China's AI Stock Crash Wipes Out $480 Billion in Four Months

China's STAR 50 index, the country's main gauge of AI stocks, has fallen 40% over the past four months, erasing nearly $480 billion in investor wealth. The decline marks the first major break in the global AI trade, and it arrived in the market that pushed AI speculation hardest.

Wall Street is still debating whether artificial intelligence is the biggest bubble of this generation or the greatest technological leap since the internet. The Nasdaq has barely moved. China has already answered the question the hard way.

How the AI spending boom was built

The boom began after OpenAI launched ChatGPT in late 2022. US hyperscalers — Amazon, Google, Meta and Microsoft — raced to build data centres packed with specialised AI chips. Silicon Valley executives said openly that they would rather build too much capacity than fall behind.

Most of the chips came from one supplier. Nvidia holds over 70% of the AI chip market. Its annual sales grew nearly fivefold in two years, and its valuation briefly rose from $400 billion to $2.8 trillion.

Hyperscalers and chipmakers have now accumulated $3.7 trillion in combined debt and off-balance-sheet lease commitments to build data centres covering an area almost the size of Manhattan.

Does the revenue justify the spend?

One estimate puts the revenue the AI industry needs at roughly $600 billion a year to generate adequate returns on hardware investments. Actual customer spending on generative AI is expected to reach just $64 billion this year.

Some reported revenue may also rest on circular financing. A tech giant invests billions in a startup such as OpenAI or Anthropic. The startup then buys cloud services from the same tech giant. The revenue is real, but demand may depend partly on investment money circulating inside the ecosystem rather than independent customers.

The first crack appeared on what Wall Street called AI's "Black Monday." Chinese lab DeepSeek released an open-source reasoning model that showed world-class AI could be trained at a fraction of OpenAI's cost using older chips. Nvidia shed $600 billion in market value in a single day. This week, Alphabet lost nearly $500 billion in market value after its heavy AI spending raised concerns about cash flows.

Why did China's bubble burst first?

Beijing fuelled its rally differently. Over two years, state-backed funds poured money into domestic chipmakers, robotics companies and AI labs to revive a sluggish economy and cut dependence on American technology. In a market dominated by retail traders, government backing acted as a stamp of approval. Millions of individuals rushed into AI stocks.

At the peak, AI stocks accounted for 42% of China's entire stock market value and 57% of all trading volume. STAR 50 companies traded at nearly 250 times annual profits. At those levels, small disappointments trigger exits.

Unitree, a state-backed humanoid robot maker, showed the pattern. Its August IPO was oversubscribed 8,000 times. Its first post-IPO report showed net profit down 17% as R&D costs doubled. The stock has since halved.

Several domestic chipmakers have lost half their value since July. Optical hardware makers Zhongji Innolight and Eoptolink plunged 60% on fears Washington could restrict Chinese components in US data centres. Emergency share purchases by a state-owned firm failed to stop the slide.

Is a price war coming?

Chinese AI firms face a familiar problem. As in the solar and electric vehicle industries, subsidies pushed too many companies to build similar AI models and hardware. Intensifying competition may now force price cuts, leaving firms fighting for market share with thin profits.

Investors have a second worry: even if AI transforms the economy, infrastructure builders may not profit. The 1996 telecom boom offers the template. Wavelength Division Multiplexing let a single fibre-optic cable carry 16 data streams instead of one. Convinced internet traffic would triple annually, telecom firms borrowed about $500 billion to lay millions of miles of cable. The internet was real, but capacity far outran demand. When the dot-com bubble burst, 85–95% of buried cables sat unused as "dark fibre," and dozens of telecom companies went bankrupt.

Today's AI infrastructure carries a greater risk. Fibre cables can wait underground for decades. AI chips lose economic value quickly as newer, more efficient processors arrive. If computing capacity grows faster than demand, rental prices for that capacity could fall sharply.

China's $480 billion wipeout may be the first casualty of a global AI reset. The technology can still endure and transform work the way the internet did, while the companies that spent trillions on it struggle to earn a return — a harder problem than building the next model, and one that chipmakers and cloud providers on both sides of the Pacific will now have to answer for in their coming earnings reports.