China's biggest technology companies are expected to double capital spending on AI infrastructure this year to 932 billion yuan ($139 billion), while the recurring revenue of the country's AI model businesses totals roughly $10 billion, according to a Rhodium Group report that the New York Times highlighted on November 20. The revenue-to-spend ratio is under 8%, a financing gap the research firm says may constrain China's AI buildout more than chip export controls do.
The comparison with the United States is unflattering on both sides of the Pacific. Rhodium estimates US data center investment at around $800 billion this year, putting China's buildout at 15-20% of the American level. The gap in monetization is wider: Rhodium puts the combined annual recurring revenue of all Chinese AI models at about $10 billion, roughly 10% of the recently reported ARR levels of OpenAI and Anthropic.
The report lands the week Anthropic raised more than $80 billion in what the NYT framed as a record IPO, underscoring how differently the two ecosystems fund the same race.

Cash flow went negative first
The spending surge is arriving just as China's hyperscalers stop generating the cash to cover it. Alibaba, Tencent and Baidu posted combined free cash flow of 170 billion yuan in 2025. In the first half of 2026 the same three firms were at negative 16 billion yuan, per Rhodium's calculation of operating cash flow minus capex.
The pattern mirrors the US, where the combined free cash flow of Microsoft, Amazon, Alphabet, Meta and Oracle fell from $191 billion last year to $14 billion in H1 2026. The response differs. US hyperscalers turned to the bond market, raising a net $163 billion in H1 2026 through debt, up from $90 billion in all of 2025. Chinese firms barely use bonds at all. Their marginal financing so far has come from equity issuance and bank loans.
That makes the buildout hostage to stock market appetite in a way the American one is not.
Why the revenue is structurally low
Two choices define Chinese AI economics: open weights and low prices.
Most Chinese frontier labs ship open-weight models. A third-party cloud provider can deploy Kimi or Qwen and sell API access without paying the lab that built the model. Moonshot is in talks with Microsoft, Amazon and Google about revenue-sharing arrangements that could reach 30% of serving revenue, but Rhodium notes that when the cloud partner is Chinese, the arrangement just reallocates existing revenue rather than creating new revenue.
Pricing reflects the same squeeze. Citing Artificial Analysis, Rhodium reports Chinese frontier models generally cost $0.04 to $0.5 per task, with the best around $0.5 to $1. Claude-class US models run $2 to $4 per task. Even after most Chinese labs raised prices in August, including DeepSeek, the gap persists.
The margins show it. Zhipu's gross margin fell from 41% in 2025 to 26% in H1 2026; MiniMax's fell from 25% to 18%, even as their API revenues grew. Zhipu's model-as-a-service ARR grew 20-fold, from $74 million in January to $1.6 billion in August, which is fast growth off a small base. Alibaba Cloud's adjusted EBITA margin reached 10% in H1 2026 against operating margins of 39% at Amazon and 34% at Google Cloud.
Consumer apps, the other hoped-for monetization path, are losses so far. Alibaba's newly separated AI app segment produced 3 billion yuan of revenue and a 14 billion yuan adjusted EBITA loss in Q2. Rhodium cites the ByteDance datapoint: Doubao has more than 200 million daily users and takes in under one million yuan a day, mostly e-commerce commissions. Douyin, by comparison, generates implied daily revenue above one billion yuan from roughly 750-800 million users.
The equity window is doing the work
Where bonds do not reach, equity has. Announced equity investment in China's AI sector hit a record 282 billion yuan through August 21, more than double the 2023-2025 lows, according to Rhodium. Frontier labs alone tapped 179 billion yuan in the first eight months of 2026, against 9 billion yuan of PE/VC money in all of 2025. Zhipu raised about 4 billion yuan in its January Hong Kong IPO and a further 27 billion yuan in a July H-share placement. Alibaba announced in August an HK$80 billion new share placement, its first since 2019, with 100% of proceeds earmarked for AI.

The valuations riding on that window look stretched by the same report's own numbers. Zhipu trades at roughly 46x ARR against 34x for OpenAI and 21x for Anthropic. DeepSeek and Moonshot sit at an estimated 163x and 167x. Moonshot's reported valuation went from $4 billion at the end of 2025 to $50 billion by August. IPOs for Moonshot and DeepSeek are pending.
State money is in the picture but is not backstopping the labs. Government guidance funds and state platforms accounted for 25% of announced AI equity investment this year, down from 30% in 2024-2025, and concentrated in chips and servers, where more than 60% of equity funding since 2023 traces to state-affiliated sources. Rhodium expects state-led investment to keep favoring hardware over frontier labs.
The constraint that replaces the chip curb
For most of 2025 the binding constraint on Chinese AI capex was compute: Tencent's spending sagged as export controls choked its chip access. Rhodium's read for 2026 is that compute is loosening, with rising SMIC domestic supply and Nvidia H200 shipments resuming from August, while ByteDance's upgraded capex plan is the largest marginal contributor to this year's increase.
What takes the place of the chip shortage is money. The report's outlook section is blunt about the dependency: continued expansion likely depends on equity market conditions, and the stabilization of China's stock market matters because a cooling market closes the main financing channel. Rhodium points to the post-listing collapse of robotics firm Unitree's shares within a week of its debut as the kind of event domestic investors are watching.
Two smaller tail risks round it out. Financial leases now touch 54.3% of newly added compute center projects in H1 2026, and the bankruptcy of one leveraged GPU lessor in Sichuan already cost 15 licensed leasing companies direct losses. And Rhodium flags the rising risk that renewed US-China tension closes off both chips and the remote-access loopholes Chinese firms have used to rent overseas GPU capacity.
For the Chinese economy overall the stakes are modest: Rhodium estimates AI investment at only 1.5-2% of national fixed investment by 2027, too small to offset the property slowdown. For the US-China AI rivalry, the report reframes the question. The bottleneck is moving from who can get GPUs to who can fund them.
Sources
- New York Times: China's A.I. Ambitions Are Running Into a Money Problem (Nov 20, 2026) - nytimes.com
- Rhodium Group: Examining China's AI Financing (Sep 10, 2026) - rhg.com
- Techmeme pick: Rhodium on China capex and the AI lab revenue gap (Nov 21, 2026) - techmeme.com
