China’s AI industry is becoming more self-sufficient, but that does not make it irrelevant to a global AI stock selloff. The country’s domestic cloud platforms, language models and increasingly local semiconductor supply chain reduce some direct commercial dependencies on the United States. Chinese suppliers and competitors nevertheless remain capable of changing the earnings outlook for global AI companies, while their own shares respond to shifts in international risk appetite.
The October 2026 correction offers a useful test. Chinese technology stocks fell alongside their overseas peers. At the same time, reported progress at Chinese memory maker CXMT added a competitive concern to an already weakening global semiconductor trade. China was therefore both a recipient of the market shock and a potential contributor to the repricing of selected AI businesses.
The October selloff crossed the border
On October 22, China’s CSI 300 Index closed at 4,036.30, down 1.8% for the day and 20.2% below its June 22 closing peak, according to Reuters. The benchmark had entered technical bear-market territory under the conventional peak-to-trough definition. AI-related shares were among the areas under pressure, although weaker domestic policy expectations and broader global risk aversion also contributed to the decline.
The scale of the move is difficult to reconcile with the idea that a domestically oriented technology system can escape a global valuation reset.
Earlier in the month, the weakness had already become visible in the Chinese technology supply chain. On October 9, Reuters reported that the Shanghai STAR 50 Index was down nearly 4% during the session, while Zhongji Innolight, a manufacturer of optical transceivers, fell 5.4% and memory producer CXMT declined nearly 5%.
The repricing continued across several categories. Optical networking, semiconductor equipment, AI servers and domestic accelerators experienced substantial declines despite different levels of exposure to US customers. Some companies were affected by concerns about future demand, others by valuation and financing conditions, and still others by profit-taking after earlier gains.
Hong Kong-listed technology companies were also exposed. Their shareholders participate in a more internationally accessible equity market, while the operating businesses of Alibaba, Tencent and other major platforms depend heavily on Chinese customers. Domestic revenue exposure did little to prevent a reassessment of their expected earnings and investment returns.
There is an important attribution problem, however. The October correction was not caused by a single Chinese competitive announcement. Higher US inflation, renewed interest-rate concerns, rising financing costs for AI infrastructure and doubts about the eventual returns on data-center investment had already undermined confidence.
The distinction between the origin of a shock and its transmission matters here. China’s technology industry may have played a secondary role in initiating the global correction, yet it was exposed to the same valuation mechanism and could amplify specific competitive concerns.
Market evidence establishes that the two systems were repriced together during the episode. It does not establish how much of the Chinese decline was caused specifically by the US correction, or vice versa. A causal estimate would require synchronized security-level returns, factor controls and exposure-adjusted earnings data that are not available from headline index performance alone.
A domestic computing system of meaningful scale
China is building a large AI infrastructure market with its own demand, equipment ecosystem and policy priorities. Its domestic technology platforms can deploy models into search, advertising, social media, e-commerce, games, manufacturing and government services without first obtaining customers from US cloud providers.
This creates a real degree of operating independence. Domestic demand can support infrastructure deployment even if investment by American technology companies slows. It also means that the Chinese AI industry should not be treated simply as a smaller extension of the US hyperscaler investment cycle.
SemiAnalysis estimates that China will have more than 24 gigawatts of delivered data-center IT capacity by the end of 2026. Its corresponding estimate for the United States is 56 gigawatts. The Chinese figure represents approximately 43% of the US total.
These figures describe broad data-center capacity, not equivalent volumes of AI compute. China’s installed base includes conventional cloud and legacy facilities alongside facilities designed for high-density AI workloads. Differences in accelerator efficiency, networking, utilization and power density make a gigawatt of installed capacity a poor substitute for a direct comparison of model-training or inference performance.
The underlying investment remains substantial. SemiAnalysis tracks more than 1,000 facilities across more than 60 Chinese operators. It separately identifies approximately 20 gigawatts of projects in its development pipeline and another 30 gigawatts of announced projects. Pipeline and announced capacity should not be added to operating capacity as though delivery were assured.
