Executive Summary

The report is called "Examining China's AI Financing." The word revenue is not in the title. Of the four key findings set out at the top, only one is about revenue. The other three say that China's AI investment runs at 15 to 20 percent of the American level, that funding constraints now sit alongside chip shortages in holding back new investment, and that Chinese AI firms lean on equity and bank loans rather than bonds. The item the press put in its headlines was one of those four.

Start with the scale. Rhodium takes thirteen listed Chinese companies as its sample and estimates that their AI infrastructure investment grows 103 percent this year to 932 billion yuan, roughly $139 billion, and passes 1.2 trillion yuan, roughly $193 billion, in 2027. The report does not state that it used a different exchange rate for the two years, so the yuan figures are the safer ones to read first. US data center investment for the same year is estimated at about $800 billion. By growth rate the Chinese number is explosive; by absolute size it is about a fifth. Both sentences are true of the same table. The report then shrinks this scale once more against the Chinese economy as a whole. Even in 2027, it says, the entire AI sector amounts to 1.5 to 2 percent of total Chinese investment, and a smaller share of GDP than that. Too small, in other words, to fill the hole left by cooling property and infrastructure investment.

Money is going out faster than it comes in. Free cash flow at Alibaba, Tencent and Baidu combined fell from 170 billion yuan in 2025 to negative 16 billion yuan in the first half of 2026. What that free cash flow figure subtracts is spelled out in a footnote: operating cash flow minus capital expenditure, with Tencent's media content spending and lease liability repayments excluded from the calculation. Tencent's own second-quarter operating cash flow included roughly 51 billion yuan of AI-related prepayments, a sum close to that quarter's 59 billion yuan of capex. Which line item goes in which column changes the size of the hole. And the hole is not a Chinese phenomenon. Over the same period, combined free cash flow at Microsoft, Amazon, Alphabet, Meta and Oracle dropped from $191 billion to $14 billion. What differs is not the size of the gap but the instrument used to fill it. The five American firms raised net debt from $90 billion to $163 billion, most of it long-dated bonds. China did not go to the bond market.

The report lines up the funding channels the Chinese AI sector uses, largest first. Cash generated by operations comes first, then equity issuance, then loans and bonds, with asset-backed securities, REITs and finance leases at the end. If you isolate the money arriving now, the order changes. At the margin, the report says, the money comes from equity issuance and new lending.

The loans and bonds row deserves a second look, because bonds are used sparingly for reasons that have nothing to do with their cost. The rates the report sets side by side point the other way. Long-dated yuan bond coupons at Tencent and Baidu ran around 2.6 percent in the first half of 2026, while the average effective rate on long-dated bonds at Amazon, Alphabet, Meta, Microsoft and Oracle was about 4.4 percent. Even so, net inflows at the surveyed companies in 2025 were 639 billion yuan from loans against 341 billion yuan from bonds, excluding convertibles. The reason the report gives lies on the issuer side. China's bond market favors state-owned issuers, which come with implicit guarantees and tangible collateral, and state-owned issuance has held above 60 percent of the market since 2022. Even after the central bank and the securities regulator began encouraging technology innovation bonds in May 2025, direct issuance by technology firms stayed limited. The company that raised capex most aggressively this year is ByteDance, and it too went elsewhere: in September it reportedly drew a $29.6 billion offshore syndicated loan from about thirty banks. Still, the report's conclusion does not treat this channel as shut. Hyperscalers and carriers, with ample cash flow and data centers to pledge, should be able to fund more capex with bonds. The ones who will struggle are the frontier labs and the smaller data center operators.

Where the state's money goes is another of the report's findings. Announced equity investment in China's AI sector reached a record 282 billion yuan through August 21, 2026, and the state-linked share of it was 25 percent. That 25 percent is a number that came down, not up. It stood at 30 percent in 2024 and 2025, and fell as private investor appetite grew and a June State Council directive tightened the creation of new government investment funds. County- and district-level governments can no longer establish new funds as a rule. The pre-pandemic average, though, was 13 percent, so the current share is still double that.

Split by segment, the picture changes again. The window has to be stated first here too. The shares below come not from 2026 alone but from a chart covering 2023 through August 21, 2026 combined. Of the equity investment that went into AI chips and servers, 47 percent came from government guidance funds and investment platforms, and banks, most of them state-owned, added another 14 percent. Together that is more than 60 percent. AI cloud is at 30 percent. Large language models and AI applications, by contrast, are dominated by private capital. The note attached to the chart also states that the segments overlap to some degree. The state holds up the hardware, and the companies that build models are left to the stock market.

