Inside China's Machine · Where China's Tokens Go, Episode 1

On 3 June 2026, COL Group (中文在线), a listed Chinese digital-content company, posted investor-relations record No. 2026-004. It is the write-up of a roadshow and phone call held in Beijing on 1 and 2 June, attended by about 40 named brokers and funds. Partway through, the record makes the case that tokens are flowing to entertainment. In English it reads: “From the distribution of industry call data, the shares of token consumption by scenario bear this out. Of all domestic token consumption, AI short drama and video generation ranks first at 55%; then e-commerce and marketing (livestream selling, ad marketing) at 24%; software development, which was once pinned with hopes by the Vibe Coding vogue, is only 15% even with Claude-type coding advances.”

Add the three shares and you get 94. That is our arithmetic, not the record’s. The remaining 6% is not assigned to anything. A leftover is not unusual in a pie chart. But a reader handed a national breakdown would expect either the full 100 or a stated “other”, and this one has neither. The gap is a small sign of something larger: the record gives the shares and nothing they could be checked against.

Our judgment is that the 55% cannot bear weight as a share of China’s tokens. It is a claim without a published measurement behind it, and on the evidence available it most likely describes some customers rather than the country. The record, in that one line, is the earliest document in the trail that states the split. It cites “industry call data” (行业调用数据) and gives no period, no sample, no unit and no data provider. Every later source in the trail either repeats it or narrows its scope. That does not make video small. IDC, working independently, puts video-generation models at 32.7% of China’s model-as-a-service revenue in the first half of 2026. Video is a large and fast-growing source of paying demand. The useful question is not whether the figure is 55%. It is “55% of what”.

What the record says, and what it leaves out

The record is eight pages long. Its attachment list reads “附件清单：无”, meaning none, and no source line appears anywhere in it. The category that ranks first is a merged one: “AI short drama and video generation”, not short drama alone. Short drama is the vertical-screen, minute-long serial format that Chinese apps sell by the episode. Video generation could cover advertising clips, e-commerce product video and anything else a video model renders. The record does not say how the two halves split.

The 15% belongs to software development. Some later press coverage turned it into research and development. One outlet, 财闻, wrote on 19 October that “only 15% really goes to R&D”. The record’s word is software development, which is a narrower thing.

The record also gives two reasons video might use so many tokens. First, it says AI drama with realistic human characters at 1080p usually uses two to three times the tokens of comic-style drama at 720p. Second, production teams repeatedly “draw cards” (抽卡): they regenerate a shot again and again and keep the one that works. The record adds that AI short drama and video generation token use “has reached over twice that of software development”. That is consistent with 55 against 15, but it is a restatement of the same claim, not a second source.

One number in the record does have a clear owner. The company says its own weekly-average call volume is up more than 20% on the previous period. That is a statement about COL’s own meter. The 55% is a statement about the industry’s meter, and the record does not say who read it.

The trail of the number

What happened next is the part that would surprise someone who has seen the 55% quoted as settled. The number moved from outlet to outlet. Its scope changed along the way, and the trail it leaves ends at the investor call.

On 4 June, Securities Times reported the record. It called the figure “the industry token call data the company cited” and gave only the 55%. In July, a note circulated that was attributed to Zhongtai Securities’ media and internet team. According to Shanghai Securities News, that note ranked first-half 2026 “B-side inference token consumption”, with short drama and video at 55% and marketing at 24%. B-side means business customers, and inference means running models rather than training them. The paper described the note’s scope as “clearly smaller than national”.

On 30 September, CAICT, the China Academy of Information and Communications Technology, released its Token Economy Development Research Report (2026). Tide News and Shanghai Securities News, both quoting it, say it gives 55% as the share of “whole-industry token consumption”. They also say it footnotes a Zhongtai Securities research report dated 20 May 2026, part five of a series on token economics in AI applications. That footnoted report is the earliest dated document in the trail, so the Tide News reporter went to it. By that reporter’s account, the Zhongtai report says token economics first reached scale in AI drama, marketing and e-commerce, and that for AI drama, token use is “nearly half of production cost”. According to the reporter, it contains no 55% split. So this rests on one named outlet’s reading.

On 23 October, Shanghai Securities News reported that it had found no authoritative data from before June corroborating the 55%. A brokerage official the paper quoted said the figure’s first explicit appearance “should be” the COL call. A compute practitioner, 程伟, told the paper the share “may be a conclusion drawn from one public-cloud platform’s data”, and that large internet firms’ in-house training compute, private deployments and research-institute compute “should not be counted”. The paper presented this as his speculation, not as data.

Read as a chain, the trail offers different scopes rather than repeated confirmations of one: an industry-wide share in the CAICT report, business-customer inference in the July note, perhaps a single cloud platform in compute practitioner 程伟’s speculation, and at the investor call only “industry call data”. Nobody along the trail has published what the number was measured on.

What a share of tokens can count

Start with the unit, because everything rests on it. A token is the unit a model provider meters. For a text model, it is roughly a word fragment. A video model’s provider meters tokens too, but a clip uses far more of them: ChinaTalk cites about 100,000 tokens for a five-second 720p Seedance clip, which it likens to roughly 200 chatbot answers. The record itself says realistic-human 1080p drama uses two to three times the tokens of comic-style 720p drama. So a share of tokens partly reflects how heavily each piece of work registers on the meter, not only how much work the customer is getting done.

Next, the denominator. “All domestic token consumption” sounds definite until you look for the totals it would be a share of. Three series exist. IDC’s states its scope, the public-cloud model-as-a-service market. The National Data Administration’s and CAICT’s do not state theirs.

The National Data Administration’s head, Liu Liehong, told a State Council Information Office briefing on 24 March 2026 that daily token consumption had passed 140 trillion. He said that was more than 1,000 times the 100 billion of early 2024, and more than 40% up in three months from 100 trillion at the end of 2025. In a separate monitoring report, CAICT put domestic daily consumption at about 500 trillion in June 2026, almost four times December 2025. It attributed part of the growth to more model-as-a-service platforms selling tokens through public-cloud APIs. IDC counts 2,857.9 trillion tokens across China’s model-as-a-service market in the whole first half of 2026.

Our arithmetic on those stated figures does not line up neatly. Going from 140 trillion a day in March to 500 trillion a day in June would be roughly a 3.6-fold jump in about three months. IDC’s half-year total works out to an average of about 15.8 trillion a day over 181 days. That sits far below even the March national figure of 140 trillion. IDC’s counts only public-cloud model-as-a-service tokens and the other two do not say what they count, so the gap shows the totals are not interchangeable; it does not show which is right. ByteDance’s Volcengine, meanwhile, reported its Doubao models alone at more than 120 trillion tokens a day in March and more than 180 trillion on 23 June. If Doubao and the national figure measured the same thing, one vendor would have been most of the country in March. More likely they measure different things. Nothing in these sources reconciles the three. And a share of “all domestic tokens” cannot be checked against two totals whose scope is unstated and a third narrower than all domestic tokens.

The other reading, heard fairly

There is a serious case that the 55% is right in direction even without a published measurement. It goes like this. Video tokens are heavy: higher resolution and realistic characters raise the count, and every usable shot may sit on a pile of discarded takes. Text workloads are many but light. So it would not take a large number of video customers to outweigh a large number of chat and coding users on a token meter.

Some ground-level signals fit this. A token-service platform interviewed by Tide News named three main customer types: foreign-trade service firms writing marketing copy and building overseas sites, AI short-drama firms, and technology firms, the last mostly “in a gradual introduction phase”. On that platform, a 55/24 split between video and marketing would not look strange. Production budgets point the same way. Jiemian reported that a 12-episode short drama “burns nearly RMB 200,000 of compute cost”, though it did not give the basis for that figure.

On this reading, the real question is not the token share but what a token share is worth in revenue. That leads to the independent data, which partly supports the rival reading and partly limits it.

IDC’s own summary of China’s model-as-a-service market for the first half of 2026 puts revenue at RMB 10.63 billion, up 889.4%, on 2,857.9 trillion tokens, up 432.5%. By model type, large language models brought in 47.0% of revenue, video-generation models 32.7%, and other models 20.2%. IDC calls video generation an important “second growth curve”. It also says internet companies’ own workloads (recommendation, search, dialogue and coding) remain the “base” of the market’s revenue, with office automation and customer service adding growth.

So video is roughly a third of what the market was paid, measured by an outside firm over a stated period. That is not a niche. It is also not a majority, and it is a share of money, not of tokens. The two can diverge legitimately: if a video token is priced differently from a text token, the token share and the revenue share will differ even when both are measured perfectly. Of the two, IDC’s is the only one with a published scope.

Two other figures circulate as rebuttals, and neither settles it. On 19 October, ByteDance’s Doubao PR lead, 刘星, wrote on Weibo that about 10% of ByteDance’s total compute goes to training and inference for video-generation models. But that is a share of one company’s compute, including training. The 55% is a share of industry tokens. Neither number can be converted into the other, in either direction.

Separately, Zilan Qian at ChinaTalk built an outside estimate, “assuming, generously”. She assumed that all 430,000 micro-drama titles released between January and August 2026 were fully AI-generated, at 200 minutes of finished footage each, computed at 1080p and 24 frames per second. On her numbers that gives about 251 trillion tokens, or about 2,500 trillion once multiplied by ten for discarded takes. Her piece separately cites about 100,000 tokens per five-second 720p Seedance clip. Set against her figure of about 34,000 trillion tokens of national consumption over eight months at 140 trillion a day, she gets roughly 7.4%. It is a bounded estimate resting on her stated assumptions, not a measurement. The title count it starts from comes from China’s National Radio and Television Administration, and the briefing as reported gives no definition of “AI drama” or counting unit. Her denominator is also the national figure whose scope is itself unstated.

Notice that the estimate covers short drama only. COL’s category also includes all other video generation. So the 7.4% and the 55% are not even counting the same slice of activity.

What would decide it

The two readings make different predictions, and one kind of document would test them.

