A real selloff has hit China’s AI-related shares. But the simplest version of the story—China’s artificial-intelligence bubble bursts, then brings down Wall Street—gets ahead of the evidence.
There are two overlapping episodes. An extended Chinese technology-stock retreat was already well established in October. Fresh selling on Friday, November 6, 2026, then followed a sharp US technology-market decline the previous session. That second episode ran substantially from Wall Street into Asia, not simply the other way around. By Friday’s US close, American technology shares were recovering some ground.
The deeper question is not whether Chinese researchers have suddenly lost their ability to build useful AI. It is whether investors paid prices that assumed too much future profit, too little competition and too easy a path from national technology ambition to shareholder returns.
Below, we separate market facts from forecasts, examine the structural problems and explain what named market professionals actually said about this interconnected trade.
Research scope: Reporting available through November 7, ahead of Sunday publication. Sources include Reuters read through licensed republications, AFP reporting, accessible South China Morning Post text, and primary material from the Shanghai Stock Exchange and US Bureau of Industry and Security. Some newspaper pages expose only introductory text; those limits are identified below. Market figures are attributed observations, not an independently reconstructed exchange-price database. No social-media AI summaries are used as evidence.
Image: Hangzhou, home city of DeepSeek, illustrates China’s AI ecosystem rather than the companies or exchanges discussed below. Photograph by CatOnMars, CC BY 4.0, unchanged.
1. What actually fell—and when?
The autumn repricing came first
In an October 26 analysis, Reuters reported that Shanghai’s STAR 50 had fallen 40% in less than four months, with roughly $480 billion in market value erased in the retreat it described. China’s broad blue-chip index was down about 20% from June highs.
These are dated observations from that report, not November 6 closing calculations. Nor is $480 billion a measure of cash withdrawn from bank accounts: market capitalization is the price assigned to outstanding equity. It can shrink without that amount of money literally leaving the financial system.
Reuters reported that optical-component makers Zhongji Innolight and Eoptolink had fallen roughly 60% from their respective June and July highs, while chip developers MetaX and Moore Threads had halved since July. The optical suppliers’ retreat was linked partly to Washington moving toward restrictions on Chinese equipment in US data centers. A prospective restriction must not be confused with a fully implemented ban.
There was no single obvious trigger for the broader Chinese reversal, Reuters said. Traders cited high prices, tax bills and possible overseas restrictions. That cautions against explaining months of selling with one subsequent political event.
November 6 added a global shock
The accessible portion of the South China Morning Post’s November 6 market report gives these closing moves:
- Hong Kong’s Hang Seng Index: down 1.3%, to 24,481.86.
- Mainland China’s CSI 300: down 1.4%.
- The Hang Seng Tech Index: down 1.8%.
- Shanghai’s STAR 50: down 2.6%.
A Reuters report carried by Business Recorder captured the earlier, more severe trading phase. It said the STAR 50 and an AI index had fallen as much as 5%, while the Hang Seng Artificial Intelligence Index was down 4%. It reported Zhongji Innolight down 8%, Eoptolink down 7%, and Zhipu down as much as 15.4% at that stage.
Those individual-stock moves and early-session readings are not closing returns. The Reuters story’s CSI 300 observation was explicitly from the midday break. The generic mainland “AI Index” was not identified by provider in the accessible article, so it is not treated here as a precisely reproducible benchmark.
A 5% intraday decline and a 2.6% closing decline can both be accurate. They describe a market recovering part of its losses during the session. Equally, October peak-to-date drawdowns and November daily returns answer different questions and should never be added together.
2. “Chinese AI stocks” is not one asset class
The categories matter because their customers, ownership structures and investor bases differ.
Mainland-listed shares include technology suppliers traded in Shanghai and Shenzhen. STAR 50 is a Shanghai STAR Market index, not a basket of every Chinese AI company. The Shanghai Stock Exchange’s own explanation describes an index selecting 50 large, liquid securities. Its innovation remit includes information technology, biomedicine and high-end equipment. Calling it a pure AI index overstates its precision.
Hong Kong-listed technology and AI shares trade in a different market, with different international capital flows. Hang Seng, Hang Seng Tech and Hang Seng Artificial Intelligence are different baskets. Their percentage changes cannot be substituted for one another.
US-listed Chinese ADRs are another category again. A Chinese internet company with American depositary receipts is not the same exposure as a mainland GPU designer. The sources checked here do not establish a comprehensive November 6 return series for Chinese AI-linked ADRs. Consequently, no blanket claim is made that they all suffered a particular percentage loss.
