While the rest of the world argues about whether artificial intelligence is a bubble, China has already run the experiment. The STAR50 — the Shanghai index of the country's AI and semiconductor champions — is down about 40% in less than four months. Roughly $480 billion of market value has gone. The broad blue-chip index is about 20% below its June high, and the market's fall is the steepest since the 1990s outside a crisis.
The useful question is not whether China had a bubble. It did, and it was bigger than anything in the West: AI stocks reached 42% of Chinese equity market capitalisation and 57% of turnover, according to Goldman Sachs. The useful question is what actually broke, because it wasn't the thing most people mean by "AI".
What fell, and how fast
- Unitree, the robot maker that was 8,000 times oversubscribed at its August listing, has halved.
- MetaX and Moore Threads, chip designers that jumped 700% and 400% on debut last year, have halved since July.
- CXMT, the memory champion, is down more than a quarter in a month.
- Zhongji Innolight and Eoptolink, which make the optical components that connect AI clusters, are down about 60% from their June and July highs — partly because Washington is moving to ban them from supplying US data centres.
No single event started it: traders cite high prices, a tax bill and the threat of a foreign ban on Chinese hardware in data centres. That is what a bubble looks like when it goes — not one pin, but the absence of the next buyer.
The buildout is real, which is the point
Here is where China differs from the dot-com story that gets retold every time. The infrastructure isn't imaginary. China has more than 24 gigawatts of datacentre capacity — more than Europe, more than the rest of Asia — and a reported Rmb2 trillion state plan to connect it, with at least 80% of the technology sourced domestically. Huawei already holds about 28% of China's accelerator market against Nvidia's 34%.
The capex is real too, and it is why the free cash flow disappeared. Alibaba, Tencent and Baidu spent about $20 billion between them in the second quarter, more than double a year earlier, and for the first time on record all three posted negative free cash flow. Across the first half, the three went from Rmb170 billion of positive cash flow in 2025 to a negative Rmb16 billion.
So the shares did not fall because China stopped building. They fell because the equity prices implied a demand curve that the physical supply chain cannot deliver on schedule.
The tell was in the earnings
Cambricon is the cleanest example. The company is China's most important AI chip designer, the beneficiary of a national push to replace Nvidia, with chips inside models from DeepSeek and Moonshot. Its third-quarter revenue more than doubled to 3.53 billion yuan, and net profit rose 148%.
It still missed analyst expectations by 44%, because it could not get enough chips made or enough high-bandwidth memory to ship. Operating cash flow was negative. Rival Hygon grew 45% and spends nearly 30% of revenue on research.
That is the paradox at the centre of the whole trade: demand is not the constraint, fabrication is. You cannot value a chipmaker on the demand it can see when the bottleneck is the thing it cannot buy.
The part nobody prices: utilisation
China has built over 500 intelligent computing centres, and by the government's own numbers only about 36.8% of national AI compute was in use in 2025. Enterprise general-purpose compute utilisation is reported as low as 10–15%. The United States, for comparison, has near-zero datacentre vacancy because almost everything under construction is pre-leased.
Some of that gap is structural: the "East Data, West Computing" policy puts training in provinces with cheap power and few customers nearby. But 37% utilisation does not support a 250x multiple, and barely supports the 100x where the STAR50 sits now, against roughly 30x for the Nasdaq 100.
Why it matters outside China
Beijing wanted a "slow bull": a steady market that would draw household savings into stocks and fund the AI push without the boom-bust cycle that has burned Chinese savers before. State funds were sellers into the rallies to keep gains in check. It became, in one Beijing fund manager's words, "a speculative crazy bull" — and the small investors who bought because the state seemed to be behind it are holding the losses.
The practical consequence is funding, not sentiment. If the IPO window closes, companies that planned to raise equity for compute will borrow instead — just as their free cash flow has turned negative.
The counter-case deserves a hearing: Goldman Sachs argued in August that China's AI sector has no overall bubble, and the buildout is genuinely state-funded and continuing. That is probably right about the infrastructure and wrong about the equities. Both can be true: a real industrial programme financed through a market that priced it like a lottery. The market corrected first. The factories are still running.
What to watch
Cambricon's fourth-quarter deliveries will settle whether the miss was a supply blip or a ceiling. Watch the MSCI-style index decisions and the IPO queue for signs of how long the funding window stays shut, and watch utilisation — if the state's compute hubs fill up, the bear case weakens; if they don't, the next leg down is about demand, not valuation.
Sources
- In China, the AI bubble has already burst — Reuters, 26 October 2026
- The Chinese AI infrastructure boom — SemiAnalysis
- China's data centre buildout: who benefits — CGS International
- Cambricon's profit soars 148% but revenue misses by 44%
- China's AI sector has no overall bubble: Goldman Sachs — The Standard
- China built hundreds of AI data centers. Now many stand unused — MIT Technology Review
- 500座智算中心利用率调查 — 第一财经
No comments yet