If the 55% describes a slice of customers, such as business-side inference on one public cloud, then a broader split should show video well below a majority of tokens. Internet companies’ own recommendation, search and chat workloads running on cloud model services would carry the weight that IDC already sees in revenue. If the heavy-token reading is right, a vendor serving a broad range of uses should still show video at or near a majority of tokens, and a revenue split for the same customers and period would show what those tokens were paid. If video were a majority of tokens but around a third of revenue on the same base, the gap between the two shares would then be the interesting number.

The figure to watch is a cloud or model vendor’s own split of tokens by use, with the period and the denominator stated. Until a method is published, anyone sizing China’s AI demand should read the 55% as a claim about an unspecified customer base, not an established national share. Video should be read as neither a settled majority nor a niche. The firmer number available puts it at roughly a third of what the model-as-a-service market was paid in the first half of 2026.

Video tokens are heavy partly because of the redraws, and what a redraw costs is the next question. For now, the 94 hinted from the start at a figure with nothing published to check it against, and the point to take away follows from that. A share of tokens measures how much of someone’s meter a use consumes; until you know whose meter, it is not yet a fact about China.

Sources

1. COL Group (中文在线), Investor Relations Activity Record No. 2026-004 (posted 3 June 2026; roadshow and call of 1 and 2 June), Eastmoney-hosted copy of the filing, https://pdf.dfcfw.com/pdf/H2_AN202606031823217652_1.pdf, captured 2026-11-07. The 55 / 24 / 15 sentence, the attachment list (”none”), the 1080p and draw-card remarks, COL’s own weekly call volume.

2. Securities Times (证券时报), 4 June 2026, https://stock.10jqka.com.cn/20260604/c677226962.shtml, captured 2026-11-07. Its report of the record.

3. Shanghai Securities News (上海证券报), “谁在制造词元消耗对立”, 23 October 2026, via Eastmoney, https://finance.eastmoney.com/a/202610233904643058.html, captured 2026-11-07. The July note, the brokerage official’s “should be”, 程伟’s remarks.

4. Tide News (潮新闻), 26 October 2026, https://tidenews.com.cn/news.html?id=3596271, captured 2026-11-07. Its account of the CAICT report and the 20 May 2026 Zhongtai Securities report; the token-service platform’s customer types.

5. 财闻 (Caiwen), 30 September 2026 (CAICT report released) and 19 October 2026 (the 15% as “R&D”; 刘星’s Weibo post), https://www.caiwennews.com/article/1634675.shtml and https://www.caiwennews.com/article/1657301.shtml, captured 2026-11-07.

6. IT之家 (ITHome), 24 October 2026, https://ithome.com/1/017/518.htm, captured 2026-11-07. CAICT’s June 2026 monitoring figure of about 500 trillion tokens a day.

7. Xinhua, 24 March 2026, https://www.news.cn/tech/20260324/050774b26bc54f429fce2d21e284949a/c.html, captured 2026-11-07. National Data Administration: more than 140 trillion tokens a day.

8. Xinhua, 3 April 2026 (via Science and Technology Daily), https://www.news.cn/tech/20260403/c4f28c9bcccd4ed98456c07122f03a0b/c.html, and 10jqka, 23 June 2026, https://news.10jqka.com.cn/20260623/c677638966.shtml, captured 2026-11-07. Volcengine’s Doubao daily token figures.

9. IDC, China MaaS market, first half of 2026 (IDC blog, 29 October 2026), https://www.idc.com/resource-center/blog/中国maas市场实现近9倍增长-高增速下厂商加速涌/, captured 2026-11-07. Revenue, tokens and the split by model type.

10. Jiemian (界面新闻), October 2026, https://m.jiemian.com/article/15193592.html, captured 2026-11-07. The 12-episode short drama’s reported compute cost.

11. ChinaTalk, Zilan Qian, 30 October 2026, https://www.chinatalk.media/p/is-half-of-chinas-ai-making-slop, captured 2026-11-07. The 7.4% estimate and its assumptions.

12. Xinhua (Economic Information Daily), 18 September 2026, https://www.news.cn/tech/20260918/1fc21cfba8334e9bba130af8eb50b565/c.html, captured 2026-11-07. The National Radio and Television Administration’s count of 430,000 micro-dramas, January to August 2026.

Inside China’s Machine · Where China’s Tokens Go, Episode 1

Disclaimer: This article is for informational and educational purposes only. It is not investment, financial, legal or tax advice, and it is not a recommendation or solicitation to buy or sell any security. Do your own research, and consult a licensed professional before making investment decisions.

Inside China’s Machine · Tokenomics, Episode 10 of 10

The Answer Price = price per token × tokens per attempt ÷ success rate
The Chinese state buys some of the country’s AI and pays part of other buyers’ bills, so who buys most cannot yet be known from public data, and the number that would settle it has not been published.

On 3 June 2025, Shenzhen prepared a submission under a national scheme that helps companies pay for computing. One sentence in it puts two numbers side by side: 拟上报核算后算力服务总额11.68亿元，算力服务申请补助总额1.29亿元. In our translation: a proposed submission of verified compute services totalling RMB1.168 billion, and requested subsidies totalling RMB129 million. The first number is the value of the verified compute services included in the proposed submission. The second is the amount the state was asked to contribute towards that bill. The record does not say who the buyers are. It shows spending submitted for government support, not a government purchase.

That sentence shows the problem with the most common question about Chinese AI demand, which is whether the state or the market is buying. The state does buy. It also pays part of the bill when someone else buys. Paid commercial channels are growing at the same time, and their filings do not say who the customers are. Our view is this: who buys most of China’s AI, or leads its demand, cannot be known from public data yet. The reason is not a shortage of numbers. China publishes plenty. Its numbers sit at different stages, at different layers of the supply chain, and in different units, and none of them separates the buyer from the payer. The number that would settle the question is specific: deduplicated final-buyer invoices at one layer of the supply chain, for one year, in one currency and on one tax basis, with reimbursements shown separately. Until that exists, anyone who says the state leads demand, or the market does, is offering a view.

The series’ core thread, the Answer Price, is the cost of a successful task: price per token, times tokens per attempt, divided by the success rate. A token is a small unit of text that a model processes and providers commonly charge for. It is a buyer’s number, and that matters here. A voucher can reduce what the buyer pays without reducing what the provider receives. The buyer’s bill and the state’s support then describe parts of the same spending, not two separate purchases.

The state buys, in cases you can read

Start with the plain case, where a government body buys.

A Chinese public body that wants to buy something usually runs a tender, a published competition in which suppliers bid. The public record of the result, naming the winner and the price, is called a tender award. An award is a promise to buy. It is not delivery, not acceptance of the delivered work, and not payment. Those come later, and for the cases below the records omit them.

Here are some awards. A hospital in Changzhou awarded RMB4,488,000 on 1 April 2025 for a model and compute platform. The award bundles the two and gives no figure for the model alone. A Beijing research platform awarded RMB1,960,000 on 18 September 2025 for hardware: AI servers plus management and storage nodes. An award for a project called Tanmai model, phase I, came to RMB2,438,000 on 12 January 2026. It bundled data, a platform, models, compute and cybersecurity, and its text ties the project to national policy: 围绕发展新质生产力核心目标，践行国务院国资委”AI+”专项行动部署. In our translation: centring on the goal of developing new quality productive forces and implementing SASAC’s “AI+” special-action deployment. SASAC is the State-owned Assets Supervision and Administration Commission of the State Council. The sentence shows how the buyer describes the project. It does not establish what motivated the purchase.

Look inside one award and the category “AI spending” starts to blur. On 29 October 2026, a package recorded as NIM compute and infrastructure was awarded for RMB9,246,200. The summary lists four BaoDe PR425KI G2 servers at RMB1,360,000 each. By our arithmetic that is RMB5,440,000, or 58.83 percent of the package. The rest is other items, and nothing in the record lets you reclassify those as spending on models. The summary also gives no count of accelerator cards, which do much of the AI computation, and no cost for them. So even a well-documented state purchase is a bundle in which the AI part has no clean edge.

Then there is a contract shape that is easy to misread. In March 2026, Guangxi police awarded a compute rental in three lots: China Mobile RMB15,178,276.80, China Unicom RMB10,850,000 and China Telecom RMB10,850,000. By our arithmetic the total is RMB36,878,276.80. The term matters more than the total: 租赁期为36个月, the rental period is 36 months (our translation), and it starts on accepted delivery. Spread evenly, the package works out to roughly RMB1.02 million a month. That figure is our arithmetic and nothing more. It is not an invoice, not 2026 cash spending, and not a price per GPU-hour, a unit measuring the use of one graphics processor for one hour. The clock starts only when police accept delivery, and the award notice does not give the acceptance date. The three carriers are also not necessarily delivering the same compute. Two of the lots carry the same headline price, but a matching price does not show matching capacity.

The Guangxi case also shows why the buyer’s identity is harder to pin down than it looks. At the service layer, the police are buying rented compute from carriers. At the hardware layer, those carriers may themselves be buyers. A record at one layer answers only that layer’s question.

Four stages, and why the big numbers sit at the early ones

A framework contract is an agreement that fixes suppliers and prices in advance. The actual purchases come later, as separate orders drawn against it. In this sample, the largest numbers are framework numbers. They describe the terms for possible purchases, not money already changing hands.

China Development Bank’s record of 9 March 2026 shows quotes of RMB63,479,700, RMB66,008,598 and RMB60,349,800, including tax, for three product routes, each with five years of maintenance. The record says the bank will 与3条产品路线供应商分别签订框架合同进行采购: sign procurement framework contracts separately with suppliers for the three product routes (our translation). The quotes are real. The record shows no orders drawn against them.

A bigger example comes from the operators themselves. In a filing, Digital China disclosed a framework with China Mobile for AI servers. It was signed on 22 August 2024 and expired on 31 August 2025, and lot 1 named 842 AI servers and supporting products. The filing says the actual amount executed is set by the specific purchase contracts: 实际执行金额以具体采购合同为准. Digital China’s 2024 annual report gives the package bid quotation as RMB19,116,746,188.47 excluding tax, with an awarded share of 10.53 percent (中标份额10.53%). Multiplying the two gives about RMB2.01 billion. That is our arithmetic, an illustration of how the package might be allocated. It is not executed orders and not revenue.

