Chinese technology companies expect to spend 932 billion yuan ($139 billion) on artificial intelligence infrastructure in 2026, up 103% from the previous year. But revenue from AI labs would cover less than 8% of that spending, according to Rhodium Group: expansion is moving much faster than the business that should sustain it.
The estimate comes from a Rhodium Group analysis of 13 publicly traded Chinese companies, including major cloud providers, telecom operators, labs, and independent data centers. Planned spending would exceed 1.2 trillion yuan ($193 billion) in 2027. Nevertheless, that remains a fraction of US investment in data centers, which Rhodium estimates at around $800 billion in 2026.
Scale matters, but so do who can pay for it and what revenue they have. The report puts annual recurring revenue (ARR) from Chinese AI models at around $10.7 billion, roughly one-tenth of the levels OpenAI and Anthropic have reported combined. The contrast does not, by itself, describe each company's profitability, but it does show the gap between the scale of spending and the business directly associated with the models.
A major investment, with cash under pressure
The three major cloud companies analyzed by Rhodium—Alibaba, Tencent, and Baidu—went from generating 170 billion yuan in free cash flow in 2025 to recording a negative balance of 16 billion in the first half of 2026. AI investment is not the only cause of that pressure: operating cash flows also matter, and these companies' traditional businesses depend in part on Chinese household consumption.
The shift is significant because those businesses—such as e-commerce and advertising—have helped finance technological expansion. Rhodium warns that slowing consumption could reduce the profits and cash available to provide that cross-subsidy. In its view, the question is not only how much companies plan to spend, but how long they can cover the gap with businesses outside AI.
The tension is not unique to China. Rhodium estimates that combined free cash flow at Microsoft, Amazon, Alphabet, Meta, and Oracle fell from $191 billion in 2025 to $14 billion in the first half of 2026. But the financing route differs: major US tech companies relied more on debt, especially long-term bonds, while Chinese companies depend relatively more on equity and bank loans.
| Indicator | China | United States |
|---|---|---|
| Investment in AI infrastructure, 2026 | 932 billion yuan ($139 billion); Rhodium forecast | Around $800 billion in data center investment; Rhodium estimate |
| Scale of investment | 15% to 20% of the US level, according to Rhodium | Rhodium's benchmark for comparison |
| Free cash flow of the major companies analyzed | Alibaba, Tencent, and Baidu: −16 billion yuan in the first half of 2026 | Microsoft, Amazon, Alphabet, Meta, and Oracle: $14 billion in the first half of 2026 |
| Financing gaining importance | Equity and loans | Debt, mainly long-term bonds |
Fewer chips are not the only bottleneck
The availability of advanced accelerators has shaped the debate about Chinese AI, particularly because of US export restrictions. Rhodium says difficulties obtaining chips limited some companies' spending, including Tencent's, during 2025. It also notes that domestic supplies increased and access to Nvidia H200 chips appeared to ease, although new tensions between Washington and Beijing could close off access.
However, Logan Wright, a Rhodium partner and co-author of the report, believes financial constraints could weigh more heavily in the long term. That shifts the focus: access to computing capacity is not enough if revenue, cash, or available capital cannot sustain investment. Rhodium puts China's AI infrastructure development at 15%-20% of the US level, even with the strong growth projected.
Dependence on external financing is already apparent in the differences between types of companies. Frontier labs have negative operating cash flows because they invest heavily in research and development and their business models remain immature. So far, much of their computing has come from renting capacity from cloud providers; if they build their own data centers, a larger share of spending will be reflected as capital investment.
Venture capital and equity placements have gained momentum. Rhodium counts 9 billion yuan in private equity and venture capital investment in Zhipu AI, MiniMax, DeepSeek, and Moonshot during 2025; in the first eight months of 2026, investments, IPOs, and private placements totaled 179 billion. The firm also notes that Chinese companies have not yet turned to bonds in the way their US counterparts have.
Open models: more adoption, less revenue capture
China's bet on open-weight models makes it easier for third parties to deploy them, but can make it harder for their creators to charge for that use. If a cloud provider offers access to a lab's model, it does not necessarily have to pay the developer. Rhodium notes that Moonshot and Alibaba are trying to negotiate revenue-sharing agreements with major users of their open models.
Revenue sharing, if it happens, changes who receives the money, but does not always expand the market: when the provider and lab are Chinese, it may simply redistribute existing revenue. If demand comes from abroad, exposure to international customers and regulatory conditions increases. Rhodium says more than half of Moonshot's revenue came from foreign users as of late 2025; it also attributes 75% international ARR through March to Kuaishou's Kling, and 61% of MiniMax's revenue in the first half of 2026 to foreign users.
Price competition adds further pressure. TNW reports that Z.ai cut the price of GLM-5.3 Flash in half for two weeks to attract users to its own servers and that, in November, around 20 rival providers offered its model at a lower price, according to OpenRouter data cited by the outlet. For developers selling APIs, low prices can make testing and deployment easier, but make it harder to turn usage into sufficient margins to fund training and capacity.
Exceptions show that efficiency and the profitability of a service are not the same as the solvency of an entire company. Rhodium reports that DeepSeek reached an 83% gross margin on its API service in July, comparable to the above-80% level attributed to Anthropic. By contrast, the report cites a 25% margin for Zhipu's open platform and API in the first half of the year, and a 45% margin for Moonshot's API service the previous December.
“We simply can't afford to train the largest models available right now.”
— Liang Wenfeng, CEO of DeepSeek
The bet depends on financing that is not yet secured
Liang made that comment while discussing the capabilities of Huawei chips, according to a leaked transcript of a conversation with investors cited by The New York Times and reported by TNW. He said Huawei chips were approximately two years behind Nvidia's and that it would take four Huawei 950 chips to match one Nvidia GB300. In a market where computing is already expensive, that difference also means greater pressure on the capital needed to train cutting-edge models.
The cost of equipment reflects that scarcity: sellers on Goofish, a Chinese secondhand marketplace, were listing Nvidia DGX B300 servers for nearly $3 million, close to eight times the US price, according to The New York Times calculation reproduced by TNW. These are marketplace listings, not a uniform price, but they illustrate how access to hardware can make expansion plans more expensive.
Public financing could help ease some of the gap, although Rhodium says state investment will probably continue to focus more on chips than on frontier labs. The report also warns that continued expansion will depend on stock market conditions and support from state-guided public funds. Adding to that uncertainty, private investment in Chinese AI totaled $12.4 billion in 2025, compared with $285.9 billion in the United States, according to Stanford's AI Index cited by TNW; the index itself warns that the Chinese figure may understate the total because it does not fully include those state funds.
For companies building products on these models, the dilemma is not simply whether to choose a cheap API or an open model. Rhodium documents that Chinese labs depend on equity financing and that some are negotiating revenue-sharing agreements for the use of their open models. Rhodium forecasts that infrastructure spending will exceed 1.2 trillion yuan in 2027; the financial question is whether cash and capital can keep pace.
Sources: TNW | China, Rhodium Group