The business-model differences are just as important. Optical transceiver suppliers serving American data-center construction depend on US customer spending. Domestic accelerator designers may depend more on Chinese procurement and import substitution. Model developers face inference costs, customer retention and price competition. Consumer platforms may finance AI from established advertising or commerce businesses.
“AI” is a common narrative. It is not a common income statement.
3. Valuations: a cheaper share is not automatically cheap
Reuters’ October analysis put STAR 50 at approximately 100 times trailing earnings, down from nearly 250, against around 30 times for the Nasdaq 100 at the time. These are reported historical comparisons, not current quotes or equivalent portfolios.
An index containing young, capital-intensive businesses can show a high multiple because aggregate profits are small. Sector composition and accounting complicate international comparisons. Still, the economic point survives those caveats: a large fall can remove exuberance without creating a low valuation.
A hypothetical illustrates the risk. A business earning one unit of profit with a share price of 100 trades at 100 times earnings. A fall to 60 leaves it at 60 times if profits are unchanged. If expected earnings simultaneously deteriorate, the apparent discount becomes smaller still.
Investors were also buying a concentrated theme. Reuters cited Goldman Sachs research putting AI-related shares at 42% of market capitalization and 57% of turnover across the mainland A-share and Hong Kong H-share universe it examined. Those figures depend on Goldman’s classification of AI exposure; they are not AI’s share of Chinese economic output.
Reuters cited HSBC research saying Asia funds remained overweight China despite selling over the preceding three months. That leaves scope for further exposure reductions, but does not prove selling will continue.
Leo Wang, partner at Qianchuang Capital, told Reuters that overemphasizing technology weightings distorted capital allocation and drained other sectors. His argument concerns market structure—not proof that every technology business is overvalued.
4. The commercial problem: adoption without pricing power
An industry can become more technologically capable while becoming less attractive to shareholders.
Open-weight models let users download and adapt model parameters, subject to their licenses. This can accelerate adoption and reduce dependence on one API provider. It also makes a durable premium for basic model access harder to defend.
The Week’s November 6 overview, drawing on the Financial Times, contrasts Chinese labs’ greater willingness to distribute models with Western firms’ emphasis on charging for proprietary access. This is a tendency, not an absolute distinction: both markets contain mixed strategies. The underlying FT article was not accessible in full during this research, so that comparison is explicitly second-hand.
Wang Zhuo, partner at Shanghai Zhuozhu Investment Management, offered Reuters a sharper bearish interpretation: Chinese chipmakers and AI suppliers would eventually face price wars like other manufacturers. That is a forecast about future competition, not evidence that every supplier is already discounting.
Competition can hurt at several levels. Model developers may charge less per token while still paying for chips, electricity and engineering. Cloud operators may discount computing capacity to improve utilization. Hardware suppliers may face customers demanding lower prices just as component costs rise.
The useful metric is not simply tokens served or models downloaded. It is the margin left after delivering the service, followed by enough cash generation to cover development and infrastructure spending.
The contrary possibility is important: cheaper AI may expand usage sufficiently to lift total demand. Applications that were uneconomic at high prices can become viable. That benefits users and potentially efficient suppliers. It does not guarantee the original equity valuations were justified.
5. Chip constraints create a cost squeeze
China’s self-reliance drive creates potential demand for domestic chips. It also exposes the difficulty of replacing an entire computing stack.
In a September 10 investigation, Reuters sources said Huawei had lifted the indicated price of its Ascend 950DT accelerator card above 250,000 yuan, 20–50% above quotations two months earlier depending on contract terms. Sources described a 20–30% increase for Cambricon’s planned next-generation processor.
These were reported quotations for particular products, not audited industry selling-price averages. Reuters attributed the pressure to expensive high-bandwidth memory, including supplies obtained through grey-market channels. The companies did not respond to requests for comment.
HBM feeds processors data fast enough to keep costly computing units productive. A capable chip design cannot fully solve a memory shortage, packaging constraint or software compatibility problem.
The primary-source policy background is clear. In December 2024, the US Bureau of Industry and Security announced controls on certain HBM products alongside semiconductor equipment and software restrictions. BIS explicitly identified HBM as important for AI training and inference at scale. This historical announcement establishes the policy’s origin and rationale; it is not a complete statement of every licensing rule or exception in November 2026.
The apparent contradiction—chip prices rising while investors fear price wars—disappears when time horizons are separated. Scarce components can lift hardware prices now. Competition can still compress future margins. Higher revenue per card does not guarantee higher profit if input costs rise faster.