This case also shows the layer problem. China Mobile signed a framework to buy AI servers. But a compute operator buying servers may rent out the capacity, so the final buyer may sit further down the chain. A purchaser at the hardware layer does not tell you who the final buyer is at the service layer. Count a server framework as demand, then count its customers’ rental bills as demand too, and the same machines can appear twice.

Sometimes a tender stops partway. One college ended a procurement 因采购需求以及采购标的有重大调整，故终止本次采购活动: because of material adjustments to the procurement requirements and the items to be procured (our translation). That shows a plan changing. It does not show demand falling away; the record shows no re-tender.

The award, the framework, the order and the payment are four different stages, though not every purchase passes through a framework. The ledger below keeps each case at its own stage. Notice that none of these records gives a payment figure.

So the first half of the proof holds. The state demonstrably buys AI and compute, with dates, amounts and named suppliers. The same records show how weak a claim that “the state leads demand” would be as a national statement. These are cases, mostly at the award stage. They cannot be summed. Some sit at the hardware layer and some at the service layer, and adding amounts across layers can count one supply chain more than once.

The state also pays when someone else buys

The second half of the proof is the more interesting one, because it breaks the question in two.

A compute voucher is a coupon from a government scheme that pays for part of a company’s bill for computing. The company still signs the contract with a compute provider and still receives the invoice. The voucher then covers a slice of that invoice. Ask who bought, and the answer is the company. Ask who paid, and the answer is the company and the state, in proportions that depend on the scheme.

Shenzhen’s line from June 2025 is the clearest illustration, because it gives both the value of the services submitted and the support requested against them.

By our arithmetic, the requested subsidy is 11.04 percent of the verified services. It is tempting to read the other 88.96 percent as private money. The record does not support that. It does not say who the buyers are, and some of them may be state-controlled. Other support may also apply. It also does not say whether the applications were approved, or whether any vouchers were redeemed, meaning money was actually paid out against invoices. An application is a request. Redemption is the cash arriving.

Elsewhere the published numbers are the rules of the schemes, and the rules are ceilings. A ceiling is the most a scheme will pay. It is not what it usually pays. Shenzhen’s 2026 training vouchers say 以训力券单次抵扣不超过服务合同金额的50%: a single use of a training voucher offsets no more than 50 percent of the service contract (our translation). The limit rises to 60 percent for eligible start-ups less than a year old. Each beneficiary is capped at RMB10 million a year, must have at least RMB500,000 of eligible annual spending, and must buy from compute providers it is not affiliated with. Hubei’s 2025 scheme provides a 10 percent subsidy on actual compute expense, capped at RMB1 million per computing user a year: 给予10%的补助，单个用算主体每年累计最高补助100万元. Liaoning’s scheme for 2026 to 2028 lets combined local and provincial support reach no more than 30 percent of actual compute-contract spending (累计给予算力合同实际发生金额不超过30%的补贴), capped at RMB2 million per beneficiary a year. Beijing’s 2023 to 2025 scheme, in a 2024 reprint, set a 20 percent ceiling on the contract. That was its ceiling then. It is not a 2026 figure.

Shanghai’s notice of 7 May 2026 closes the loop with the private market. It supports 调用非关联方云平台部署的第三方大模型API产生的费用: the cost of calling third-party large-model APIs deployed on non-affiliated cloud platforms (our translation). An API, or application programming interface, lets one piece of software request services from another; here, customers send requests to a model over the internet and pay for usage. Under that clause, a private company buying model services by the call, from a cloud platform it is not affiliated with, can have part of that bill covered by public money. A buyer that looks entirely commercial can be partly subsidised.

Support also comes as funds and as cheaper inputs. Beijing’s Shijingshan district announced a compute fund on 7 February 2025, 设立规模1.9亿元, established at a scale of RMB190 million (our translation). That is the size of the fund. It is not the money actually paid in and not the value of services bought. Fujian in 2024 said it would 协调中国联通福建省分公司提供优惠的机柜租赁价格、电费补贴和其他优惠政策: coordinate China Unicom’s Fujian branch to provide favourable rack rental prices, electricity subsidies and other preferential policies (our translation). The support is arranged through the provider’s prices and subsidies. The rate per kilowatt-hour is not public, and neither is the amount paid.

Shenzhen’s first training-voucher issue, in March 2025, was reported on a government site as 总额近2亿元，惠及近40家人工智能和具身智能机器人企业: nearly RMB200 million in total, benefiting nearly 40 AI and embodied-robotics firms (our translation). That is an allocation, not redeemed cash. It may also overlap with the national-scheme submission above, so the two cannot be added.

One scheme does report redemption, and it shows how much care the numbers need even when they get that far. Xinhua reported that in Guizhou, as of 31 March 2026, 299 of the 333 voucher certificates issued under the 2025 scheme had been redeemed, for RMB98.02 million, and that 算力券资金执行率达到98%: the voucher fund execution rate reached 98 percent (our translation). By count, 299 of 333 is 89.79 percent, by our arithmetic. That is the share of certificates redeemed, not the share of money spent: certificates need not represent equal amounts. The 98 percent is a measure of fund execution whose denominator is not published. The two percentages measure different things, and the published information does not reconcile them. Guizhou has allocated RMB140 million for 2026, and dividing the 2025 redemptions by it would set one year’s results against another year’s budget.

A national action document names fiscal funds and government procurement among its policy instruments: 充分发挥财政资金、政府采购等政策作用, make full use of the policy roles of fiscal funds, government procurement and other instruments (our translation). It sets targets of more than 70 percent adoption of AI terminals and agents by 2027 and more than 90 percent by 2030. Those are targets. They describe what the document wants. They do not describe what has happened.

Now return to the Answer Price. A voucher at Shenzhen’s 50 percent ceiling would halve the buyer’s effective price on that invoice. With the task, token use and success rate unchanged, the buyer’s cost per successful task would fall with it, while the provider could still receive the full invoice amount. If you measure demand from the buyer’s side, the voucher makes the company look like a price-sensitive commercial buyer getting a good deal. If you measure from the seller’s revenue, the whole invoice looks like customer demand. Either way the state’s share disappears unless reimbursements are shown separately. There is also a trap on the other side: count the voucher as state spending and the whole supported invoice as customer spending, and the government’s slice is counted twice.

So “who buys” and “who pays” are two different questions. The documented cases answer the first in places. The voucher schemes show why the first answer does not settle the second.

Read more

Inside China’s Machine · Tokenomics, Episode 9 of 10

The Answer Price = price per token × tokens per attempt ÷ success rate
China now has a working second supply of AI compute that hedges an import channel which has opened and closed, but no public record yet shows whether its tokens cost less or more than imported NVIDIA ones.

On 24 April 2026, DeepSeek published the prices for its new V4 models as a single image. Pro’s list prices were CNY 1 per million input tokens served from cache, CNY 12 per million input tokens not in cache, and CNY 24 per million output tokens. Tokens are the small units of text a model reads and writes. Under the numbers, the lab explained why Pro was scarce: 受限于高端算力，目前 Pro 的服务吞吐十分有限 (limited by high-end compute, Pro’s service throughput is currently very limited). Then it made a forecast that named a chip: 预计下半年昇腾 950 超节点批量上市后，Pro 的价格会大幅下调。 (Pro prices are expected to fall substantially after Ascend 950 supernodes are available in volume in the second half.)

Ascend 950 is Huawei’s next generation of AI chip. A price list that ties a future price cut to a domestic chip’s arrival is a lab’s forecast, not a measurement. It is still a useful sentence, because it states plainly what China’s own supply of compute is being asked to do: turn a scarce, expensive model into a cheap one.

Our view is this. China has built a working alternative supply of AI compute. It has hardware systems documented in dated publications with published serving numbers, a lab and a cloud operator that say their production traffic runs on domestic hardware, and, alongside that domestic supply, an import channel for NVIDIA chips that policy has opened and closed more than once. That last fact is what makes the alternative matter. What no one can show yet from public records is whether a finished token from the domestic supply costs less or more than one from the NVIDIA hardware China can import. The arithmetic that would decide the question has not been published for both sides on the same terms. The hedge is visible. A cost advantage is something nobody can yet show, and nobody can yet rule out.

The alternative is real

An accelerator is a chip built for the arithmetic that AI models run on; NVIDIA calls its accelerators GPUs, and Huawei calls its accelerators NPUs, for neural processing units. HBM, high-bandwidth memory, is memory stacked in layers right next to the chip so the model’s weights, the numerical values that encode what it has learned, can be read fast enough to keep the arithmetic busy, and some of the trade policy in this story is about HBM.

The clearest domestic system on record is CloudMatrix384, described in a Huawei-led technical paper (with SiliconFlow co-authors; version 3, 19 June 2025). It joins 384 Ascend NPU packages with 192 CPUs across 48 compute nodes in 12 racks, plus four racks for communication (16 in all). Each NPU package holds two dies, two separate pieces of silicon, so the system has 384 packages, not 768 cards. The power it consumes and the links that move data between chips belong in the system’s economics too. Like the other papers below, this is a company describing its own system, not an outside measurement.

A system that exists is not the same as one that serves paying users. On 17 September 2026, Zhipu AI (Z.ai), which makes the GLM models, wrote that GLM-5.3-Flash runs on a system of more than 100,000 Chinese-made accelerators: “All production inference for GLM-5.3-Flash runs on this system.” Inference is the work of running a trained model to answer requests. That is a company statement about its own fleet; no independent audit of the fleet is public, and the company does not name the chip.

Huawei’s 2025 roadmap puts the next generation, Ascend 950DT, in the fourth quarter of 2026. That is a roadmap date, not a delivery report; Huawei said on 17 September 2026 that Ascend 950 supernodes were on sale early. A supernode joins many accelerators into a closely connected system.

