US think tank Rhodium Group breaks down China's AI financing: model ARR about $10.7 billion, less than one-tenth of OpenAI and Anthropic combined
US think tank Rhodium Group estimates in its report "Examining China's AI Financing" that the combined ARR of Chinese AI models is about $10.7 billion, of which ByteDance accounts for about $4 billion, Alibaba about $2.4 billion, Moonshot about $1 billion, MiniMax about $800 million, and DeepSeek about $500 million, while OpenAI is at $40 billion and Anthropic at $65 billion. The report notes that Chinese models are generally open-source and lower-priced, with value more dispersed across cloud services, advertising, chips, and overseas revenue; for example, Tencent's Q2 marketing services revenue rose 22% to 43.6 billion yuan and was attributed to AI. On financing, major US companies have turned to the bond market, while major Chinese companies rely on operating cash flow and bank loans, and model companies rely on PE/VC and IPOs; Moonshot's valuation rose from $4 billion at the end of 2025 to $50 billion in August this year, and DeepSeek's valuation-to-ARR ratio is about 163 times.
Why read this
AI score 58
The original article breaks down the differences in AI revenue, pricing, and financing structures between China and the US, and notes that the value of Chinese AI is dispersed across cloud, advertising, and overseas markets.
Topics
Body · AI translation
[[**Chinese AI model ARR is about $10.7 billion, less than one-tenth of OpenAI and Anthropic combined; token consumption is large but hard to monetize, and both China and the US are betting funds far exceeding cash flow, with returns not yet proven.**
**Key points**
**1. China's AI revenue is only about one-tenth of the US's**
Chinese models' combined ARR is about $10.7 billion, versus OpenAI's $40 billion and Anthropic's $65 billion; in March this year, national daily average tokens reached 140 trillion, with Doubao exceeding 180 trillion, but low prices and free usage make it hard for revenue to cover scale.
**2. Chinese AI monetizes indirectly through open source and ecosystems**
Alibaba open-sources Qwen to sell cloud, ByteDance connects to Douyin advertising and e-commerce, Tencent's AI gains are hidden in advertising, and DeepSeek relies on High-Flyer's A-share trading; the model layer is hard to rent-seek, and value is scattered across cloud, chips, and overseas.
**3. China and the US diverge in financing bets**
US big tech cash flow is declining and turning to the bond market, while Chinese big tech relies on operating cash flow and bank loans, and model companies rely on PE/VC, IPOs, and the stock market; Moonshot's valuation reached $50 billion, and DeepSeek's valuation/ARR is about 163x.
**4. Profits and underlying risks remain large**
80% of OpenAI and Anthropic's enterprise revenue comes from the top 1% of customers, and enterprises are reportedly switching to cheaper open-source models; China's cloud and model gross margins are low, 54.3% of computing power projects involve financial leasing, and AI will account for only 1.5%-2% of China's total investment by 2027. 2026-11-03 09:16 A US think tank did the math on China's AI and found some interesting insights... TOP Innovation District Research Institute quick overview
This article is from the WeChat public account: TOP Innovation District Research Institute, author: AI Research Group
Previously, the US research institution Rhodium Group released a report on China's AI financing titled "Examining China's AI Financing," which is well worth reading closely.
For example, Doubao has more than 200 million users every day. But its daily revenue is less than 1 million yuan, and mainly e-commerce commissions. Douyin, from the same company, has about 750 million to 800 million domestic daily active users, and its daily advertising revenue is estimated at more than 1 billion yuan.
The gap between them is more than a thousandfold.
The annualized recurring revenue (ARR) of all Chinese AI models totals about $10.7 billion—ByteDance about $4 billion, Alibaba about $2.4 billion, Moonshot about $1 billion, MiniMax about $800 million, DeepSeek about $500 million; by comparison, OpenAI is $40 billion and Anthropic is $65 billion, and the two US giants easily exceed $100 billion.
Numerically, this conclusion holds up.
Why can revenue differ so much?
According to estimates by Artificial Analysis, after price increases, Chinese frontier models mostly cost $0.04 to $0.5 to do a standard task, with the best few at $0.5 to $1; top US Claude costs $2 to $4, and high-end GPT costs $1 to $2.
For the same dish, a Chinese restaurant often charges only a fraction, or even a few tenths, of the price.
Both revenue and price differ by an order of magnitude.