The far end of the channel list splits in two directions. Asset-backed securities and REITs are a route the regulator designed, so that independent data centers could recycle existing assets and bring down leverage. The securities regulator proposed data center REITs in its sixteen measures to support technology companies in April 2024, and last August Runjiang IDC and GDS issued the first public data center REITs on the Shenzhen and Shanghai exchanges. The report expects this channel to keep growing with policy support. Where things went wrong is the alternative that smaller operators use when they cannot get cleanly into that channel: finance leasing. By the China Academy of Information and Communications Technology's count, 54.3 percent of new computing center projects in the first half of 2026 involved a finance lease, and the first-quarter computing rental market grew 62 percent year on year to 68 billion yuan. Then Sichuan Bingji Technology, which had been buying GPUs through sale-and-leaseback to build leverage, went bankrupt, and fifteen licensed leasing companies took direct losses in the second quarter. Depreciation and market demand vary sharply from one GPU model to the next, the report notes, so further turbulence in the computing market could produce more bankruptcies and defaults.

A sentence from Logan Wright, a co-author of the report, sums the structure up in one line for CNBC. "They will be heavily dependent upon a favorable climate in the equity market—historically that's not an easy bet in China." Government funding, Wright added, has been helpful on the hardware side of the compute buildout, but will probably balk at funding the frontier labs directly. Pebblous has traced where the capital poured into inference serving ended up before. This article asks the question from the other end. Why doesn't that money come back as revenue?

The $10.7 billion is not a revenue figure. It is annual recurring revenue, ARR. CNBC describes the metric as an industry measure that multiplies the most recent month by twelve in order to capture fast growth. Given that definition, it points at something other than the money that actually came in over the past twelve months. Rhodium does not hide this. The body of the report says so directly.

Seven companies go into the sum. Four of them are frontier labs: DeepSeek, Moonshot, Zhipu and MiniMax. To those are added two large incumbents, Alibaba and ByteDance, and Kuaishou, which makes the video generation model Kling. Put each company's figure next to its reference month and the character of the sum becomes visible.

Reference months reconstructed by reading Figure 7 of the Rhodium report against CNBC's coverage. Zhipu's $1.6 billion covers the MaaS business only, as footnote 4 of the report states; on-premises deployment and other revenue are excluded. Listing status splits three ways: Zhipu and MiniMax are AI companies listed in Hong Kong, Kling is a division of a listed conglomerate, and DeepSeek, Moonshot and ByteDance are private at the company level.

Reading. The seven figures in the table add to $10.8 billion. The $10.7 billion printed in the report is a rounding difference. One thing follows from this. Zhipu told investors on September 16 that its latest ARR was $1.8 billion, and substituting that figure would put the total at $11.0 billion, which does not match the printed figure. The $10.7 billion in the headlines holds only when Zhipu is entered at $1.6 billion of August MaaS revenue.

The denominator behind the 10 percent is worth checking too. Dividing $10.7 billion by the combined $105 billion of OpenAI's $40 billion and Anthropic's $65 billion gives 10.2 percent. It is not a comparison against either company on its own. Taken separately, the ratios are 27 percent of OpenAI and 16 percent of Anthropic. This distinction slips easily in the retelling. One Korean-language report turned the same fact into a company-by-company comparison and printed the total as $10.5 billion.

Now put the reference months on a horizontal axis. You can see that the seven figures in the sum are not a photograph of a single moment. Kuaishou Kling's figure stops in March, DeepSeek's is June, ByteDance's is July, and the other four are August. That sum is then set beside OpenAI in August and Anthropic in July.

One more thing needs separating. These seven companies are a set assembled to compare revenue, and the sample whose cash flows the report dissects is a different one: thirteen listed companies, made up of three hyperscalers, Huawei and four telecom carriers, two frontier labs and four independent data centers. The two sets share some members but they are not the same list. That the paragraphs about financing and the paragraphs about revenue stand on different samples is worth re-checking every time this report is cited.