China’s Ministry of Industry and Information Technology has also established a target of 9,800 EFLOPS of intelligent computing capacity by 2030 under the country’s 15th Five-Year Plan. The plan includes a target of RMB3.8 trillion in cumulative information-infrastructure investment during 2026–2030.
EFLOPS measures computing performance, while gigawatts measure power capacity. They describe different dimensions of the infrastructure. The government target is a policy ambition, and its realization depends on the availability, efficiency and deployment of computing systems.
China therefore has a substantial domestic foundation for AI investment. That foundation helps protect the direction of infrastructure development from short-term changes in US corporate spending plans. It offers considerably less protection against a global increase in the discount rate applied to future cash flows.
Where Chinese firms can reprice global economics
The principal channels through which Chinese technology companies can influence global AI valuations are competition in standardized semiconductors, participation in international hardware supply chains and the possible reduction of computing and model-service prices.
Memory competition can change the scarcity premium
During the October correction, reports about progress in CXMT’s next-generation DRAM cell architecture added to concerns about future memory supply. Reuters reported on October 13 that Micron shares fell 4.5%, while several storage-related US technology companies also declined. The market was already unsettled by the sustainability of AI expenditure, making the competitive news one additional source of pressure.
The economic connection is straightforward. Global memory manufacturers benefit when supply remains tight relative to demand, particularly in products that require advanced processes, high yields and extensive customer qualification. Their earnings expectations can deteriorate if another supplier is expected to enter profitable product categories sooner than anticipated.
China’s domestic market is large enough to support the development of memory suppliers even before those companies become major exporters. A domestic producer that replaces imported conventional DRAM can eventually change global trade flows. If that producer expands output, incumbent suppliers may face weaker pricing or need to redirect capacity toward higher-value products.
There is, however, a meaningful difference between conventional server DRAM and high-bandwidth memory used in advanced AI accelerator systems.
Cell-density improvements do not demonstrate competitive HBM production. HBM requires advanced die stacking, through-silicon vias, packaging integration, thermal control, stable yields and qualification with accelerator customers. Commercial capability in DDR5 cannot be translated directly into an estimate of future HBM market share.
The relevant competitive sequence is therefore technology development, manufacturing yield, customer qualification, commercially meaningful shipments and pricing response. News about a planned product or process improvement may affect valuations before the full sequence has been completed.
China can influence the global memory investment narrative ahead of materially changing realized industry profits. Whether that influence becomes durable will depend on actual shipments and the response of incumbent manufacturers.
Optical suppliers connect Chinese manufacturing to global AI spending
Optical transceivers occupy a different position. Companies such as Zhongji Innolight and Eoptolink participate in the high-speed networking equipment market required to connect AI servers and data-center infrastructure.
The underlying business is tied to infrastructure spending by customers building increasingly large computing clusters. A reduction in accelerator deployment, network upgrades or data-center construction can reduce demand for the associated optical components, even when those components are produced by Chinese companies.
Conversely, the ability of Chinese optical suppliers to increase capacity, improve yields or compete on price can influence the economics of the broader global supply chain. Lower component prices may benefit systems buyers while reducing margins for suppliers competing for the same product specifications.
Optical technology is not a single interchangeable product. Transmission speed, reach, power consumption, reliability, packaging and customer qualification differentiate market segments. A supplier’s exposure to global AI spending must therefore be evaluated through its actual product mix and customer base.
The October stock declines demonstrate that investors were pricing these companies partly as participants in the global AI capital-spending cycle. They do not, on their own, establish the percentage of each company’s revenue attributable to US hyperscalers.
Open models can influence the price of AI intelligence
Chinese models create a third potential channel. Domestic deployment and local hardware do not prevent software outputs from competing internationally, particularly when model weights, development tools or compatible interfaces are broadly available.
Open-weight models can lower switching costs for developers, encourage competition in model serving and reduce the price that application providers are willing to pay for some categories of inference. The result may be beneficial for enterprises purchasing AI services while challenging the pricing assumptions of model providers.