Then there is the channel that makes all this a hedge rather than a hobby. On 2 December 2024, the US Bureau of Industry and Security put export controls on high-bandwidth memory above a bandwidth-density threshold. In April 2025 the US government told NVIDIA a licence was required to export its China-market H20, and NVIDIA booked a $4.5 billion charge. On 13 January 2026, the US Commerce Department said H200-class licence applications would be reviewed on a “case-by-case basis provided certain security requirements are met.” On 26 August 2026, NVIDIA said it had shipped H200 products worth less than 1% of its total data-centre revenue to customers in China under licences that quarter, and that its forward outlook held no China data-centre compute revenue.

The pressure has come from the other side too: on 5 November 2025, Reuters, citing anonymous sources, reported guidance that state-funded data centres should not use foreign AI chips, and that projects less than 30% complete would have to remove installed foreign chips or cancel purchases. Reuters said it was unclear whether the guidance applied nationwide, and no Chinese official text of it has been found, so it remains a press report of guidance, not a published rule.

Put those records in a row and the pattern is easy to see. A Chinese buyer planning compute for 2026 could not count on any one import route staying open for a whole year. Against that, a domestic system that serves production traffic is worth something even before anyone knows its price per token, in the same way insurance is worth something before anyone knows whether the house will burn. That is the hedge, and it is the part of the case that the public record does show.

Read more

Inside China’s Machine · Tokenomics, Episode 8 of 10

The Answer Price = price per token × tokens per attempt ÷ success rate
Chinese labs’ open weights give buyers a credible alternative supplier, but the cost of a finished task is set by cache hits, token use, service terms and utilization, and the public record cannot yet show whether they capped anyone’s price.

On 20 January 2025, DeepSeek announced its reasoning model R1 with this sentence: “今天，我们正式发布 DeepSeek-R1，并同步开源模型权重。” That means today we officially release DeepSeek-R1 and release its model weights at the same time. A few lines later came the part a buyer’s lawyer would notice: “完全开源，不限制商用，无需申请。” That means fully open, no commercial-use restriction, no application.

The thing given away needs a precise name. A model’s weights are the trained numbers that make it work. Publish them, along with the code that runs them, and anyone with suitable machines can run the model, adapt it or sell access to it. DeepSeek released R1’s weights and a technical report, with code and models under MIT, a short licence that permits almost any use. It did not publish a full training corpus in any source we reviewed. You could run R1 and modify it, but not, from what was published, rebuild it from scratch. In this series, only the second counts as open source. The first thing, which is what DeepSeek released, is better called open weights, and the difference matters for what follows.

Here is our view. Published weights from Chinese labs give a buyer a real alternative and a fallback: another supplier, or the option of running the model themselves. They do not tell the buyer what a finished piece of work will cost. That depends on how much of each request the provider can serve by reusing earlier work, how many tokens the model spends, which tier and rate limits apply, and, for anyone running their own machines, how busy those machines stay. Tokens are the small units of text a model reads and writes. So the comparison that matters is the bill for the same task, not the price per token. Two larger questions remain open. One is whether open weights pushed down what proprietary labs charge. The other is whether the lab or the company serving its weights holds the bargaining power. Neither can be settled from public records yet, and the reasons are specific enough to name.

The alternative is real

Within weeks, R1 was on sale from other companies. A host is a company that runs someone else’s model on its own machines and sells access to it. On 18 March 2025, Fireworks described its R1 price in a line that leaves nothing to interpret: “Matches standard DeepSeek pricing ($0.55/$2.19 per million tokens).” Those are list prices for input and output tokens. Tencent’s cloud, from 26 February 2025, listed R1 at CNY 4 per million input tokens and CNY 16 per million output tokens, the same as DeepSeek’s own price. Baidu’s Qianfan platform, on 5 February, listed it at CNY 2 and CNY 8, half the lab’s price on paper. The same announcement offered a free period of about two weeks, so for a while that nominal price was not what anyone paid.

A buyer who had heard of R1 at launch could choose among the lab and at least three hosts within about two months. Those hosts charged the lab’s price or less, and the buyer did not have to download or run the model. The alternative arrived as a set of suppliers.

The newer case is sharper and less tidy. Moonshot’s Kimi K3 lists at USD 3 per million input tokens and USD 15 per million output tokens, against USD 5 and USD 25 for Anthropic’s Claude Opus 4.8 (list prices dated 17 July and 28 May 2026). That makes K3 40 percent cheaper per token. Artificial Analysis, an independent evaluator, ran both on its Intelligence Index on 17 July 2026 and measured USD 0.94 per task for K3 against USD 1.80 for Opus 4.8, about 48 percent lower. That is the cost of running the evaluator’s tasks, not a token-price comparison. K3 scored 57 on that index, and the two models are not of identical quality, so this is a cost gap between different products.

Nor can the gap be credited to open weights. K3 had its API price by 17 July, and its weights repository licence was committed on 27 July. The cheap price came first, the licence ten days later. MiniMax launched M3.1-Flash as an API on 19 October 2026 with a promise rather than a release: “并将于稍后开放权重” (meaning: the weights will be made available later). Cheap Chinese APIs and open Chinese weights can arrive together. They are still two different things, and a buyer can have the first without the second.

The bill is not the token price

Cost is measured with one rule of thumb, the Answer Price: price per token × tokens per attempt ÷ success rate. It is the cost of a finished task, not of a token. We have no measured success rate for the task below, so everything that follows is cost per attempt. We can compare token prices and token use, but not the cost of a successful answer.

Take one hypothetical task: 16,000 tokens of input, such as a long document or a codebase excerpt plus instructions, and 1,000 tokens of output. These counts are invented for illustration and belong to no real buyer. At list prices, the modeled cost per attempt is USD 0.105 on Opus 4.8 and USD 0.063 on K3. On OpenAI’s o1, whose list price on 1 November 2026 was USD 15 / 60 per million input and output tokens, it is USD 0.30.

Now add one thing a per-token price does not show. A cache hit happens when the start of a request matches text the provider has just processed, such as the same long system prompt or the same files. The provider reuses that work and bills those input tokens at a lower cached rate. K3’s cached-input rate is USD 0.30, a tenth of its normal input price. Apply it to the same task, varying only how much of the input qualifies:

• With no cache hits, the K3 task costs USD 0.063.

• At half the input served from cache, it costs USD 0.041.

• At 90 percent, it costs USD 0.024, which is 38 percent of the uncached bill.

Moonshot says its own coding workloads see cache hit rates above 90 percent. That is a company statement about its traffic, not a promise about yours. On this task, moving from no cache hits to a 90 percent share on K3 saves nearly as much as switching from uncached Opus 4.8 to uncached K3. The cache share depends on how a buyer structures their requests, and it does not appear on any price list.

The second lever is how many tokens the model uses. Start from the uncached bills and keep the task’s shape at sixteen parts input to one part output. K3 could then spend 1.67 times as many tokens as Opus 4.8 before the two bills matched. Hold K3’s input at 16,000 tokens, and its output could grow from 1,000 to 3,800 tokens before it cost what the Opus reference costs. That margin sounds comfortable until you see how much token use moves. On the Intelligence Index, K3 used 21 percent fewer output tokens than its predecessor K2.6. On Artificial Analysis’s agentic workload, AA-Briefcase, K3 used 120,000 output tokens per task against 42,000 for K2.6, 2.86 times as many, across 83 turns, at an average of USD 10.57 per task. There, K3 cost more per task than Opus 4.8. The same model, measured by the same evaluator in the same month, used fewer tokens than its predecessor on one workload and almost three times as many on another. Those two results describe different kinds of work and should not be averaged, and a list price says nothing about either.

Read more

Inside China’s Machine · Tokenomics, Episode 7 of 10

The Answer Price = price per token × tokens per attempt ÷ success rate
In long-context agent work, the memory a server stores and re-reads for every token sets serving cost more than model size does, and the public record cannot yet show whether the smaller caches China’s labs have built reach the price list.

On 10 September 2026, DeepSeek released V4.1-Flash, a replacement for its cheaper model. The Chinese release gave memory a section of its own: “新一代模型大幅减少了 KV Cache 缓存的大小，与上一代模型相比，对 HBM 的需求减少到 1/4，对 SSD 的需求减少到 1/8。在 Agent 使用场景中，缓存命中的费用往往占比较高，对 KV Cache 的压缩大幅降低了 Agent 类任务的使用成本。” In our translation: the new model substantially reduces the size of the KV cache; compared with its predecessor, demand for HBM falls to one quarter and demand for SSD to one eighth; in agent use, cache-hit charges often make up a large share of the cost, and compressing the KV cache substantially lowers the cost of agent tasks.

A lab selling a model told its customers that the thing it had made cheaper was the stored memory of the conversation.

Here is our view. In long-context agent work, the cost of a token depends less on how many weights a model has than on the state the server must hold and re-read for every token it writes: its look-back over the transcript. Labs in China and elsewhere have shrunk that state, and their own reports show that shrinking it lowers what long-context work costs to serve. DeepSeek, Kimi, MiniMax and Qwen have published numbers on it. Whether this gives China’s labs a lasting edge on price cannot be known from public accounts yet. The reason is specific: the saving and the price list never move alone. Architecture changes, retraining and tariff changes arrive in the same week, and no source reviewed here gives the cost figure separating them.

The look-back, measured in bytes

Start with what a language model does when it writes. It produces one token at a time. A token is the unit a model reads and writes, roughly a word fragment; how much text one token covers depends on the model family’s tokenizer, so a token in one family is not the same amount of text as a token in another. To choose each new token, the model runs a step called attention: its layers compare the new position with earlier positions in the context and decide what to weight. Doing that from scratch each time would mean recomputing the whole transcript for every new token. Servers avoid this by keeping, for each earlier token, a small record of what attention needs from it. That record is the KV cache. The name comes from the “keys” attention compares against and the “values” it combines. The cache grows with every token of context, and while the model is using it, it lives in the GPU’s high-bandwidth memory, HBM, the scarcest and most expensive memory in the machine.

The weights are the model’s learned numbers. They take a fixed amount of memory per server: loaded once and shared by every user on the machine. The cache is different. It belongs to one conversation, it grows as the conversation grows, and at long contexts, or across many users, it can exceed the room the weights take. So the question for a long-running agent is how many bytes each model stores per token of context, more than how big the model is.