Tokens are like the "characters" in the model's "per-character billing." According to the National Data Administration, in March this year the national daily average token consumption reached 140 trillion, compared with only 100 billion in early 2024. In June, Doubao alone exceeded 180 trillion daily calls, and Volcano Engine's share of China's public cloud large-model service market by token reached 49.5%. But its daily revenue is less than 1 million yuan.
Chinese large models generally bet on open weights, with models made public so anyone can download and run them, like a chef posting the recipe at the door.
When third-party cloud vendors take them to deploy and sell APIs, it becomes hard to collect rent at the model layer, which is tantamount to openly admitting:
US labs roughly share one cash register: consumer subscriptions plus enterprise APIs. ChatGPT Plus, Claude Pro, Gemini Advanced, plus token-billed interfaces. With the same exit, they increasingly look alike.
China does not have this exit. Consumers have been accustomed to free services for decades and are probably the world's best at milking benefits.
The wool comes from the pig, so each company has to find its own "pig," and business models have been reverse-engineered into different shapes.
Alibaba's money is in the cloud, so Qwen must be open source. The model is given away, like free samples at a supermarket entrance, while computing power and inference are charged for, and the math works.
ByteDance makes money from advertising, e-commerce, and content, so the video generation model Seedance is more important than the language model Seed, because it plugs directly into Douyin's distribution and transactions. Whether it ranks first does not matter; what matters is being closest to the cash register.
Zhipu relies on government and enterprise customers and the capital market, so it adopts the MIT license (the most permissive open-source license, allowing commercial use) plus a compliance narrative.
MiniMax bets on multimodal consumer products and the capital market.
DeepSeek's ATM is in A-shares: the High-Flyer Quant behind it makes money through trading. So it is more like a research institution, daring to pursue aggressive architectures others dare not. People without KPIs often go the farthest.
Tencent is more direct: the money AI earns sits right in advertising revenue: second-quarter marketing services revenue rose 22% to 43.6 billion yuan, which the company attributed to AI upgrades continuously improving the advertising effectiveness and pricing of the WeChat ecosystem. This money is not on any ARR chart, but it is indeed earned by AI.
At the developer conference in May, Baidu's Robin Li said that AI first truly took off thanks to applications, that the protagonists of the next stage are agents and applications, and that the weight of model subscriptions is declining.
After the national token data was published, the combined market value of AI chip companies such as Cambricon and Hygon at one point exceeded 2.5 trillion yuan.
But Rhodium notes that some of the ledgers "elsewhere" are thinning on their own.
In 2025, 57% of several cloud giants' revenue was tied to domestic consumption, coming from e-commerce, entertainment and advertising.
Alibaba's e-commerce group saw adjusted EBITA (roughly equivalent to core-business profit) fall 44% in fiscal 2026, mainly because of price wars with JD.com and Meituan. In the second quarter, Alibaba broke out its AI applications segment separately for the first time; revenue for the quarter was 3 billion yuan, with adjusted EBITA a loss of 14 billion yuan.
The free cash flow (the money left in their pockets after subtracting investment) of Alibaba, Tencent and Baidu went from 170 billion yuan in 2025 to negative 16 billion yuan in the first half of 2026.
So they still have to go overseas.
Since the end of 2025, more than half of Moonshot AI's revenue has come from overseas users; as of March, 75% of Kuaishou Kling's ARR came from overseas customers; and in the first half of this year, 61% of MiniMax's revenue came from overseas.
In fact, Chinese manufacturing has walked the same road:
build scale and squeeze costs at home, earn profits abroad.
Moonshot AI is negotiating with Microsoft, Amazon and Google, hoping to take up to 30% of the revenue from Kimi K3-related services on the three clouds. The talks are still early; the revenue-share ratio, data access and how token usage is audited have not been agreed.
They can negotiate because K3 has 2.8 trillion parameters and its weights are public, yet the vast majority of enterprises cannot run it and must rely on big cloud providers for distribution.
Rhodium warns that if the cloud providers are Chinese companies, the revenue share is merely a redistribution of existing revenue among enterprises; if the money comes from overseas customers, the labs become more dependent on external demand, and geopolitical and compliance risks rise accordingly.
This overseas path is much narrower than the one solar and electric vehicles took back then, and there are far more people standing on both sides of the road.
The financing ledgers on both sides
OK, now back to the question we were pursuing:
For Microsoft, Amazon, Google, Meta and Oracle in the United States, free cash flow fell from $191 billion last year to $14 billion in the first half of this year, while net debt financing over the same period was $163 billion, mainly long-term bonds. When deposits run dry, they borrow long money.
But there are risks, too.
