For most of 2026, "is AI a bubble?" was a question about the United States. In late October it got an answer, but from somewhere else. China's STAR 50 index, the mainland benchmark for AI and semiconductor companies, fell about 40% in less than four months, wiping out roughly $480 billion of market value, according to a Reuters analysis headlined "In China, the AI bubble has already burst". Over the same stretch the broader CSI 300 index fell about 20% from its June peak, into bear-market territory.
Headlines about an "AI bubble burst in China" are now everywhere, and they tend to blur two things together: a stock market that ran far ahead of itself, and a technology that people and companies keep using. This piece separates them. It sets out what actually fell, why it fell, what Beijing is doing about it, and what it means if your business builds on AI models, Chinese or otherwise.
Nothing here is investment advice. It is about building systems, not buying shares.
What actually fell
The fall was steep because the rise was. Reuters, citing Goldman Sachs, reports that AI-related stocks made up 42% of China's market capitalisation and 57% of turnover across mainland A-shares and Hong Kong H-shares. At its peak the STAR 50 traded at close to 250 times the previous twelve months' earnings. After the fall it still sits near 100 times, against about 30 for the US Nasdaq 100. The Nasdaq itself was only about 6% below a record it set three weeks earlier.
The losses were spread across the whole AI supply chain:
| Company or index | What it does | Reported fall |
|---|---|---|
| STAR 50 | AI and chip benchmark | about 40% from its early-July record |
| MiniMax | AI model maker (Hong Kong) | about 88% from its March peak, now below its IPO price |
| Zhipu AI | AI model maker (Hong Kong) | about 85% below its June high |
| Zhongji Innolight, Eoptolink | optical parts for data centres | about 60% from June and July peaks |
| MetaX, Moore Threads | GPU makers | about half their value lost since July |
| Unitree | robotics | about half since it began trading in August |
Sources: Reuters via The Kathmandu Post and Caixin.
The two listed model makers are the clearest case. MiniMax debuted in Hong Kong on 9 January 2026, peaked at HK$1,330 on 18 March and, by 23 October, traded at HK$160.7, below the price it listed at. Bloomberg puts the combined loss for MiniMax and Zhipu at about $150 billion from their highs. Unitree shows how much of the buying came from ordinary savers: its listing was 8,000 times oversubscribed by retail investors.
Why it burst
Reuters found no single trigger. Traders pointed to a combination of things that had been building for months.
Valuations that assumed everything would go right. A market trading at 250 times earnings has priced in years of flawless growth. Any doubt about demand is enough to move it a long way, and in a market dominated by small investors it moves fast. One Beijing fund manager told Reuters the rally had "evolved into a speculative crazy bull" and that "now we see the bubble bursting".
Export risk. Washington is moving towards barring Zhongji Innolight and Eoptolink from supplying US data centres, and traders cited the wider threat of Chinese equipment being banned from data centres abroad. For companies whose growth story depended on selling into the global AI build-out, that is a direct hit to future revenue.
A price war that the model makers started themselves. This is the part that matters most for anyone buying AI. On 22 May 2026 DeepSeek permanently cut the price of its flagship V4-Pro model by 75%, to 3 yuan (about $0.44) per million input tokens. Competitors had to respond. MiniMax launched its flagship M3 model on 1 June and halved its API fees a week later; JPMorgan read that as a sign the market had not accepted its premium pricing. Cheap tokens are good for customers and bad for the margins of the people selling them: in the first half of 2026 MiniMax's gross margin was 17.9%, and Zhipu's 26.4%, while MiniMax still lost $358 million on $117 million of revenue.
Tax bills and leverage curbs. Traders also cited tax bills, and regulators had already been cracking down on leveraged and high-frequency trading, which removed some of the borrowed money that had pushed prices up.
The other bubble: data centres nobody uses
The stock market is the visible part. Underneath it is an older problem: China built a great deal of computing capacity that is not being used.