6. Computing access is political on both sides
Recent reporting complicates the idea that Washington alone determines Chinese access to computing power.
On November 5, Reuters reported that approximately 600,000 Nvidia H200 chips were stored in Hong Kong awaiting shipment to mainland China, citing two people familiar with the matter. The report said Beijing was withholding large-scale import approval despite demand from technology companies.
That stockpile is an anonymously sourced report, not an inventory independently counted for this article. Reuters said the relevant Chinese authorities, Nvidia and the technology companies it contacted did not comment.
The strategic tension is understandable: allowing imports can accelerate model development, while limiting them can support domestic alternatives and reduce dependence on American technology. Training and inference may receive different policy treatment.
For Chinese buyers, these choices affect deployment schedules, working capital and hardware selection. For US suppliers, addressable demand depends on Chinese import decisions as well as American export permissions.
A warehouse full of chips is not automatically recognized supplier revenue. A reported order is not completed delivery. Nor can one inventory figure be translated directly into an earnings forecast without prices, contractual terms and accounting details.
7. Profitability and funding: what state support cannot settle
The evidence does not support claiming that the whole Chinese AI sector is unprofitable.
The accessible introduction to a November 6 South China Morning Post report says several domestic AI chipmakers crossed into profit in the third quarter, while memory costs and supply-chain problems clouded the outlook. Only introductory text was available here; the detailed company accounts and banks’ margin models were not inspected. This article therefore does not reproduce precise quarterly earnings claims from inaccessible text.
The important distinction is between an accounting milestone and durable financing capacity. A profitable quarter does not establish positive free cash flow after capital spending. Revenue growth does not establish diversified customers. State procurement can improve demand while creating dependence on official priorities.
Hong Hao, managing partner and chief investment officer at Lotus Asset Management, told Reuters in October that the retreat would obstruct companies seeking to list. This is the funding feedback loop: falling valuations make new equity harder to raise, affecting research and infrastructure financing.
AFP’s November 6 reporting described renewed official support through liquidity tools and state-backed fund activity. It cited Bloomberg data for 120 billion yuan flowing into nine ETFs on Tuesday. That remains AFP’s attributed reporting, not an independent identification of every buyer.
Charu Chanana, chief investment strategist at Saxo Markets, told AFP that weak domestic demand, intense competition and disappointing earnings had pressured Chinese technology stocks. She also said policy support, lower valuations and improving sentiment could encourage renewed interest.
Stephen Inness of SPI Asset Management offered a more cautious assessment: stabilization measures had so far been insufficient against a weak domestic economy and doubts about AI earnings.
Policy can support liquidity. It cannot by itself establish sustainable returns on every GPU cluster or model-training program.
8. What named strategists say about the global consequences
The strongest evidence concerns shared vulnerabilities—not a precise forecast that China will cause a particular US decline.
In Reuters’ November 6 China report, Jason Hsu, founder and chairman of Rayliant Global Advisors, emphasized that Chinese sentiment was heavily influenced by US AI performance. Dong Baozhen, chairman of Lingtong Shengtai, described the US market as the backbone of the global AI rally and questioned whether other AI bull markets could hold up after its slump. These are views about dependence on the American AI trade, not numerical S&P 500 targets.
Dilin Wu, a Pepperstone strategist, said China could not completely escape aggressive Wall Street selling, but cautioned against treating its whole AI supply chain as one trade. She distinguished self-reliance-driven domestic chip businesses from other exposures. Huawei, named in that discussion, is not itself a publicly traded equity investors can simply buy on an exchange.
The American evidence also challenges a one-direction crash narrative. Reuters’ November 6 US closing report said the Nasdaq gained 1.1%, the S&P 500 rose 0.31%, and the Philadelphia Semiconductor Index advanced 4.2% on Friday. Nvidia rose nearly 3%. Those were partial rebounds after Thursday’s rout, not proof the underlying concerns had disappeared.
Gerry Sparrow, president and portfolio manager at Sparrow Capital Management, told Reuters that much of the feared disruption to the AI buildout probably would not occur, emphasizing uncertainty surrounding political change. This offers a substantive counterargument to the bearish reading.
Reuters described investor fears after the US midterms about regulatory scrutiny and data-center constraints. Fear of future legislation is not legislation already enacted. Nor does a market reaction establish that one political factor explains every trade.
9. Four channels from China’s problems to American equities
The following is this article’s analytical framework, not an attributed analyst price forecast.
Valuation comparisons. If Chinese AI businesses were priced for unrealistic profits, investors may revisit similar assumptions elsewhere. US companies with established cash-generating operations are not equivalent to newly listed specialist chipmakers. Neither are they immune to lower expected returns on AI investment.