The same property survives in the revenue chart, and Rhodium flagged that one as well. The report puts cloud revenue at hyperscalers and carriers at about 13 percent of total revenue, and a footnote explains that the 13 percent covers Alibaba, Baidu, Huawei and the three telecom carriers, with Tencent and ByteDance excluded because they do not disclose cloud revenue. The note on the revenue-mix chart goes a step further. Tencent's cloud revenue is reported together with fintech, including WeChat Pay, and other business services, so it was classified as mixed, and the proportion of cloud, data center and AI revenue at hyperscalers is therefore understated. Here the original report flags the bias in its own figure.

Interrogating a metric is not the same as dismissing the number. The growth rates the same report gives are remarkable on their own terms. Zhipu's MaaS annual recurring revenue went from $74 million in January to $1.6 billion in August, a twentyfold rise, and MiniMax quadrupled between February and August. On a slope like that, multiplying the most recent month by twelve opens an unusually wide gap. That gap is the subject of the next section. Pebblous has looked at the token usage statistics China publishes at the national level before; where that piece examined usage numbers issued by a government body, this one examines a financial metric investors use to price companies.

Zhipu occupies a special position among the seven. It is listed on the Hong Kong exchange, so it has to publish audited numbers every half year. That means two kinds of revenue figure exist at once for the same company and the same business. One is the $1.6 billion of August MaaS annual recurring revenue that Rhodium used. The other is the 825 million yuan of API and MaaS revenue for the first half of 2026 that Zhipu filed with the exchange.

Taken from Zhipu AI (2513.HK) interim results for 2026. The percentages compare with the same period a year earlier; only the gross margins are stated as absolute levels.

Reading. To measure the two numbers with the same ruler, the numerator and the denominator have to cover the same business. Converting the 825 million yuan of first-half API and MaaS revenue at the exchange rate the report used for 2026 gives about $122.9 million, and annualizing that half-year figure gives about $245.9 million. Dividing the $1.6 billion of August MaaS annual recurring revenue by that comes to roughly 6.5x. That is the distance between two numbers describing the same business at the same company.

The same arithmetic on MiniMax gives a different degree. MiniMax's total revenue in the first half of 2026 was $116.6 million, which annualizes to $233.2 million. Dividing the $800 million August ARR by that gives about 3.4x. This is why Zhipu's 6.5x cannot be copied onto another company. How far the two figures separate depends on where that company currently sits on its curve.

This property flows straight into the multiples in the next section, because ARR is the denominator when valuation is divided by it. When the denominator is set higher than realized revenue, the multiple looks smaller. So far we have been watching what happens at the bottom of that fraction.

What about profitability? Gross margins at both listed AI companies fell. Rhodium puts Zhipu at 41 percent down to 26 percent, and MiniMax at 25 percent down to 18 percent. There is one thing to watch in that comparison. The first figure in each pair is full-year 2025 and the second is the first half of 2026. Zhipu's own results coverage compares the same 26.4 percent against 50.0 percent in the year-earlier half. The trailing number is identical while the baseline period differs, so mixing the two comparisons changes the size of the decline.

Reading that decline purely as decay misses something the report places right next to it. Zhipu shifted weight away from enterprise general-purpose model services, its highest-margin business at 42 percent in the first half, toward the open platform and API side at 25 percent. The reason is in the report: the addressable market for the first business is smaller than for the second. Part of the margin decline is a choice rather than an accident. Over the same period Zhipu acquired the AI infrastructure specialist Xcore Sigma for a one-gigawatt data center built on domestic chips, a move aimed at raising API margins, and the 80 percent drop in token inference cost since January points the same way. There is a number attached to what that move is aiming at. Zhipu's API business margin is reported to be capable of reaching 50 to 60 percent under ideal conditions on its own infrastructure, roughly double the 25 percent the first half delivered. MiniMax turned in the same direction: its enterprise share of ARR rose from 30 percent in August 2025 to 80 percent in August this year.

The sentence from the report that traveled furthest was not the revenue comparison but the multiples. Valuation divided by annual recurring revenue comes to 163x for DeepSeek, 50x for Moonshot and 46x for Zhipu, against 34x for OpenAI and 21x for Anthropic. Nearly every citation of those five numbers carries the same gloss: the valuations in the numerator belong to private companies, cannot be verified, and so neither can the multiples. That premise is wrong. The valuations are drawn as bars in a chart on page 17 of the report, and the axis labels state, company by company, which month each one belongs to.