Those bytes can be calculated from each model’s published configuration. Figure 1 does this at one common precision, BF16, which spends two bytes per stored number, for a single sequence. The result is logical stored state. It is not what a server allocates at runtime, not a bill, and not a measure of quality.

The figure shows that model size and memory per token barely track each other. Llama-3.1-70B and Qwen3-32B store state the plain way: every layer keeps keys and values for every token, with eight shared “KV heads”, sets of keys and values that several attention heads can use, to save some room. That is how they reach 40 and 32 gibibytes (a gibibyte is about 1.07 billion bytes) at 131,072 tokens of context, for one user. The models below them on the chart use one of three tricks to keep less.

The first is the latent. DeepSeek’s multi-head latent attention, or MLA, does not store full keys and values. It stores a compressed summary per token, a “latent” vector of 512 numbers plus 64 for position, and reconstructs what attention needs when that state is used. That is how a model as large as DeepSeek-V3 lands at 70,272 bytes per token.

The second is the window. Some layers look back only over the last stretch of tokens and forget the rest. Once the window is full, they hold a fixed amount of state however long the conversation runs. Gemma 3 uses “5 local layers for every global layer, starting with a local layer as the first layer of the model,” with a local window of 1,024 tokens. gpt-oss alternates full layers with windowed ones; in the model card’s words, “the bandwidth is 128 tokens.”

The third is linear attention. Instead of a record per token, a layer keeps a running summary of fixed size and updates it as each token passes, the way a running total replaces a list of receipts. Qwen3-Next does this in 36 of its 48 layers, which is why its per-token growth is the smallest on the chart. Those layers still need a fixed block of recurrent state, even for a short conversation; the saving grows with the transcript.

Two readings follow, and the second is the one most readers miss. First, the design choice moves the memory requirement by more than an order of magnitude at the same context length, which no choice of model size does on this chart. Second, the leaders are not all Chinese. At 131,072 tokens, OpenAI’s gpt-oss-120b stores 4.504 GiB against DeepSeek-V3’s 8.578 under these assumptions, and Qwen3-Next stores less than both. The comparators are already optimised designs, not a naive baseline that China beats by default. The design space is shared.

What shrinking it buys, in the labs’ own numbers

Fewer bytes per token matter because HBM limits both how many conversations a GPU can serve at once and how quickly it can read back what it holds. A smaller cache means more users per GPU, longer contexts per user, or both. The evidence that this turns into throughput comes almost entirely from the labs themselves. Each result is a provider-authored report, not an independent audit, and each has its own boundary.

DeepSeek’s V2 paper of May 2024 is where MLA first appeared, and it reported the effect on the whole system: the design cut the “KV cache by 93.3%, and boosts the maximum generation throughput to 5.76 times” against DeepSeek’s own earlier 67B model, on eight H800 GPUs. The baseline was its predecessor, not a matched model using the grouped-query approach Llama uses, and the deployed cache was quantised to an average of six bits. So this is a measure of what changed between two of DeepSeek’s own systems, not of MLA in isolation.

The V3.2 paper of December 2025 went one step closer to money. It ran an actual H800 service comparison of V3.2 against V3.1-Terminus and costed the GPU time at “a rental price of 2 USD per GPU hour.” The long-context cost curves for prefill (reading the prompt) and decode (writing the answer) separate substantially as the token position climbs toward 128K. This is modeled rental cost on an assumed rate, not an invoice and not an all-in cost.

By the V4 paper of 2026, DeepSeek reported that at a 1M-token context V4-Pro uses 10% of V3.2’s KV cache and 27% of its per-token compute. That comparison depends on precision and deployment, and it is not a price ratio. Then came V4.1-Flash, the model in the opening release. Its paper reports a global cache of 890 bytes per token, against 3,514 for V4-Flash, a ratio of 3.948. On that accounting, a million tokens of context come to about 0.83 GiB of global cache. The paper’s boundary is narrower than Figure 1’s: it uses mixed precision and excludes the transient sliding-window state. On DeepSeek’s own series, it is the smallest yet.

The other Chinese labs report the same direction at different boundaries. Kimi’s October 2025 Kimi Linear paper reported “reducing KV cache usage by up to 75% and achieving up to 6×” speed against a matched MLA research baseline; the paper’s own Section 6 calls its up-to-6.3x speed figure theoretical, and it is not a number for the deployed K3. Kimi K3, released in July 2026, carried the idea into a flagship: it “以 3:1 比例混合 KDA 与 Gated MLA” (ours: mixes KDA and Gated MLA in a 3:1 ratio), so roughly three linear-attention layers for each latent-attention layer. MiniMax’s MSA paper of June 2026 reported 28.4x fewer attention operations at 1M tokens “under the same head configuration,” with prefill 14.2x and decode 7.6x faster in wall-clock time on H800, for an experimental model. Its M3 launch made a separate claim on a separate boundary, and the two should not be added together. Qwen3.8’s August 2026 paper reported “7.6×and 4.9×attention-module speedups for prefill and decode, respectively” against a dense baseline at 1M. Those are module-level results, not the whole service and not a price.

The closest non-DeepSeek serving measurement reviewed here is a report by LMSYS of 19 August 2026, which served DeepSeek V4-Pro on H20 GPUs. Its capacity-oriented profile at 1M context reached 177.48 output tokens per second per GPU, with four requests per data-parallel rank. That is a throughput operating point, not a latency promise. As modeled arithmetic only, an assumed USD 2 per GPU-hour at that rate comes to about USD 3.13 per million output tokens of GPU rent. That figure leaves out the GPUs doing prefill, replay, storage, CPU, networking, power, idle time, staff and margin. It shows that the throughput is large enough to matter. It does not show what anyone pays to serve.

So the first half of the call stands on the labs’ own evidence: the look-back can be made dramatically smaller, and smaller look-backs raise what a GPU can serve at long context. What remains is whether agents feel it, and whether anyone passes it on.

Read more

On 30 June UBTech, the Hong Kong-listed robot maker, launched a companion robot called the U1 and announced that “cumulative orders for the UWORLD U1 Series had surpassed 13,361 units”, with prices “set from 119,800 RMB”. Two months later its interim results repeated an “annual production target of 10,000 units” for full-size humanoids, and placed the U1 in the same product line as Walker, the robot it sells to factories. Put those two numbers together and the result looks like a company about to build ten thousand humanoids and already holding orders for more. The same report says something smaller. In the first half of 2026 UBTech sold 921 full-size humanoid robots, 9% of the target.

The headline numbers are not wrong. They count three different things: a production target, a tally of reservations for a cheaper consumer product, and a number of robots booked as sales. In the filings, the third count is by far the smallest. UBTech’s humanoid capacity and robot reservations sit well ahead of its paid factory sales. Building and reserving have outrun billing.

This matters to anyone whose plan rests on a unit count. A forecast for reducers, actuators or robot hands, a capacity investment, or a valuation that multiplies units by a price all inherit whichever of the three numbers they started from. Those numbers measure different things and move at different speeds.

Three counts, one line in the accounts

Take the target first. “10,000 units” is the company’s own production target, a statement about what it intends to build. Its history matters. In a release of 17 November 2025, UBTech was “targeting an annual capacity of 5,000 industrial humanoid robots by 2026, scaling to 10,000 units by 2027”. By the end of 2025, according to its annual results, it had “annualized production capacity of over 6,000” full-size humanoids. By the interim report, 10,000 had become the annual production target for this year. A capacity goal for 2027 arrived a year early as a production goal. Neither phrasing is a sales forecast. A plant built to make ten thousand robots says nothing about who will buy them.

Next come the orders. The 13,361 units are U1 reservations taken at launch. Jiemian, the Shanghai business outlet, reported in September that the U1 presale deposit was RMB 3,000 and was refundable without condition before 15 July. A refundable deposit of that size costs the buyer little and commits them to very little, roughly what a phone pre-order commits you to. The U1 is also a different machine from the one that made UBTech’s name. HK01 reported list prices of RMB 119,800 for the half-torso U1 Lite and RMB 169,800 for the full-body U1 Pro, and the company pitches the series at “commercial and consumer companionship scenarios”. Walker, the factory robot, sold at about RMB 600,000 on average in the first half, according to a Daiwa note relayed by the financial site 10jqka (Tonghuashun).

Last come sales. Of the three numbers, only the 921 units are booked as revenue: RMB 590.3 million in the first half, up 1,445% year on year. No U1 revenue appears in that half.

Following the U1 from launch night to the factory gate

The U1 shows best how far apart the counts sit, because its record runs from launch night to the start of deliveries.

On 1 July, the day after launch, HK01 reported that the chief executive aimed to complete delivery of the 10,000-plus units by 31 December. On the August results call, according to Jiemian, the chief financial officer put this year’s deliverable volume at about 1,500 to 2,000 units. Against 13,361 orders, that is a conversion of 11% to 15% within the year. In October a vice president told 21st Century Business Herald that the U1 had been reaching end customers in small batches since September, and that a cheaper edition was planned for the fourth quarter. A reported 1,500 to 2,000 deliveries against 13,361 orders, together with a cheaper model on the way, describes a launch whose reservation list ran far ahead of expected shipments this year. The company may not have expected every reservation to ship.

The balance sheet fits that picture. Deposits of RMB 3,000 on all 13,361 reservations would come to about RMB 40.1 million. Contract liabilities, the line where a company records customer payments for goods it has not yet supplied, rose by RMB 42.2 million over the half. The match is suggestive, not proof, because that line covers every product and includes a subsidiary acquired in 2026. Even read generously, it describes some tens of millions of yuan of refundable deposits, not a backlog.

Inventory also grew. At the parent company, which excludes the subsidiaries, inventory doubled in six months, from RMB 282.5 million to RMB 565.7 million. Across the group it rose 71%, and finished goods rose 57% to RMB 428.0 million. A company preparing a second-half ramp would carry more stock, so this does not show weak demand. It does show money tied up in robots before they become sales.

From the guidance to the units

UBTech’s interim report contains no numerical revenue guidance. The guidance that circulates comes from Daiwa, which reported on 9 September that the company reaffirmed full-year revenue guidance of RMB 3.5 billion. If that figure is right, and if the CFO figure Jiemian reported is right, the two together say how many full-size robots the year can hold.