Supply-chain exposure. US hyperscaler spending supports global chip, networking and optical demand. Supplier stress could signal changing customer budgets, but falling share prices alone do not establish order cancellations. Replacing restricted Chinese components could also create costs or opportunities for alternative suppliers.
Model economics. Competitive Chinese models can pressure API prices and weaken the scarcity value of proprietary access. Cheaper inference could simultaneously help American software businesses deploy AI economically. The effect differs between sellers of expensive intelligence and buyers using intelligence as an input.
Risk positioning. Global funds can reduce exposure across regions when volatility rises or financing tightens. The October positioning evidence makes this plausible. Quantifying the effect would require detailed flow and leverage data beyond the sources reviewed here.
These mechanisms justify investigating exposure. They do not establish a fixed contagion coefficient between Shanghai and New York.
10. What would confirm—or challenge—the bearish case?
A prolonged downturn becomes more convincing if falling valuations are followed by weaker bookings, deteriorating cash conversion, more dilutive fundraising and slower customer deployment. Competition alongside expensive chips would reinforce the margin squeeze.
A less damaging outcome would involve better utilization, falling deployment costs, stable infrastructure orders and measurable business value from AI applications. Chinese policy support could buy time, while diversified American companies could absorb disappointing individual projects.
There is also a mixed outcome: AI usage keeps rising and businesses benefit, but profits accrue to fewer suppliers than investors expected. Technological success and disappointing stock returns can coexist.
The next useful evidence is therefore company filings, order conversion, cash flow, utilization and actual policy decisions—not simply another dramatic trading screen. Valuation comparisons should use consistent dates and definitions rather than mixing intraday lows, closing prices and different historical peaks.
Bottom line: a warning about economics, not proof AI has failed
China’s AI-stock retreat is real, predates the latest American selloff, and exposes demanding valuations, uncertain pricing power, costly compute, financing dependence and policy friction.
November 6 also demonstrates why causality matters. American technology selling hit Asia; US shares then partly rebounded. Named market professionals disagree over structural damage versus overreaction. None of the sources reviewed establishes a reliable timetable for a China-driven US market crash.
The defensible conclusion is narrower: the global AI investment cycle is being tested on profitability and cash generation, not merely technical progress. China offers an especially sharp version of that test. American markets face related questions, but their answers depend on their own customers, balance sheets and returns on investment.
This is an explanation of market reporting and competing interpretations, not personalized investment advice.
Sources and verification notes
- Reuters, October 26, updated October 27: In China, the AI bubble has already burst, read through Investing.com. Dated drawdowns, valuation comparisons, positioning and October commentary.
- Reuters, November 6: China, Hong Kong stocks tumble as global tech selloff deepens, read through Business Recorder. Midday/intraday observations and Hsu, Dong and Wu commentary.
- South China Morning Post, November 6: Tech stocks lead declines in Hong Kong, mainland China. Accessible opening text supplied closing index figures; the full subscriber article was not read.
- AFP, November 6: Beijing seeks to bolster AI shares, read through Yahoo Finance. Policy support and Chanana/Inness commentary. Its Bloomberg and FT references remain secondary.
- Reuters, September 10: China’s AI chipmakers raise prices as high-bandwidth memory shortage bites. Direct Reuters access was unavailable; the republication was read.
- Reuters, November 5: Nvidia chips pile up in Hong Kong as China withholds import approval. Anonymous-source inventory and pricing claims remain attributed reporting.
- Reuters, November 6: Wall Street stocks rally after heavy AI-related selloff, by Chibuike Oguh and Marc Jones. US closing moves and Sparrow’s counterview.
- South China Morning Post, November 6: China’s AI chipmakers report first profits, but will the trend continue?. Accessible introductory material only; no filing-level profit verification claimed.
- The Week, November 6: Why China’s AI bubble is bursting. Accessible text used for its attributed summary of FT business-model analysis; underlying FT full text was unavailable.
- Shanghai Stock Exchange, primary source: Q&A on launching STAR Market 50. Historical design explanation, not a current constituent or price audit.
- US Bureau of Industry and Security, December 2, 2024, primary source: Commerce strengthens export controls. Historical HBM-policy background, not comprehensive current legal advice.
Limits: Proprietary analyst notes, complete exchange time series and company quarterly filings were not independently audited. Reports published at different moments can use different baselines; October drawdowns, November intraday readings and November closing returns are kept separate. No ADR returns are inferred from mainland data, and no missing figure or quotation has been filled in from social-media summaries.