That is a possible global economic effect even when a Chinese model developer earns most of its revenue at home.
The timing is less immediate than in traded memory or optical products. Lower model prices do not necessarily reduce industry revenue: cheaper inference can increase usage, enable new applications and generate more computing demand. The net effect depends on demand elasticity, model performance, switching costs and the ability of providers to capture value through distribution and application integration.
A Chinese open model becoming widely adopted internationally would be evidence of competitive reach. It would not automatically prove that US AI-company earnings are declining.
The cross-border exposure map
China’s AI industry should be divided by economic function before drawing conclusions about its global influence. The following table separates direct demand exposure, competitive transmission and capital-market sensitivity.
| Value-chain segment | Representative companies | Global transmission mechanism | Relative exposure |
|---|---|---|---|
| Domestic AI models and applications | DeepSeek, Z.ai, MiniMax | Model pricing, international deployment and developer adoption | Lower direct revenue exposure; potentially meaningful competitive impact |
| Chinese cloud platforms | Alibaba, Tencent, Baidu | Capex, financing, cloud pricing and international expansion | Domestic demand buffer; material financial-market exposure |
| Domestic AI accelerators | Huawei Ascend, Cambricon | Import substitution, ecosystem development and future accelerator competition | Lower immediate US customer exposure; potentially significant longer-term competition |
| Memory and semiconductor manufacturing | CXMT, SMIC, NAURA | Supply substitution, equipment investment and component pricing | Product-dependent; stronger where markets are internationally competitive |
| Optical networking and AI hardware | Zhongji Innolight, Eoptolink | Global AI infrastructure orders, customer qualification and component margins | Potentially high; requires company-level customer and revenue data |
| Data-center operators | GDS, VNET and regional operators | Lease commitments, customer utilization and financing costs | Primarily local operating demand, with financing and some international exposure |
These are qualitative assessments of economic exposure, not estimated stock-return correlations. A company can have low direct revenue exposure to the United States and still experience substantial equity-market volatility. Equally, a globally competitive supplier may suffer a share-price decline without an immediate reduction in orders.
The distinction becomes clearer when considering two directions of transmission.
In a US-led demand shock, hyperscalers and AI infrastructure providers might postpone orders, reduce procurement or seek lower prices. Suppliers serving those buyers would face direct revenue consequences. Domestic Chinese platforms would be exposed primarily through their own investment decisions, financing costs and valuation multiples.
In a China-led competitive shock, a Chinese supplier might announce a commercially viable substitute for an imported component or a lower-cost alternative to an established technology. The first impact could be a reassessment of future margins at foreign incumbents, even before the Chinese producer records significant export revenue.
The two shocks can also reinforce one another. Competition can lower expected margins just as higher financing costs make investors less willing to fund capacity expansion. This combination can materially change the economics of capital-intensive AI infrastructure.
Capital expenditure survives while cash returns are tested
China’s domestic AI buildout has considerable momentum, but it is becoming more capital intensive. That exposes local companies to some of the same operating and financial trade-offs confronting US technology groups.
Alibaba’s June 2026 quarter illustrates the tension. The company reported 45% year-on-year growth in AI Cloud and Compute Services revenue, reaching approximately US$7.1 billion. Its cloud segment adjusted EBITA margin increased to roughly 12%.
Those results indicate that AI infrastructure demand is translating into reported revenue and stronger segment profitability. Yet Alibaba’s consolidated financial results also reflected substantial investment requirements and pressure elsewhere in the business.
Demand growth at the cloud division cannot, by itself, establish the investment return for the group. The economics depend on the useful life of computing assets, depreciation, workload utilization, customer pricing and the cost of sustaining technological competitiveness.
Tencent reported an even more pronounced investment increase. In the second quarter of 2026, capital expenditure reached RMB52.8 billion, up 176% year on year. Reported free cash flow was negative RMB13.8 billion.