The arithmetic runs as follows. First-half revenue was RMB 1,269.1 million, which leaves RMB 2,230.9 million for the second half. Hold the other product lines at their first-half pace and the acquired subsidiary at its run-rate, and full-size humanoids need to bring in roughly RMB 1,410 to 1,463 million in the second half. Take out 1,500 to 2,000 U1 robots at about RMB 106,000 to 150,000 each. Both U1 prices are list prices assumed to include 13% VAT and are shown here without it; neither is necessarily the price actually realised. What remains supports about 1,730 to 2,170 second-half Walker-class robots at RMB 600,000 to 641,000. Add the 921 from the first half, and full-size sales for 2026 come to about 4,270 to 5,000 units, or 43% to 50% of the target. A third to two fifths of them would be companion robots. The total shifts by about 50 to 90 units depending on whether the acquired subsidiary is consolidated from March or April. The conclusion does not shift.

The result rests on two press-reported figures and is not a statement by the company. The guidance figure can also be read the other way: where the first calculation fixes U1 deliveries at the CFO’s 1,500 to 2,000 and solves for Walker-class robots, this one holds Walker at its first-half pace and solves for how many U1 reservations the guidance can pay for. If all 13,361 U1 orders were delivered this year and Walker continued at its first-half pace, revenue would be about RMB 4.0 to 4.6 billion, 1.16 to 1.32 times the reported guidance. On those assumptions, RMB 3.5 billion leaves room for roughly 5,500 to 8,200 of the 13,361 reservations to ship in 2026, 41% to 62%, so not all of them.

When units and revenue stop moving together

The U1 matters for a second reason. It changes what one unit is worth. In 2025 UBTech sold 1,079 full-size humanoids for RMB 820.6 million, about RMB 761,000 each. In the first half of 2026 the figure was about RMB 641,000, close to the RMB 600,000 average that Daiwa reports for Walker S2, down from RMB 700,000 to 800,000. The interim report puts a robot starting at RMB 119,800 into the same line. Every U1 that ships pulls the average down. The cheaper edition planned for the fourth quarter would pull it further.

A line that blends a factory robot with a companion robot priced at about a sixth as much can report rising units without a matching rise in revenue. Anyone who sizes the market by units, which is most supply-chain forecasting, needs the product mix as well as the count. Units, read with the design of each model, tell a forecaster how many reducers and hands to expect. Revenue tells you what customers are willing to pay. Since mid-2026 that one line has held two different businesses.

The case for a timing gap

The strongest rival reading holds that UBTech built capacity ahead of sales and expects deliveries to bunch in the second half, so half the target is a timing gap and not a demand gap. Its evidence is real. UBTech ended 2025 with capacity of over 6,000, and its first-half sales, annualised, came to only about 31% of that. A plant described in the press as built for ten-thousand-unit capacity reportedly began operating on 12 September. The company said in November 2025 that orders for Walker S2 had passed RMB 800 million. On this view, 2027 converts what 2026 booked.

That reading may prove right. Two disclosures would help test it at the annual results, which last year came on 31 March. The first is full-year 2026 full-size units broken out by product. The second is any figure for U1 refunds or cancellations. If second-half Walker-class units come in well above two thousand, and U1 deliveries convert most of the 13,361 into robots and not refunds, the timing reading gains ground. If units land between two fifths and half the target and a third to two fifths of them are companion robots, the target measured what the company meant to make, not what it would sell, as its wording always allowed.

I expect the gap between what Chinese humanoid makers announce and what they bill to persist for some time. Capacity can be built on a budget, and reservations can be taken in a single launch night. Paid deployment needs a customer whose factory runs better with the robot than without it. That view would be wrong if industrial units began to track the targets within a year or two of each target being set.

Treat the next robot announcement the way the filings treat this one. Ask whether it counts what the company will build, what customers have reserved, or what customers have bought. For UBTech in 2026 the three answers are 10,000, 13,361 and 921. Of the three, only the smallest is revenue.

Sources

1. UBTech Robotics, Interim Report 2026 (24 September 2026), https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0924/2026092401451.pdf, captured 2026-11-04. First-half revenue by product line, full-size humanoid revenue and units, the production target, balance-sheet lines.

2. UBTech Robotics, Interim Results Announcement 2026 (28 August 2026), https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0828/2026082801691.pdf, captured 2026-11-04. The first-half results the interim report repeats.

3. UBTech Robotics, Annual Results Announcement 2025 (31 March 2026), https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0331/2026033102607.pdf, captured 2026-11-04. 2025 full-size humanoid revenue and units; year-end capacity.

4. UBTech release, “UBTech Launches UWORLD U1” (1 July 2026; the launch event was on 30 June), https://www.prnewswire.com/ae/news-releases/ubtech-launches-uworld-u1-the-worlds-first-full-size-mass-produced-ultra-bionic-humanoid-robot-302815285.html, captured 2026-11-04. The 13,361 orders and the entry price.

5. UBTech release, “Walker S2 Begins Mass Production and Delivery” (17 November 2025), https://www.prnewswire.com/news-releases/ubtech-humanoid-robot-walker-s2-begins-mass-production-and-delivery-with-orders-exceeding-800-million-yuan-302616924.html, captured 2026-11-04. The capacity goals and the RMB 800 million of Walker S2 orders.

6. Daiwa note as reported by the financial site 10jqka (Tonghuashun), 9 September 2026, https://stock.10jqka.com.cn/20260909/c679736946.shtml, captured 2026-11-04. The reaffirmed RMB 3.5 billion revenue guidance and the Walker S2 average selling price.

7. Jiemian, “UBTech U1 delivery” report, September 2026, https://m.jiemian.com/article/15133669.html, captured 2026-11-04. The chief financial officer’s 1,500 to 2,000 U1 deliveries and the RMB 3,000 refundable deposit.

8. HK01, report on the U1 launch (1 July 2026), https://global.hk01.com/即时中国/60365658/, captured 2026-11-04. U1 Lite and U1 Pro list prices; the chief executive’s delivery aim.

9. 21st Century Business Herald (21世纪经济报道), 14 October 2026, https://m.21jingji.com/article/20261014/herald/46d7d92d0820bda4be76d82b051b2e48.html, captured 2026-11-04. The vice president’s account of small-batch U1 deliveries and a cheaper edition.

Inside China’s Machine · Frontier, Episode 6

Disclaimer: This article is for informational and educational purposes only. It is not investment, financial, legal or tax advice, and it is not a recommendation or solicitation to buy or sell any security. Do your own research, and consult a licensed professional before making investment decisions.

On 16 July, Moonshot released Kimi K3, and for a few weeks China’s best AI model looked closer to America’s frontier than it had in almost a year. A July account of the AI researcher ’s argument put the gap at roughly three to five months, down from a debated six to nine.

’s capabilities index puts models from every lab on a single scale. In that file, Kimi K3 scores 158.33. On 17 October DeepSeek released V4.1 Flash 1017, which scores 158.83. That is half a point higher, and it makes DeepSeek’s model the best Chinese model in Epoch’s file. By the measure that matters for anyone planning around it, DeepSeek’s model is also further behind.

The measure is time. An American model first reached Kimi K3’s score 133 days before Kimi K3 shipped. By the time DeepSeek’s newer, slightly better model shipped, America had already offered that level for 226 days, about seven months. Seven months is also the average Epoch itself reported in January: “Chinese models have trailed US capabilities by an average of seven months, with a minimum gap of four months and a maximum gap of 14.” Kimi K3 caught the gap near its floor. On two Chinese models and five dates, “nearly caught up” describes a moment in July, not a trend.

This matters to anyone whose decision assumes convergence. You might be sizing demand for domestic chips, choosing which models a product will run on, or arguing about whether export controls are working. Each of those decisions needs to know how far behind the top Chinese model sits, and whether that distance is shrinking. On today’s evidence it is not shrinking.

The target did not move

The calculation is simple. Take a Chinese model’s score. Find the first American model that scored at least as high, and count the days between the two releases. For Kimi K3, the first American model above 158.33 was OpenAI’s GPT-5.4 Pro, released on 5 March at 159.81.

For DeepSeek’s October model, the answer is the same GPT-5.4 Pro. The American model being chased did not change. The Chinese score barely changed. Only the release date changed. DeepSeek shipped three months after Kimi, so the lag grew by those three months, 93 days, almost exactly. On a clock, matching last spring’s American level in October puts you further behind than matching it in July. A lab that improves by half a point while the calendar moves a quarter of a year has lost ground.

The March level is still out of reach. No Chinese model in the file has reached 159.81, so on 31 October that score has been American-only for 240 days. Epoch’s 90% interval for DeepSeek’s model runs from 156.71 to 162.20. That band takes in 159.81, so the shortfall from the March level sits inside the margin of error. The gap at the very top does not. The upper end of DeepSeek’s band is below 165.88, the lower end of the interval for GPT-6 Astra, the best American model.

Who shipped last

A second measure compares scores rather than dates: at each date, the American best to date minus the Chinese best to date. This points gap tells the same story with more noise. It was 10.40 at the end of December and 11.70 at the end of March. On 16 July, the day Kimi K3 arrived, it was 4.78. By the end of September it was back to 10.73, and on 31 October it stood at 10.23. Ten months after December, the gap is almost exactly where it started.

Each move tracks a release. GPT-5.4 Pro widened the gap in the spring and Kimi K3 narrowed it in July. Then three American models arrived between July and early September, the last of them GPT-6 Astra on 3 September, and lifted the American ceiling from 163.11 to 169.06. Between July and October the Chinese ceiling rose by half a point. July’s narrow gap was real. It was also the moment just after a Chinese release and just before three American ones.

What a seven-month lag sells

If the best Chinese model usually sits behind the American frontier, think about what a Chinese lab offers a buyer who could also use the American model. It is not the top of the capability curve. It is something else: a licence, a price, or availability where American models are not offered. The licences in Epoch’s file already differ in this way. DeepSeek’s October model comes with open weights and no restriction on use. Kimi K3’s weights are open but non-commercial. Alibaba’s Qwen 3.8 Max is available only through its API. The licences show the labs differing on terms; whether that is a strategy is an inference, since the index measures neither pricing power nor sales.