According to Morgan Stanley estimates, capital expenditure at big US companies has already eaten up nearly 100% of operating cash flow, versus an average of only about 40% over the past decade. Global AI-related bond issuance is expected to approach $570 billion this year.
It is like a family spending its entire salary on renovations and still having to swipe the credit card.
And the bond buyers are starting to hesitate. According to statistics from Apollo chief economist Torsten Slok, for every $1 of bonds a big company issues, it received nearly $5 in subscriptions in February, but that fell to less than $2 in July. In July, for a $25 billion Amazon bond issue, the longest-maturity tranche had to offer 18 to 21 basis points more in interest before it could be sold.
China relies mainly on operating cash flow, equity financing and bank loans, and hardly uses bonds.
This year, ByteDance, which has the largest increase in capital expenditure, relied mainly on overseas loans; according to reports, in September it obtained a $29.6 billion overseas syndicated loan from about 30 banks—a group of banks pooling money to lend to one company.
Large model companies rely on the stock market. Zhipu listed on the Hong Kong Stock Exchange on January 8, 2026, becoming "the world's first listed large-model company"; MiniMax followed on January 9, more than doubling on its first day.
While Chinese model companies line up to ring the bell, the US leaders are in no hurry.
In a September interview with Fortune, Altman confirmed that OpenAI will not go public this year, citing the current AI safety situation making the timing unsuitable and the company not feeling financial pressure.
But you may not believe it: money in China is very cheap.
In 2025, the median long-term RMB loan rate for Tencent was 2.8%, and the coupon rates on Tencent's and Baidu's long-term RMB bonds were about 2.6%. By comparison, the average effective rate on long-term bonds of the five major US cloud providers is about 4.4%.
With money this cheap, why don't Chinese big companies issue bonds on a large scale like their US peers?
In fact, they are borrowing too, just from banks.
One reason given by Rhodium: bank loans require less information disclosure, and in the banking system data centers are a very popular form of collateral.
Issuing bonds is like borrowing money publicly from the whole street—you have to lay your finances bare for all to see; going to a bank means talking behind closed doors, and pledging a data center seals the deal.
They don't have the cash flow of the big companies, nor do they hold many hard assets to pledge, so the money can only come from equity: PE/VC in the primary market, and IPOs and private placements in the secondary market.
Zhipu, MiniMax, DeepSeek and Moonshot AI together received 9 billion yuan in PE/VC investment in 2025; in the first eight months of 2026, PE/VC, IPOs and private placements combined rose to 179 billion yuan.
That is roughly a twentyfold increase in a year.
Valuations have risen even faster. Moonshot AI's valuation rose from $4 billion at the end of 2025 to $50 billion in August this year.
Dividing valuation by ARR is like asking how many yuan the market is willing to pay for a shop with 1 yuan of annual revenue: OpenAI is about 34 times, Anthropic about 21 times, Zhipu about 46 times, Moonshot AI about 50 times, and DeepSeek as high as 163 times.
Kimi and DeepSeek are not yet listed, but their money is still tied to the stock market. The reason private investors are willing to give Moonshot AI such a high valuation is that they are looking at how much it could sell for once it goes public; DeepSeek's money, as mentioned earlier, comes from High-Flyer's quantitative trading in the A-share market.
Rhodium Group partner Logan Wright told CNBC that China's frontier labs will be highly dependent on a favorable stock market. The trouble is, the stock market's mood is unpredictable.
Once you break it down to this level, the cards on both sides become clear.
American big tech is betting on the bond market,
while Chinese big tech relies on its own cash flow and banks—but as mentioned earlier, the consumer ledger supporting that cash flow is thinning.
First, look at that long bar for the US: its base is actually not wide.
Ramp's consumption data shows that 80% of enterprise revenue for both OpenAI and Anthropic comes from their top 1% of customers; nearly a quarter of Anthropic's revenue comes from the two coding tools Cursor and GitHub Copilot.
It's like a restaurant where 80% of the turnover depends on a few tables of regulars—except these tables are the best at haggling and the most motivated to switch to something cheaper. The New York Times reports that companies such as AT&T, Deloitte and Airbnb are increasingly using cheaper open-source models.
This is exactly where Chinese open-source models have their best opportunity.
Profit margins are hardly ample either. Gross margin is like selling a bowl of noodles: after deducting the noodles and toppings, how much is left. OpenAI's gross margin in the first quarter was reportedly only 39%, while Anthropic's was around 65%.
Now look at China.
Technical efficiency is in no way inferior. DeepSeek's API gross margin is reportedly 83%, close to Anthropic's 80%-plus, yet its pricing is much lower.