The chart is titled "Frontier lab valuation and valuation-to-ARR ratio, 2026 latest available by August," and its unit line reads "Billion USD, valuation-to-ARR ratio (RHS)." Bars on the left axis are valuations; dots on the right axis are multiples. The category names along the horizontal axis are more than company names. They read "DeepSeek by Aug · MoonShot by Aug · Zhipu by Aug · MiniMax by Aug · OpenAI by Mar · Anthropic by May." Four Chinese labs as of August, OpenAI as of March, Anthropic as of May. Rhodium marked, in its own chart, that "latest available" stretches as far as five months apart depending on the company.

Reading. Correcting for scale using the positions of the axis gridlines and then reading the bars reproduces all five printed multiples to within 0.1 to 4.3 percent. Measured that way, DeepSeek and Zhipu each stand at roughly $69 billion, Moonshot between $45 and $50 billion, OpenAI at about $849 billion and Anthropic at about $960 billion. The Moonshot bar matches a figure the report states in prose, a jump from $4 billion at the end of 2025 to $50 billion as of August, which makes it a cross-check. The MiniMax bar is only four pixels tall, so its precision is poor and we record it only as being in the tens of billions.

Dividing each bar by its dot recovers the annual recurring revenue that each multiple assumed. For the four Chinese companies the results broadly match the table in the previous section. The two American ones do not. The back-calculation gives about $25 billion for OpenAI and about $48 billion for Anthropic, while the same report's revenue comparison lists $40 billion and $65 billion. This is not an error but a consequence of the reference months. Rhodium paired each company's valuation with ARR from the same point in time. As methodology that is the more careful treatment, and the method note beneath the chart says as much.

What the careful treatment leaves behind is a lag between the things being compared. The four Chinese companies were measured in an August window and the two American ones in March and May windows, and over those five months both the valuations and the revenue of the two American companies rose. It is possible to calculate what happens if the two American companies are re-measured in an August window using their latest reported valuations. The valuations that go into that calculation are not Rhodium's figures, though, but round sizes under discussion and listing targets as relayed by the press, which should be stated first.

Calculated from figures the SCMP relayed from the Financial Times. Neither figure is a settled valuation nor a Rhodium estimate, so they should not be placed on the same footing as the report's own numbers.

The result is simple enough. Align the windows and the two American companies converge around 30x, which narrows the distance to Zhipu's 46x and Moonshot's 50x noticeably. DeepSeek's 163x stays far away. Matching the reference months erases some of the gap and leaves the rest standing.

4.1Two kinds of valuation sit in one row

The chart's source note splits those six values into two kinds. Four are reported prices from private rounds; two are market capitalizations observed daily in Hong Kong. DeepSeek, Moonshot, OpenAI and Anthropic are in the first group, Zhipu and MiniMax in the second. The two kinds behave differently.

The first kind carries a bias that the literature has quantified. Will Gornall and Ilya Strebulaev reconstructed the charter documents of 135 US unicorns and repriced them to reflect the contractual terms attached to preferred shares. The post-money valuations reported in the press ran 48 percent above fair value on average. After adjustment, 65 of the 135 lost unicorn status. The overstatement comes from protections granted to the most recent investors: a guaranteed IPO return in 15 percent of cases, a veto over lower-priced IPOs in 24 percent, and seniority over other investors in 30 percent. The limits of the analogy have to be stated alongside it. The sample is US venture unicorns of the 2010s, not Chinese frontier labs, and the overstatement ranges from 5 percent to 188 percent, which is far too wide to apply as a multiplier to any individual company. What carries over is only this: the literature has measured which direction a round price leans.

The second kind can be verified, and in exchange it never stops moving. Zhipu's market capitalization stood at about $71.3 billion on August 31, the date Rhodium took its figure from, and at about $44.2 billion on September 17, when the report came out. That is a fall of roughly 38 percent in three weeks. CNBC noted in its story that same day that Zhipu's share price had briefly more than tripled over the summer before returning to spring levels. The Zhipu bar drawn as of August matches the August 31 market capitalization within the margin of error. The 46x is not wrong. But on the day it was printed, the valuation was already a different number. This is not a Rhodium mistake but a property of marking anything to market.