For a forecast, the consequence is narrower and firmer. Some plans assume Chinese models are converging on the frontier. They include revenue models for Chinese labs, valuations that price in parity, and export-control arguments that treat July as proof the controls failed. Leaning on July alone means choosing the single reading most favourable to that assumption. A plan that needs the frontier itself, rather than a model that is good enough on attractive terms, should assume, on this record, a lag of seven months or more, growing by the day until a Chinese model reaches the level America passed in March.

The case for July

The strongest rival reading says the long-run lag is shrinking and July showed where it is heading. On this view, one more Chinese release closes the distance. The evidence that could support it is easy to name. A new Chinese model released now and scoring at least 163.11, the score of OpenAI’s April model GPT-5.5 Pro, would cut the lag to about six months. A model scoring at least 169.06 would close the gap at the top altogether. Neither has appeared in Epoch’s file.

I expect the lag at the top to keep swinging with release timing rather than closing in a straight line. That view would be wrong if Chinese releases began, one after another, to land within a few months of America first reaching the same level, the way Kimi K3 did once.

Both headlines were true. In July China’s best model was about four months behind. In October it is about seven, and the better model is the one further back. Nothing collapsed between the two readings; DeepSeek simply shipped in October rather than July. A claim of parity is a photograph of whoever shipped last.

Sources

1. Epoch AI, Epoch Capabilities Index scores (eci_scores.csv; latest model date 2026-10-26), https://epoch.ai/data/eci_scores.csv, captured 2026-10-31 and re-pulled 2026-11-04 (byte-identical). Every score, interval, release date and licence group used here; the day lags and the points gaps are computed from it.

2. Epoch AI, “US vs China ECI” data insight (2 January 2026), https://epoch.ai/data-insights/us-vs-china-eci, captured 2026-11-04. The January average of seven months, with a minimum of four and a maximum of 14.

3. AI Weekly, “Moonshot’s Kimi K3 Compresses Open-Weight Gap to 3-5 Months” (published 21 July 2026), https://aiweekly.co/alerts/moonshots-kimi-k3-compresses-open-weight-gap-to-3-5-months, captured 2026-11-04. A secondhand summary of Nathan Lambert’s argument.

4. Epoch AI, ECI public code and documentation (eci-public README), https://github.com/epoch-research/eci-public, captured 2026-11-04. The interval columns of the scores file are 90% intervals (the 5th and 95th percentiles).

Inside China’s Machine · Frontier, Episode 5

Disclaimer: This article is for informational and educational purposes only. It is not investment, financial, legal or tax advice, and it is not a recommendation or solicitation to buy or sell any security. Do your own research, and consult a licensed professional before making investment decisions.

Inside China’s Machine · Tokenomics, Episode 6 of 10

The Answer Price = price per token × tokens per attempt ÷ success rate
Where compute is scarce, China’s AI providers now sell capacity by the hour, by queue position and by service level, so the race to cut token prices has become a contest over who can deliver at the hours customers want.

On 24 April 2026, DeepSeek published the launch prices for the preview of its V4 models. The graphic had the usual rows: input, output and cached input, in yuan (per million tokens, by DeepSeek’s convention). Tokens are the units of text a model processes or produces; cached input is material the service can reuse rather than process afresh. Underneath it was a footnote: “受限于高端算力，目前Pro的服务吞吐十分有限”. In our translation, constrained by high-end compute, Pro’s serving throughput is currently very limited. It added that Pro’s price should fall sharply once Ascend 950 supernodes ship in volume in the second half of the year. A price list is meant to tell you what something costs. This one also told you that you might not be able to buy much of it.

Our view is that China’s AI price war is over wherever compute is the binding constraint: wherever the machines available cannot serve all the work customers want done. It has not ended in a truce. It has turned into a different contest, over capacity and service level. In the race to the bottom on token prices, when ByteDance launched Doubao in May 2024 at 0.8 yuan per million input tokens and Alibaba halved qwen-max prices that September, from 40 to 20 yuan for input and from 120 to 60 for output, the number that mattered was the price per million tokens. The price notices of 2026 carry other things: a clock, a priority fee, a rationed plan, a rise in the list price of the compute underneath. What a buyer pays now depends on when the work runs, how fast it runs and how reliably it gets admitted at all. The obvious next question is who keeps the money from this, and the honest answer is that the public accounts cannot yet say.

The clock that flipped

The clearest case is DeepSeek’s own price page, which has now kept time in two opposite directions.

In 2025, DeepSeek’s discount came at night. Archive snapshots of its page from March to June 2025 show cheaper rates between 00:30 and 08:30 Beijing time, when, DeepSeek said, load was low and it moved nodes to research. DeepSeek’s price page, as archived on 21 August 2025, announced the night discount’s withdrawal for V3.1 from 6 September 2025. Archive snapshots record what a page said on a date, not when a policy began, but the direction is clear.

The 2026 clock runs the other way. On 13 August DeepSeek announced a peak and off-peak tariff, effective 17 August, and stated the rule plainly: “闲时价格为高峰时段价格的一半” (our translation: off-peak prices are half of peak prices). Peak hours are 09:00 to 12:00 and 14:00 to 18:00 Beijing time. The 2025 night discount sold spare capacity cheaply. The 2026 tariff charges a premium for Beijing daytime hours, when, by DeepSeek’s own 2025 account, its load is highest. Both make quieter hours cheaper, but the emphasis has shifted: from selling spare capacity at night to charging more for scarce capacity when customers most want it.

On 10 September DeepSeek replaced V4-Flash with V4.1-Flash and cut its rates. The headline is accurate but needs care. V4.1-Flash is a new architecture, so this is a model change, not a like-for-like price cut. Measured against the August V4-Flash off-peak rates, input fell from 1.5 to 1 yuan, output from 4.5 to 4 and cached input from 0.05 to 0.02. The clock stayed. The new peak rates are exactly double: 2 yuan input, 8 output, 0.04 cached. Peak is now defined as working days only, so weekends are off-peak, and the current page, observed on 1 November, also excludes statutory holidays.

Set that “cut” against the April preview, which listed Flash at 1 yuan input and 2 output (down from V3.2’s 2 and 3 in January). After September’s cut, off-peak input is back where the preview started. Off-peak output is double the preview rate, and peak output is four times it. Each step involves a different model or label (preview, general release, replacement), so this is not one product repriced three times. Still, a user following Flash through these changes saw input and output prices rise, then partly fall, while cached input prices fell throughout and the tariff acquired a clock. The September cut was a partial retreat on output prices that had moved up. It was not a return to the 2024 race.

DeepSeek’s explanation for the cut points at the same constraint as the April footnote. “得益于模型架构的创新，DeepSeek V4.1 Flash 能够以更低的成本服务更多的用户”, in our translation thanks to model architecture innovation, V4.1 Flash can serve more users at lower cost. The company added that its smaller KV cache (the working memory a model keeps so it can use earlier parts of a request without recalculating them) cuts high-bandwidth memory (HBM) needs to a quarter and solid-state storage needs to an eighth of the previous generation’s. Those figures describe memory requirements, not total cost or margin. They still show what the company is saving: scarce hardware per user served. When compute is scarce, DeepSeek’s cost story is getting more users onto the same machines.

What the clock does to a bill

The Answer Price is what a buyer spends per successful task. In its simplest form:

answer price = price per token × tokens per attempt ÷ success rate

The multiplication gives the cost of an attempt. Dividing by the share of attempts that succeed spreads the spending on unsuccessful attempts across the successful tasks. Where input, output and cached input have different rates, the cost of each is calculated separately and added together. Token price is one lever of three. The 2026 notices are interesting because between them they pull all three.

Start with the first lever, using DeepSeek’s clock in a scenario. This is arithmetic on list prices, not an invoice. Take a fixed job: 1 million input tokens and 0.1 million output tokens, with no tools, retries, storage or tax. With no cached input, it costs 1.40 yuan off-peak and 2.80 yuan at peak. If 90% of the input is cached, as can happen for agents re-reading context, it costs 0.518 yuan off-peak and 1.036 at peak. Caching cuts the bill a great deal, and the clock still doubles whatever is left. Moving the job to 20:00 halves the bill. That only helps a buyer who can wait. A batch job can. A developer whose coding agent is stuck at 10:30 on a Tuesday cannot, so the clock lands hardest on the most time-sensitive work, which may be the work a buyer values most.

Now the third lever, success rate. From 13 March, Baidu’s Coding Plan began throttling some models at peak hours: “将对高峰时间段内部分模型的调用进行限流处理”, in our translation calls to some models during peak periods will be rate-limited. Baidu said calls might slow, queue or fail. A failed call never appears as a price change, but it lowers the success rate. Where a failed or queued call is billed, or has to be paid for again by a retry, dividing by a smaller success rate raises the Answer Price while the list price stays the same. The notice does not say whether failed calls are billed. MiniMax offers the reverse: a priority tier at 1.5 times its standard tariff that buys admission priority and a lower chance of request failure. That tier is a separate service, not a rise in the list price, and the fee buys a better chance that the request gets in at all. A request being admitted is not the same as a task succeeding, but it is a precondition for success. MiniMax is charging for that precondition directly.

Zhipu applies the clock to quota rather than price. Under its coding plan of 15 March, GLM-5 and GLM-5-Turbo consumed quota at three times the base rate at peak and twice the base rate off-peak (”高峰期3倍，非高峰期2倍”), though as a promotion GLM-5-Turbo counted at only 1x off-peak through April. On 30 July it began counting weekends as off-peak all day, and its plan page, seen on 1 November, puts peak at Monday to Friday, 14:00 to 18:00. A buyer on a flat monthly plan therefore gets less work from the same allowance at the busy hour, without the price ever being called a price.