After ChatGPT arrived, more than 500 city and provincial data-centre projects were announced, and by early 2025 the rent on an eight-GPU Nvidia H100 server had fallen from about 180,000 yuan a month to 75,000. Much of the building followed the national "Eastern Data, Western Computing" programme, which put data centres in the energy-rich but distant west. An analysis by the Australian Strategic Policy Institute found some of those western sites running at 20–30% utilisation, against a policy goal of more than 60%. The reasons are practical: AI that answers people in real time needs low latency, so the busiest workloads stay near the eastern cities, and remote sites struggled to find staff and customers. Provinces competing for growth had offered subsidies, tax breaks and cheap land, and built in parallel.
Claims that as much as 80% of new capacity sits idle circulate widely, but they come from local press estimates and are disputed. The more defensible reading is a mismatch: too much of the wrong capacity in the wrong places, and not enough of the right kind. Beijing has since said no new large data centres should be built in cities where existing ones run below 50% utilisation.
What Beijing is doing about it
China's leadership had been warning about this. On 9 October 2026, with the sell-off already under way, the Communist Party and the State Council issued guidelines on "new productive forces", the category that includes AI. According to Reuters' report of the Xinhua release, they tell officials to avoid "swarm-like" rushes into high-tech sectors, investment bubbles and blind investment, and say officials responsible for major losses from blind investment will be held accountable. The same document commits to the "AI Plus" programme of using AI to upgrade traditional industries.
So the official line has two halves: keep pushing AI into the economy, and stop local governments and investors from piling into the same projects at once. The securities regulator's chairman, Wu Qing, met market participants in October to collect suggestions for improving the market, and state-backed investors had already been selling into rallies to cool them. Nothing reported so far looks like a plan to rescue share prices.
A market crash, not a technology crash
It is worth being precise about what failed. What fell were share prices that assumed AI companies would capture enormous profits quickly. What did not fall was the usefulness of the models. Chinese labs kept releasing them throughout the slide, and the price of using them fell sharply over the year. In the first half of 2026 MiniMax's revenue grew 283%, which is not what a product nobody wants looks like. The problem was the gap between that growth and the price investors were paying for it.
That is the same distinction our earlier piece on whether AI is a bubble drew for the US market. China has now shown what one version of a correction looks like: the technology carries on, the weakest providers come under pressure, and the price of using AI becomes less predictable.
What it means for a business building with AI
If you are using AI in your operations, or planning to, the China crash does not change whether a good AI system is worth building. It does change a few things about how to build one.
Today's model price is not a contract. DeepSeek cut prices by 75% in May, then in August announced a "significant" increase to its API prices and added surcharges at peak hours. A price war run by companies losing money cannot last forever. If your business case only works at this quarter's token price, it is not a business case yet. Work out the cost per task at two or three times today's price and see whether it still pays.
Your model provider might not be there in two years. A company whose shares are down 88% and that loses three dollars for every dollar of revenue may be bought, merged or shut. Build so the model can be swapped: keep the prompts, the evaluation set and the business logic in your own system, and treat the model as a component behind an interface you control, not as the foundation.
Measure before you scale. Beijing's own instruction to its officials, prove the project fits before you pour money into it, is good advice for a company too. Start with one workflow, measure the time or errors it saves, and expand only when the numbers hold up. Our guide to deciding whether your business needs AI at all covers how to choose that first workflow.
Know where your data goes. Export controls and data-centre bans are now moving share prices. If you use a model hosted abroad, check where your data is processed, what happens to it, and what you would do if access to that provider were restricted. Open-weight models you can run yourself, or through a provider in your own jurisdiction, reduce that exposure.
Don't read a market crash as permission to wait. The companies that lost value were mostly priced for a future that has not arrived. The businesses using their models are still getting the work done, at prices that have fallen sharply. A correction makes careful buyers better off, as long as they build so that no single provider's fortunes can take their system down with it.