4.2Six in the chart, two by the last outlet

The chart holds six multiples. The report's prose carries five. MiniMax's roughly 19x never makes it into a sentence, and neither do the reference months. Go one step further down and something else falls away. Counting what survives at each step exhaustively reveals a pattern.

A comparison of the chart on page 17, the prose on page 16, and the coverage that carried them onward. Only the values actually printed at each step are counted.

The order of the losses runs from the lowest multiple up. The 19x that lived only in the chart goes first, the 46x next, and the one that survives to the end is the most extreme, 163x. There is no point speculating about intent here. Rhodium drew all six in the chart and spoke about five of them in prose; the outlets carried as many as their space allowed. Stated structurally: each step down the citation chain drops the lower numbers first, and the qualifiers go before the numbers do. The loss of information does not begin at the outlets. It has already begun inside the original report.

Why doesn't the money come back to the lab? The answer attached to that question most often is open weights, and the report starts there too. Chinese frontier labs have backed open-weight releases, and that has limited their potential revenue. The sentence that follows reads as below.

No condition is attached to that sentence. In the license files themselves, there are conditions. This does not make the report wrong; it means the report left the conditions unstated. Reading the license files of five models directly and separating them by scope gives the following.

Compiled by reading the LICENSE files in the Hugging Face repositories directly (checked September 19, 2026). What was checked is the current file in each repository, so these clauses should not be generalized to earlier or later releases. Access to the Qwen3.8-Max repository was restricted, so the license was confirmed from another repository carrying a license file of the same title. The Moonshot license exempts internal use and access through official products or certified partners; the list of certified partners and the terms attached to them could not be verified.

5.1The threshold starts at $20 million

Here is how the clauses are worded. The Moonshot license states that if the licensee or any of its affiliates operates a Model as a Service business and the aggregate revenue of the licensee and its affiliates exceeds 20 million US dollars over any consecutive twelve months, the licensee must enter into a separate agreement with Moonshot AI before using the software or its derivative works for any commercial purpose. The Alibaba license has the same structure with a $50 million threshold, and it defines MaaS inside the same document: giving a third party access to language model inference or fine-tuning in a manner that allows that third party to exercise meaningful control over the inputs, parameters or training data.

This is where the report's next paragraph is worth reading on top of the licenses. Rhodium writes about revenue sharing as something currently being negotiated: Moonshot is in discussion with Microsoft, Amazon and Google about these arrangements, and its share of revenues could reportedly run as high as 30 percent. But the two companies named in that sentence, Moonshot and Alibaba, are precisely the ones that have already written the hook into their licenses. A negotiation here does not create a right that did not exist. It sets the price of a clause already on the page. The size of the thresholds also makes clear that these clauses are not aimed only at the very largest operators. Twenty to fifty million dollars is a level most companies in the inference business will cross. The third-party infrastructure business of serving open models that Pebblous examined earlier sits exactly there.

Even a concluded negotiation does not guarantee collection. The report says so itself.

For it to be new revenue, it has to come from abroad. And that path brings its own risks. Pebblous has worked through what follows a move to self-hosting once already, in a cost calculation for running a Moonshot model in house, and mapped how unevenly open-weight adoption spreads across countries in a report on the geography of that adoption.

5.2More than half the revenue arrives from outside

The overseas shares the report cites for three companies are high. The three figures rest on different denominators, though, so they cannot be packed into a single sentence. The table marks which of the three is a share of revenue and which a share of ARR.

Figures the Rhodium report cites with footnotes to Chinese media. The three rest on different metrics over different periods and cannot be combined into a single average.

The report's reading runs like this. Generating new revenue rather than reshuffling domestic revenue means leaning on overseas demand, and that raises geopolitical and compliance risk in turn. When one route to collection opens, risk attaches somewhere else.

5.3The cost of solving one task varies by multiples

The third strand is price. Borrowing estimates from the AI comparison firm Artificial Analysis, the report records where Chinese frontier models sit on a per-task basis even after the August increases. This is not a token rate card. It is a weighted average of what it costs to solve one of the tasks used in an intelligence index evaluation, computed from input, cache-hit, cache-write, reasoning and output token prices, divided by the number of tasks and then weighted by the index.

Bands from Figure 8 and the prose of the Rhodium report. Per-model figures could not be confirmed because the underlying source is an interactive chart, so only the bands are carried here.