What holds it back is pricing power.
Alibaba Cloud's adjusted EBITA margin was 9% in 2025 and 10% in the first half of this year; the cloud businesses of Amazon and Google had operating margins of 39% and 34% respectively in the first half of this year. Alibaba CEO Eddie Wu believes Alibaba Cloud's margin can rise to 20%, but even so, it would still be lower than major US cloud providers.
The gross margins of Zhipu and MiniMax fell from 41% and 25% in 2025 to 26% and 18% in the first half of this year.
Revenue is rising, margins are falling.
So this year, everyone has started trying to raise prices.
Since the beginning of the year, most labs and cloud providers have raised prices. Around May, Doubao tested three subscription tiers—68 yuan, 200 yuan and 500 yuan—on the App Store, and public opinion immediately exploded. No one bothered to calculate whether it was worth it; the first reaction was: will the free version be deliberately made worse?
In August, DeepSeek also raised prices and introduced peak-hour pricing, like surge pricing for ride-hailing during the evening rush.
This is the first time the Chinese market has seriously tested: how much extra are users actually willing to pay for "smart"?
Finally, let's talk about the parks.
Since 2023, 47% of equity investment in AI chips and servers has come from government guidance funds and investment platforms, 14% from predominantly state-owned banks, and state-related sources combined account for more than 60%. In AI cloud, direct government investment accounts for about 30%. Financing for large models and AI applications, by contrast, is mainly private capital.
The reason is simple: chips and servers are hard assets that can be pledged for loans and counted in statistics; the state also gives strategic priority to supporting hard tech and bottleneck areas.
Risks are gradually accumulating in financial leasing. Data from the China Academy of Information and Communications Technology shows that 54.3% of newly added computing center projects in the first half of this year involved financial leasing; in the first quarter, the computing power leasing market grew 62% year on year to 68 billion yuan.
A company in Sichuan used sale-and-leaseback plus leverage to buy GPUs. Sale-and-leaseback is like selling your own car to a dealership and then leasing it back to drive, extracting a lump of cash. After it went bankrupt, 15 licensed leasing companies were directly hurt in the second quarter.
GPUs are different from factories and land: different models depreciate at very different rates and face very different market demand. Like second-hand phones, no one can say for sure how much a batch of cards will be worth after two or three years.
In August 2025, Runze Technology and GDS each issued China's first publicly offered data center REITs on the Shenzhen and Shanghai stock exchanges respectively, slicing the rental income from server rooms into small pieces and selling them to investors.
The strongest card the parks hold is electricity.
In 2024, China added about 429 gigawatts of power generation capacity, versus about 51 gigawatts in the US. Bank of America research believes China's industrial electricity prices are 30% to 60% lower than those in developed markets, and its power grid is relatively young.
Still, the gap in the underlying base remains stark.
According to the Stanford AI Index, in 2025 the US had about 5,427 data centers, while China had about 449.
Chips and memory remain China's hard constraints.
The US lacks power; China lacks chips.
Cheap electricity and large tracts of land are the parks' real advantages. To turn them into revenue, two hurdles remain: obtainable chips, and demand willing to pay. Clear both, and the server room is a production line; fail, and the server room is just a piece of collateral on a financial leasing contract.
Still, even by 2027, the entire AI industry will account for only about 1.5% to 2% of China's total investment, far from enough to offset the shock from the slowdown in real estate and infrastructure.
The top part of Rhodium's chart records the money OpenAI and Anthropic collect directly from customers. In the US, the money almost all flows through the single outlet of subscriptions plus APIs, so it is recorded clearly.
The small colored bars below record only a small portion of the value of Chinese AI that can be directly priced. The rest is scattered across WeChat ads, Alibaba Cloud bills, chip companies' market values, electricity meters in western server rooms, and invoices to overseas customers.
What Rhodium calculates is the portion recorded under model companies' names.
The way bets are placed is different too.
The US bets with bonds and the budgets of a few big customers; China bets with bank loans, the stock market and state capital. Both sides believe today's investment will turn into future cash flow, and neither has yet proved it.
Those few small colored bars, looking back in a few years,
may be a seed, or they may be a ceiling. AI industry signal channel: Frontier Tech. This content comes from the internet, original link; the views represent only the author and do not represent Huxiu's position. For copyright issues, please contact hezuo@huxiu.com, and we will verify and handle it promptly. Everyone who is changing the world and wants to change it is on the Huxiu APP[ ]