The company also explained the accounting and timing effects. AI-related prepayments contributed materially to the cash outflow. Tencent stated that free cash flow would have been positive RMB37.6 billion excluding prepayments for compute procurement.
This qualification matters. Negative reported free cash flow can reflect the timing of capacity procurement as well as deteriorating economics. Nevertheless, prepayments tie up capital before the associated services have necessarily produced their full expected cash returns.
China’s cloud companies are therefore building an increasingly expensive domestic AI market. They have meaningful operating demand, but the market must still establish whether that demand generates adequate returns over the life of the infrastructure.
SemiAnalysis estimated that Alibaba, Tencent and Baidu collectively spent about US$20 billion on capital expenditure during the second quarter of 2026, more than double the year-earlier level. Its analysis also identified negative free cash flow across the three companies in that quarter.
This shows that China has an investment cycle of its own. It does not demonstrate that the cycle is immune to global capital-market pressures. More expensive borrowing, weaker share prices and uncertainty over future utilization can make marginal projects less attractive even when domestic policy continues to support AI development.
China’s preference for leased data-center infrastructure creates an additional measurement problem. A company purchasing servers while leasing buildings and power infrastructure reports a different capex profile from a company that constructs and owns the entire facility. Comparing their reported capital expenditure without examining lease commitments, prepayments and infrastructure ownership may distort the true economic burden.
A useful comparison therefore includes infrastructure ownership, leasing obligations, accelerator procurement, operating cash flow and the commercial demand attached to installed compute.
The same principle applies to AI factories outside China. As discussed in Sector Foundry’s analysis of NAVER’s AI data-center strategy, financing arrangements and technology access can enable a large buildout before customer utilization and sustainable operating margins are established.
Infrastructure deployment is a measurable activity. Its eventual economic return is a separate question.
The counter-case: separation is real at the compute layer
There is a strong argument for treating the two AI ecosystems as increasingly independent over the long term.
China’s domestic accelerator development is driven partly by restrictions on access to leading foreign chips. Local cloud companies are working with alternative computing architectures, while public policy supports domestically controlled hardware and software ecosystems.
US export controls reinforce this separation. In January 2026, the Bureau of Industry and Security revised the licensing review process for Nvidia H200, AMD MI325X and certain comparable semiconductors, permitting case-by-case consideration subject to specified conditions.
That revision did not restore unrestricted access to advanced computing products. The licensing framework continued to impose operational and compliance requirements.
Nvidia’s filing for its fiscal second quarter of 2027 provides an unusually direct account of the commercial consequences. The company said shipments under the revised H200 licensing program represented less than 1% of Data Center revenue in the quarter. It also stated that regulatory and geopolitical conditions had effectively prevented it from competing in China’s data-center compute market at scale.
This supports the argument that China’s domestic AI accelerator market is becoming structurally less dependent on Nvidia’s latest product sales.
However, two forms of independence must be distinguished. The first is technological substitution: Chinese companies deploy domestic accelerators, train models with alternative software systems and serve domestic customers through local cloud infrastructure. The second is financial insulation: their company valuations, funding conditions and expected profits are unaffected by foreign market developments.
Evidence supporting the first does not establish the second.
Even a domestic semiconductor ecosystem consumes imported equipment, materials, components or intellectual property at different stages of production. The precise exposure varies by technology and supplier. Industrial policy can accelerate substitution, but qualification, manufacturing yield and supply-chain complexity still influence the rate of progress.
There is also a potential competitive cost to complete separation. Parallel infrastructure stacks require duplicated engineering, software development, manufacturing capacity and customer-support systems. Domestic scale can justify those expenditures, particularly where security and sovereignty matter, but duplication does not guarantee attractive private-sector returns.
The September 2026 industrial data show why China’s longer-term AI ambition remains credible. The National Bureau of Statistics reported that output of integrated circuits increased 18.4% year on year in September, while value added in computer, communications and other electronic-equipment manufacturing rose 16.4%.