The report treats this price difference as the main reason Chinese AI companies earn such thin profits. As structural background for having to price low it cites deflationary pressure, a customer base with low willingness to pay, fierce price competition, and the ubiquity of open-weight models. At the API layer alone, though, the picture differs. DeepSeek reportedly posted an 83 percent API gross margin as of July, close to Anthropic's low-to-mid 80s. At the company level the two separate again, DeepSeek at 45 percent and Anthropic in the mid-60s. OpenAI's company-wide gross margin in the first quarter was lower still at 39 percent, and in August the structure shifted as enterprise revenue passed half of the total. Most hyperscalers do not break out cloud margins separately; only Alibaba Cloud is visible, at an adjusted EBITA margin of 9 percent in 2025 and 10 percent in the first half of 2026. Over the same period Amazon went from 35 to 39 percent and Alphabet from 24 to 34 percent. Alibaba chief executive Eddie Wu believes Alibaba Cloud's margin can reach 20 percent, and even at that level it would sit below the operating margins of the major American providers. All of these figures are ones Rhodium cites from other outlets and analysis firms with footnotes, rather than compiling itself.

Everything so far has been the seller's situation. Move over to the buyer's side. That a supplier is loss-making and leaning on the equity market is not a problem that stays with that supplier. For an organization that has wired a pipeline to the supplier's model, it comes back as unit-price risk. And that risk is not hypothetical. It has already materialized once.

DeepSeek spent the first half of this year saying it would hold API prices low, running against the industry's upward drift. Then it signaled an increase on August 6, posted the official notice on August 13, and put it into effect at midnight Beijing time on August 17. Peak-hour pricing arrived at the same time. Peak runs from 9am to noon and from 2pm to 6pm Beijing time, and everything outside those hours is exactly half the peak rate.

Yuan rates as announced by DeepSeek on WeChat. They line up with the dollar conversions Western outlets reported. The token rates in this table and the per-task costs in the previous section are measured with different rulers and cannot be swapped inside one sentence.

The line that stands out is cache-hit input, up more than tenfold. The pattern most exposed to that line is the one that throws the same prompt repeatedly, meaning agent-style usage that attaches a long system prompt to every call. A call pattern designed around a cheap price becoming the most expensive pattern the moment the price moves is a scene Pebblous has watched before, in the marks an agent's retry loop left on the invoice.

The report adds a sentence here. V4.1 Flash prices are expected to come down in September, but output token prices will still sit above where they were before August. A cut is not a restoration. Nor is this only DeepSeek. The report notes that most Chinese frontier labs and hyperscalers raised model prices from early this year, with DeepSeek initially running against that current. Zhipu also raised prices several times during the year, and Tencent's Hunyuan raised some tiers by more than fourfold.

6.1In consumer apps, users do not convert into revenue

Why prices had to rise becomes clearer from the consumer app numbers. ByteDance's Doubao has more than 200 million daily users, and daily revenue below one million yuan, most of it e-commerce commissions. A single user generates less than 0.005 yuan of revenue per day, which is under a tenth of a US cent. On a transaction-value basis it is under 0.05 yuan per user, so mixing the two inflates the figure tenfold. The comparison is Douyin, at the same company: 750 to 800 million daily users and an estimated daily advertising revenue above one billion yuan. Alibaba also broke out an AI application segment in the second quarter, with 3 billion yuan of quarterly revenue against a 14 billion yuan net loss on an adjusted EBITA basis.

So the report's conclusion names the price explicitly.

The second of those two branches is running right now, and the report puts numbers on how far it opened this year. Private equity and venture capital money into Zhipu, MiniMax, DeepSeek and Moonshot came to 9 billion yuan across all of 2025; from January to August 2026, counting listing proceeds and private placements, it jumped to 179 billion yuan. About twentyfold. Zhipu alone raised roughly 4 billion yuan in its January Hong Kong listing and 27 billion yuan in a July H-share placement. Alibaba announced a HK$80 billion share placement in Hong Kong in August, its first new issuance since listing there in 2019, and said the entire amount would go to AI. In mid-September Zhipu launched a roughly $5 billion raise combining new equity and perpetual bonds, its second large raise in two months, while lifting its year-end ARR forecast from $2.4 billion to $3 billion and telling investors its latest ARR was $1.8 billion. The first branch, price, moved once in August. The report does not assume this funding climate stays friendly either. It lines up the sharp fall in Unitree's share price a week after listing, which has domestic investors worried, reports that the securities regulator has warned banks about listing volumes from low-quality companies, and the Moonshot and DeepSeek listings sitting on the horizon at still-uncertain valuations. Everything about the listings is secondhand rather than confirmed, and Moonshot said it does not comment on market rumors. Nor does the report put weight on the expectation that the state will hold up the market. The so-called national team funds have historically bought blue-chip ETFs to support the index rather than propping up a handful of already-surging technology stocks, and that method works poorly on technology stocks that pay no dividend. So the picture the report projects is this. Support for the top firms may continue while the rest of the AI sector gets cut off from equity funding, and future capex is constrained accordingly.

