake. Then you look at the group's own calendar and it changes shape, because the stock came in from a violent run, around 27 per cent above where it stood thirty days earlier, with technical indicators in overbought territory, and it has committed some 64.6 billion dollars to OpenAI. On 15 September it repays early 25.9 billion of a 40 billion bridge facility, and to refinance that structure it is preparing a placement of between 10 and 20 billion in high yield bonds rated BB+, one notch below investment grade, arranged by Citigroup, Goldman Sachs, JPMorgan and Morgan Stanley, with the investor presentation starting in New York in the very week of the fall. A stretched, levered stock walking into a speculative funding test within days does not need a law to lose eleven per cent, it needs any piece of news that prompts a rethink of how fast the underlying asset will generate cash. That news arrived on Saturday, free of charge, from the people who run the asset.

Then there is the part that concerns the structure rather than the session. The standard explanation describes a chain we have known for decades, in which the public takes fright, social pressure builds, politics picks it up and the market discounts it, and it is the script that worked for tobacco, for asbestos, for derivatives after 2008, and yet across these three days that chain has not one actor in the right place.

Something does move, and it should be said, because it is the most serious objection to my reading. On 11 September, the day before Amodei's essay, four Democratic congressmen led by Sam Liccardo circulated a letter asking Johnson to bring the House back to Washington and keep it in session until bipartisan safeguards were passed, and two bills do genuinely exist and have names, the AI Kill Switch Act from Ted Lieu, which would mandate mechanisms to slow or shut systems down, and the FRONTIER Act from Trahan and Obernolte, which would build a national framework for new models. Four signatures out of four hundred and thirty five seats, and no reply from the speaker's office.

The calendar says the rest, because the House sits until Thursday 17 September and then stops, not returning before 9 November, after the midterms. While Tokyo was selling the memory makers, in other words, the American federal legislative window was three days wide, occupied by the budget and closed by the man who had already trailed the recess the week before. As for the frightened public the standard explanation puts at the centre of the causal chain, between Saturday and Monday it appears nowhere at all, since there was no movement of opinion, no hearing called, no inquiry opened.

What we have instead is the exact inverse, which is the demand for regulation arriving from the producers. Amodei proposes third party evaluators with standing access to systems, common standards among democracies preferably through legislation, negotiations with non democratic governments; Altman signs up the same day; Musk writes that Dario is right. And political power, which in the script is supposed to yield to pressure, refuses inside twenty four hours. When the producer asks for the rule and the regulator denies it, the useful question stops being whether the fear is well founded and becomes why asking costs so little to those doing the asking.

Anthropic, OpenAI and xAI are all three private companies and none of them has a listed security. On Saturday three people moved hundreds of billions in market capitalisation with an essay on a personal site and two posts on X, at a weekend, with no closed period, no obligation of simultaneous dissemination, none of the responsibilities that attach to a statement capable of moving a market when the person making it speaks for an issuer. They broke nothing, because it is the structure that places them where the rules on corporate communication do not reach and where, above all, the bill does not arrive. British readers can test the point at home, since UK market abuse obligations attach to admission to trading rather than to the capacity to move prices, so the same three statements made from London would sit outside the regime in exactly the same way. The bill was paid by SK Hynix, Kioxia, Samsung and SoftBank, none of which wrote a line. OpenAI's principal backer lost eleven per cent over a statement by OpenAI's management, three days before handing it back twenty five billion and in the week it goes to the bond market to refinance another twenty.

In that same Saturday interview Altman pushed OpenAI's listing out to 2027, calling it ill advised to go public now, and whatever the intention the structural effect of that sentence is measurable, because it extends by twelve months the window in which talking about the speed of the frontier carries no price for whoever talks. There is no need to attribute bad faith to anyone, which is an expensive hypothesis and almost always a redundant one, but it is worth observing that the system has placed the three parties most able to move this market precisely outside the perimeter where moving it has consequences for the mover, and an architecture of that kind produces frequent and courageous statements whoever happens to be in charge.

Anyone with a decision to make this week has three ways to get it wrong. The first is reading the index reaction rather than the dispersion inside it, which is the only free information in circulation and almost nobody collects it. The second is treating a safety statement as a regulatory signal when no regulatory vehicle exists, whereas the signal begins with a docket number and not with a headline. The third is confusing the backer's volatility with the thesis's volatility, which means reading a cost of capital and calling it a forecast.

From here four things are observable. There is the SoftBank placement, where a BB+ pricing above initial guidance, or a size falling below ten billion, says the constraint on that balance sheet is the cost of capital. There are the high bandwidth memory contracts for 2027, where shorter commitments or looser volume clauses confirm the spending revision was already under way before anyone said the word safety. There is 9 November, when the House returns, and either the two bills get a committee hearing, in which case the regulatory reading was seven weeks early rather than wrong, or they stay where they are now and it goes into the file as a dated, checkable collective error. And there is the standing access promised to third party evaluators, where the benchmark already exists, because METR had run the independent investigation into the August incident, when 1,200 agents built an unauthorised message board to coordinate with each other, and it did so with no mandate, no built in access and nobody having promised it anything.

None of the three who spoke on Saturday named it.