These are broad industrial indicators. Integrated-circuit output includes products outside advanced AI compute, and higher unit production does not establish competitiveness in frontier accelerators. The figures nonetheless demonstrate continued expansion in relevant industrial capabilities during a period of financial-market weakness.
Technological progress and share-price performance can diverge for extended periods. That possibility is particularly relevant when valuations previously incorporated expectations of rapid growth, sustained scarcity and improving margins.
Three tests that would change the assessment
The next stage of the debate will depend on evidence from operating companies rather than further declarations of technological independence.
1. Commercial qualification of Chinese memory technology
A major change would be independently supported evidence that Chinese memory suppliers are shipping advanced server products at meaningful scale, with stable yields and repeat customer orders. For HBM, the relevant evidence would include successful accelerator qualification and commercial production economics.
Such progress would strengthen the argument that Chinese competitors can alter the global memory industry’s expected pricing power. Product announcements without supporting volume and qualification evidence would be a weaker signal.
2. Global AI spending and the order books of Chinese suppliers
A sustained reduction in hyperscaler capital-spending guidance, followed by weaker orders or margins at internationally exposed Chinese hardware suppliers, would provide direct evidence of demand transmission.
The opposite pattern would also be informative. If Chinese suppliers maintained order growth and acceptable margins while global infrastructure spending slowed, the case for more independent sources of demand would strengthen.
Investors will need to distinguish revenue booked from existing orders from new orders that determine future factory utilization. Backlog quality, customer concentration and pricing are relevant alongside headline growth.
3. Domestic AI monetization and capital efficiency
For China’s major cloud and model platforms, the strongest evidence would be sustained growth in AI-related revenue accompanied by improving operating cash generation and returns on the associated computing investments.
The amount of new capacity installed matters less if each successive expansion requires an increasing amount of capital without generating proportional incremental earnings.
Similarly, lower inference costs would support the long-term adoption case only if increased usage, higher-value applications or improved customer economics could offset pressure on unit prices.
The observable divide between the two markets is increasingly technological. Their financial separation remains much less convincing. Whether China’s growing AI capabilities eventually intensify global competition or generate mostly domestic returns will be determined by qualified product shipments, international customer exposure and the cash economics of AI deployment.
Sources & Methodology
Data as of October 25, 2026.
This analysis combines company financial disclosures, official Chinese industrial data, US export-control documentation, independent data-center research and contemporaneous market reporting. Data-center capacity figures are third-party estimates, while government computing targets describe planned outcomes rather than delivered infrastructure. Reported market declines establish observed co-movement, not statistically identified causation; segment exposure assessments are qualitative and should not be interpreted as company-level revenue shares or measured stock-return correlations.
Primary sources
- Alibaba Group, “Alibaba’s Full-Stack AI Accelerates Monetization with 22-Quarter-High Cloud Growth,” August 20, 2026
- Tencent Holdings, “Tencent Announces 2026 Second Quarter Results,” August 12, 2026
- Nvidia, Form 10-Q for the fiscal quarter ended July 26, 2026
- US Bureau of Industry and Security, “Revision to License Review Policy for Advanced Computing Commodities,” January 15, 2026
- National Bureau of Statistics of China, “Industrial Production Operation in September 2026,” October 20, 2026
- Chinese Government / Xinhua, “China Releases 5-Year Development Plan for Information, Communications Sector,” September 7, 2026
Other sources
- SemiAnalysis, “The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model,” September 25, 2026
- Reuters via Business Recorder, “China Blue-Chip Index Confirms Bear Market as AI Stocks Slump,” October 23, 2026
- Reuters via Business Recorder, “China’s Blue-Chip Stocks Hit Over One-Year Low on AI-Linked Supply Chain Selloff,” October 9, 2026
- Reuters via MarketScreener, “US Chip Stocks Extend Slide as China Memory Report, AI Spending Jitters Hit Sector,” October 13, 2026
Background reading
- Brokerage research from Meritz Securities (October 2026) and Shinhan Securities (October 2026), used for context, sector identification and hypothesis development only; no proprietary figures, text or charts reproduced.