6.2Five things to check in the contract

When a customer says that switching to Chinese open weights costs a tenth as much, the thing to verify is not the rate card. The five rows below are drawn only from what this report states and from the license files we read ourselves. This piece did not check which clauses count as an industry standard, so none are listed here.

Not one of the five rows asks about model performance. They are all questions about contracts, records and portability, and the material needed to answer them mostly sits inside the organization already. What you actually buy, when you wire a pipeline to cheap weights, is not the unit price printed today. The share you take on is how that price moves from here. Money a supplier has not yet collected has not disappeared. It is sitting somewhere, and who bills for it, when, and under what name is the question this report hands on.

This entire report stands on one number: the most recent month of revenue multiplied by twelve. That number set the price of companies, set the gap between two countries, and set newspaper headlines. What Pebblous does sits exactly at that point.

7.1An undefined yardstick is the more dangerous one

At DataClinic we attach a grade and a defect type to each individual record. The first step in that work is not assigning the grade. It is defining what the grade measures. Figures accumulated without a defined yardstick look more precise the larger they get, which is what makes them dangerous. The $10.7 billion is that danger realized at national scale: a sum of seven figures carrying six different reference months, and not one step in the chain of citation attached a sentence asking which month.

7.2Metadata has to travel with the value

The sharpest evidence in this case is Rhodium's own chart. When Rhodium drew the multiples, it wrote each company's reference month straight into the axis labels. On the revenue-mix chart, it noted that Tencent's cloud revenue is bundled with other businesses and even wrote down which direction that biases its own figure. The metadata was attached at the source. And then the prose of the same report stripped the labels and listed the multiples alone, and the outlets dropped one low number after another from that prose. The lesson is not that definitions and reference dates should be recorded at the source. Rhodium recorded them. Recording is not enough; the qualifier has to be bound to the value so that it moves when the value moves. That is exactly what AI-Ready Data demands, and what happens to training data here happened to a financial metric instead.

7.3Customers ask about performance, and the answer is in the contract

The five rows in Section 6 are a starting point for a conversation rather than a product pitch. When a customer decides to switch models, the place we can help is not choosing the model. It is organizing the records that come with the decision. Which version of the weights is held where, what threshold the license file for that version states, and which line on our invoice moves first when the supplier changes its price. None of the five rows is a model performance question. They are all questions of data and records, and they are the work Pebblous already does with customers.

7.4What this article does not say

Pebblous does not take the position that Chinese AI is in trouble or that America has won. The same report also records growth rates, the shift toward enterprise customers, and the revenue-sharing negotiations, and we have carried those with equal weight. Where we stand is on opening up the back of a single number. What ruler measured it, when, and what was left out or added in. Doing that work on somebody else's financial metric tends to make readers ask the same questions of the AI metrics in their own organization. That is the one shift this article is after.

The figures in this article come from two streams. The ARR and valuation ledger in sections 1 and 2 was carried over after checking it directly against the Rhodium Group PDF, and the valuation-multiple reconstruction in section 4 cites one academic paper with an explicit limit on how it applies. Below that are the press coverage, filings, and license texts that the Rhodium report itself cites, or that this article opened and checked directly.

Primary report and academic anchor

• 1.Logan Wright, Endeavour Tian. "Examining China's AI Financing." Rhodium Group, 2026-09-17 (22-page PDF). This is the primary report this article traces. The ARR table, the Figure 12 valuation values, the Figure 9 financing breakdown, and the licensing discussion were all checked directly against this document.
• 2.William Gornall, Ilya A. Strebulaev. "Squaring venture capital valuations with reality." Journal of Financial Economics 135(1), 2020, 120–143 (NBER WP 23895, 2017). nber.org/papers/w23895 — Reconstructs the contractual terms in the charters of 135 U.S. unicorns and re-prices each. Reported post-money valuations run 48% above fair value on average (published-version figure). Section 4 cites it only for the limited claim that the literature has quantified which direction a round-price denominator is biased — it does not use this ratio to recompute any individual company's multiple.

