Has the Bubble Really Burst (in China?)

China’s STAR 50, the index of the fifty largest tech stocks on the Shanghai exchange, peaked at 2,255 points in early July and closed on 26 October at 1,315.5, about 40% lower according to Reuters (my own calculation gives 42%). The broader CSI 300, which tracks large companies in Shanghai and Shenzhen, is in a technical bear market, roughly 20% below its June peak, according to a commentary I found online, and Hong Kong’s tech sub-index is down more than 30% this year. Headlines now say the bubble has already burst. The price is a fact. “Burst” is already an explanation, and there are at least four on offer.

The first is that China followed the world. On 22 June both the SOX, the US index of chip makers, and South Korea’s Kospi peaked, and Samsung and SK Hynix, the two largest memory-chip makers, weigh more than half of the Kospi. The next day it lost almost 10% after a forecast from Broadcom, a major AI chip supplier, that the market found short. On 28 October Hynix reported profits above forecasts, with its memory chips sold out until 2029, and its shares jumped 11%, though they are still about 15% below a month ago, according to a report I read; an analyst at KB Securities said the market is looking at next year and at doubts over the sustainability of AI spending, not at today’s earnings. The same evening Alphabet, Google’s parent, reported $51.4bn spent in the quarter, mostly on AI infrastructure, and negative free cash flow of $9.1bn for the second quarter running; its shares fell more than 10% in after-hours trading, according to Corriere della Sera. Microsoft spent almost as much, $45.2bn, but generated $18.6bn in cash and rose about 3%. Google also admitted that searches in its AI mode earn less advertising than traditional ones: it sells attention, and a product that answers directly shortens the path to the click. I suspect the market has stopped asking whether the machine is wanted and asks who can afford to wait for the revenue. That is the problem of the fixed due dates I wrote about on 28 October.

The second is that AI is the wrong name for the cause. According to the same commentary, Kinger Lau of Goldman Sachs says China’s fiscal policy has turned contractionary, an investment director at Aberdeen sees no change in fundamentals apart from a tax on offshore trusts, and Macquarie analysts say a lasting rally depends on a recovery in domestic demand; some investors speculate that wealthy people are selling to pay old tax bills. The STAR 50 fell twice as far as the CSI 300, so technology amplified a wider drop. I suspect the label has a cost: calling a largely macroeconomic fall an “AI bubble” moves the responsibility from fiscal policy to speculators and to a technology. I can’t prove it, but the word assigns blame before it explains a fact.

The third is the state, which in China is also a buyer. In January the CSRC, the securities regulator, raised from 80% to 100% the margin required to buy shares on credit, and in May it asked companies and fund managers to justify their valuations. Since July a group of state-aligned funds and brokers known as the national team has been buying, and last week its purchases of exchange-traded funds were probably the most aggressive since April 2025, according to Macquarie, which infers it from an indirect indicator. An analyst at Pantheon Macroeconomics says the aim is to prevent spirals of margin calls. I suspect Beijing wanted less froth and no crash, and that in October it moved from cooling to defending because the fall went past what it wanted. When the buyer who stops the price is the state, the price no longer says only what the market thinks; it says what the buyer wants, and some of the loss passes to the public balance sheet. Domestic investors, who make up most of the trading in mainland China, are the ones Beijing wants to dissuade from fleeing.

The fourth is the one that bothers me most, and the one where I have least proof: that someone gains from the word “burst”. Incentives run both ways, and a word does not need a plan to pay. Whoever must sell shares soon prefers the bubble, if there is one, to be elsewhere: Hynix halted the US listing of its Solidigm unit last week, citing volatility, and Meta, which announced up to $95bn in shares, and Anthropic, which filed on Monday and makes Claude, as I declare, are preparing. Whoever campaigns against data centres has the opposite motive, and before the US midterm elections on 3 November the Reuters/Ipsos polls, in which 64% of Americans say building them fast is not a good thing, weigh on the debate. In Beijing, “correction” costs less than “burst”. I found no evidence of coordination among any of them, and the domestic causes above explain much of the fall without any outside interest, so this is the weakest of the four suspicions.

A limit on the numbers: I don’t know who sold and who bought, and the stock-market multiples circulating don’t say on which earnings they are calculated, so I haven’t used them to decide whether the fall is over. The deliberate-cooling hypothesis would weaken if most sellers turned out to be foreign investors rather than domestic ones.

Before calling it a burst or a correction, the question is who answers for the loss when the buyer is the state, and who has the power to decide the word.