Press coverage

• 3.Evelyn Cheng. "OpenAI and Anthropic are making 10 times more revenue than all Chinese AI models combined, research group Rhodium says." CNBC, 2026-09-17. Quotes Rhodium partner Logan Wright directly, and this article is also the source for Zhipu's (Z.ai's) revised year-end ARR forecast ($1.8 billion).
• 4."China's top AI models generate just 10% of OpenAI, Anthropic revenue: report." South China Morning Post, 2026-09-18. Carried only 4 of the 5 multiples (Zhipu's 46x is dropped) — evidence for this article's claim about information loss down the citation chain.
• 5.Park Chan. "중국 AI 총매출, 오픈AI·앤트로픽 10% 불과...몸값 '거품' 논란" ("China's total AI revenue is only 10% of OpenAI and Anthropic's — valuation 'bubble' debate"). AI Times (Korea), 2026-09-18.
• 6."Chinese AI models soar in value but make just 10% of OpenAI and Anthropic's revenue." Invezz, 2026-09-17. One of the few outlets to link directly to the Rhodium PDF. Carried only 2 of the 5 multiples (DeepSeek and Moonshot) — the sparsest case of the same information loss.
• 7."China's Moonshot in talks with Microsoft, Amazon, Google over K3 revenue sharing." Reuters, 2026-08-26. Cited directly by the Rhodium report — reports that Moonshot's share of the negotiated revenue split could run as high as 30%.
• 8."ByteDance secures $29.6 billion offshore syndicated loan for AI push." Reuters, 2026-09-04. Cited directly by the Rhodium report — the basis for the financing ledger in section 1.
• 9.36Kr report on Doubao's daily active users and revenue, 2026. The primary source a Rhodium footnote links to. It separates Doubao's per-user daily GMV (0.05 yuan) from per-user daily revenue (under 0.005 yuan) — the basis for the Doubao/Douyin diagram in section 6.1.

Filings and license texts

• 10.Zhipu AI (Z.ai, 2513.HK). 2026 interim results, Hong Kong Stock Exchange filing, 2026-08-31. Cross-checked against SCMP's figures via Tencent News and Xueqiu coverage that quoted the original numbers (2026-08-31 through 09-02). The basis for the 6.5x gap between ARR and disclosed H1 revenue in section 3, and the 86.5% API/MaaS revenue share in section 5.
• 11.MiniMax (0100.HK). H1 2026 Financial Results, PRNewswire (official company release), 2026-08-31. The basis for the 3.4x ARR-to-disclosed-revenue calculation in section 3 and the shift in enterprise revenue share (30% to 80% in one year) cited in the closing section.
• 12.DeepSeek-V4-Pro LICENSE, Hugging Face. Plain MIT — no resale revenue threshold and no attribution requirement. The primary text behind section 5's claim that DeepSeek is one of two labs releasing without conditions.
• 13.GLM-5.2 LICENSE, Hugging Face (zai-org). Plain MIT. Zhipu has the least conditional license among the frontier labs even though it earns the largest share of its revenue from its own API/MaaS business (86.5%) — the counterexample section 5 relies on.
• 14.Kimi K3 License, Hugging Face (moonshotai). The primary text of the clause requiring Moonshot's separate written agreement once a Model-as-a-Service business's trailing 12-month revenue exceeds $20 million. What Rhodium describes as an "in discussion" revenue split is, in fact, setting the value of a clause that already exists.
• 15.Qwen3.8-Max License, Hugging Face. The primary text of the clause requiring a separate license from Qwen once trailing 12-month revenue exceeds $50 million. The Qwen/Qwen3.8-Max repository itself returned HTTP 401, so this article confirmed the same license (same title, "Qwen3.8-Max License") in this repository instead.
• 16.MiniMax Community License, Hugging Face (MiniMaxAI). The primary text of the clause requiring prior written approval once annual revenue exceeds $20 million. It also specifies an attribution requirement — "Built with MiniMax M3" — across all commercial use.