The middle lever, tokens per attempt, is where providers claim to be helping. DeepSeek’s 17 October revision left “API 调用方式与价格保持不变” (invocation and prices unchanged) and said the model now shortens task runtime and output length. If it uses fewer chargeable tokens to achieve the same result on a buyer’s own workload, the bill falls at an unchanged list price. It is a company statement with no invoice comparison behind it. The lever can also move the other way. In a scenario with a fixed token mix, a 20% cut in unit price is cancelled by 25% more chargeable tokens, and if the tokens double, the bill rises to 1.6 times its earlier level. None of the filings and notices cited here gives a matched series of Chinese customer bills, so nothing here shows that agent bills rose. The point is narrower: by 2026 the headline price was one lever of three, and the notices were moving the other two.

Read more

Inside China’s Machine · Tokenomics, Episode 5 of 10

The Answer Price = price per token × tokens per attempt ÷ success rate
The highest reported margins in AI sit with chips, memory and the leading foundry, not with GPU renters or model makers, and no public filing can yet show whether that lasts.

In the second quarter of 2026, CoreWeave took in USD 2,575 million. CoreWeave rents out AI computing capacity. On those sales it reported a GAAP operating loss of USD 49 million, an operating margin of -1.903%. Two months later, Micron closed its fiscal year on 3 September. Micron makes the memory those chips cannot run without. Its cloud memory unit, which sells to the hyperscalers, the largest cloud computing providers, reported an operating margin of 76%, rounded and unaudited, for the final quarter. Both are operating margins. One belongs to a whole company and the other to a business unit, so they are not twins. Still, they answer the same plain question: after operating costs, including depreciation on equipment, how much of each dollar of sales is left?

At the memory maker, about three quarters is left. At the firm that buys chips and rents them out, slightly less than nothing is left.

Today the highest reported margins in the AI stack sit upstream, at the layers where capacity binds: chips, memory and the foundry. Upstream means the suppliers of the physical inputs; downstream means the businesses that use those inputs to sell computing, models and applications. The firms that rent that capacity, and the firms that sell models on top of it, report thin or negative operating margins; applications disclose too little to judge. Owning GPUs is not the same as keeping the surplus, and neither is selling the model that runs on them. This describes where reported margin sits now, as of the quarters that closed by October 2026. It is not a forecast. The second half of the question is whether the surplus stays upstream once new capacity arrives. The honest answer is that public accounts cannot tell yet, and the reason is worth understanding.

Why margin may sit upstream

Start with the plumbing. A company that wants to sell AI answers needs accelerators, the chips that do the computation; high-bandwidth memory, or HBM, to feed them data fast enough; and wafers from foundries, the factories that manufacture chips. Those are bought in advance, paid for in cash, and turned into revenue over years through rental, model access or application fees. Our reading: whoever holds the scarce capacity has leverage over the price at purchase, and everyone further down pays it, then competes for the buyer.

Memory and cloud sellers say supply is short. Microsoft, on its October call: “Customer demand continues to exceed available supply.” Oracle: “Customer demand for AI Cloud Training and Inferencing Services continues to grow faster than supply.” Training builds a model; inference runs it to answer a request. Samsung said its memory business set another record “by managing its product mix, with a focus on AI products, amid capacity constraints.” SK hynix disclosed on 28 October that long-term agreements cover “more than 50% of total production capacity for 2027 and 2028.” That statement is more limited than “sold out”: it covers more than half of capacity, not all of it, and long-term agreements are not fixed-price guarantees. None of these statements shows that all AI capacity is sold. What they do show is the seller’s position: at the moment of purchase, the buyer is the one waiting.

The margins match that position. They are reported under different definitions, and here each one keeps its own. Gross margin is what remains after the costs assigned to delivering the product. Operating margin also deducts operating expenses such as research and administration. Adjusted measures exclude further costs chosen by the company.

At the chip layer, NVIDIA’s consolidated GAAP gross margin was 71.1% for the fiscal year ended January 2026 (audited). For the quarter ended July 2026 it was 75.0% (unaudited; the company stated: “For the quarter, GAAP and non-GAAP gross margins were both 75.0%.”). That is the whole company’s mix, though Data Center made up USD 89,023 million of the quarter’s USD 96,221 million in revenue. Not every chip maker is so placed. AMD’s Data Center segment, which combines server CPUs with accelerators, earned a segment operating margin of 31.304% in Q2 2026 (calculated). That is a different measure from NVIDIA’s gross margin and is not ranked against it. It does show that “chips” is not one uniformly rich layer.

At the memory layer the figures are the most striking in the stack. Micron’s gross margin for fiscal 2026 was 80.716%, calculated from audited amounts. For fiscal 2025 it was 39.791%. Micron’s filing does not split the rise between higher prices for the same products and a more profitable product mix. SK hynix reported an operating margin of 80%, rounded (79.657% calculated), for Q3 2026, across all its DRAM and NAND memory, not HBM alone. Samsung’s chip division, which combines memory, foundry and logic, earned an operating margin of 72.4% (calculated).

At the foundry, TSMC reported a Q3 2026 gross margin of 66.4% and an operating margin of 58.7%, up from 59.9% and 50.8% for fiscal 2025 (audited). That covers all of its wafers, not only the AI ones.

The hyperscaler clouds sit in between, and their margins are hard to read. AWS earned a segment operating margin of 38.650% in Q3 2026, against 34.642% a year earlier. Google Cloud earned 36.214%, against 23.712%. Microsoft’s new Agents and Infra segment earned 52.813% in its first reported quarter. That segment was created in the quarter ended September, so it cannot be joined to the old Intelligent Cloud series: the businesses being measured have changed. Each of these figures combines training, serving, rental and in some cases applications. None tells us the margin on inference alone. Oracle discloses no margin at all for its cloud infrastructure, and dividing Oracle’s corporate profit by cloud revenue would mix profit from the whole company with sales from only part of it.

Read each panel separately, from top to bottom, and the pattern is clear. Outside China, most of the dots for the layers that make the scarce physical things sit between the high fifties and the high eighties, under their own definitions; AMD’s Data Center segment, at 31.304%, is the exception. The GPU clouds and the model maker that file operating results report margins near zero or below at the operating line; the hyperscaler segments are higher but mix many workloads. The third panel does not overturn that pattern: it shows that GPU clouds report much higher margins when depreciation and other excluded costs are left out. A missing dot does not mean zero profit, only that the measure is not disclosed.

Read more

Inside China’s Machine · Tokenomics, Episode 4 of 10

The Answer Price = price per token × tokens per attempt ÷ success rate
Where compute is tight, AI providers now price the working hour and the promised speed on top of the token, and the serving margin plausibly goes to whoever keeps machines busy at that speed, though no public filing yet shows it.

DeepSeek’s Chinese price page, last modified October 30, 2026, has two prices for each kind of token on the same model. For its Flash model, a million output tokens cost 8 yuan in the “高峰时段”, the peak period, and 4 yuan in the “空闲时段”, the idle period. The page then defines the peak period as Monday to Friday, excluding China’s statutory holidays, from 09:00 to 12:00 and from 14:00 to 18:00 Beijing time. Everything else is idle. That includes nights, weekends, holidays and the two hours in the middle of the day.

The clock follows the Beijing office day closely enough to leave out lunch. Nothing about the model changes at noon. What changes is who wants the machine. For seven hours a weekday, DeepSeek charges double. When the office empties, it charges half.

Half a world away, OpenAI’s price page lists four prices for one model, gpt-6.1-sol. Input costs $1 per million tokens on Batch, $2 on Standard, $4 on Fast and $12 on Ultrafast. The model and the provider are the same, and the price differs by a factor of twelve. The only thing that varies is when the work runs and how quickly the tokens come back.

That is this episode’s call. Where compute is tight, sellers no longer price only the token. They price the working hour and the promised speed. A buyer is paying for a service envelope, meaning when the work runs and how fast it returns, and that envelope has its own price schedule that sits on top of the token price. Who earns the margin from serving it is a separate question. The mechanism points to whoever can keep machines busy at the speed they promised, rather than whoever holds the cheapest chips. The public record cannot yet prove that.

One model, several prices

Once you know where to look, the pattern repeats across providers. Anthropic sells Claude Opus 5.5 at $4 per million input tokens on Standard, $2 on Batch and $8 on Fast. Google’s Gemini 4 Argon is $2 on Standard, $1 on Batch or Flex and $3.60 on Priority. On Amazon’s Bedrock in Sydney, the same DeepSeek v3.1 model is sold at half price on Flex and at 1.75 times the price on Priority. In China, Moonshot’s kimi-k3.1-pro costs twice as much on its Highspeed tier and 60 percent of Standard on Batch. DeepSeek’s own API does the same thing with a clock instead of a tier.

Two things in the figure are worth a second look.

The first is that the multiples hold when prices move. Google has already posted the Argon rates for December 1: Standard doubles to $4, Batch and Flex double to $2, and Priority doubles to $7.20. The base price changes, and the half-price and 1.8-times relationships stay exactly where they were. Google’s Flash list also doubles on a fixed date. Here the doubling leaves the shape of the tariff alone. The seller is adjusting what the token costs and separately holding what speed and patience cost.

The second is Moonshot’s flagship. On kimi-k3.1-pro, Highspeed doubles input, output and cache writes, but a cache hit costs 1.30 yuan per million tokens on both tiers. On the older kimi-k2.7-code, the cache hit doubles along with everything else. So on its flagship, Moonshot has chosen not to charge for speed on the re-reads. That choice matters a great deal for one kind of buyer. Agents mostly re-read their own notes, and those re-reads are cache hits. For an agent on k3.1-pro Highspeed, the effective premium is below 2x, because most of what it reads costs the same either way. The tier’s name says 2x, and the effective premium depends on how much of the workload is cached.

The mechanism behind all this is short. A serving machine earns the most when it runs many requests together in one batch. Batching raises throughput, but each request then waits for the others and comes back more slowly. A request that has to come back fast takes a share of the machine that could otherwise have served several slower ones, so the fast lane is priced at a multiple. Batch and off-peak pricing work the other way round. A buyer who can wait lets the seller fill hours that would otherwise sit idle, and those hours are sold cheaply because unsold machine time cannot be stored for later.

Read more

Zhipu’s interim results for the first half of 2026, filed in Hong Kong on 31 August, make two claims about the same six months. The Beijing model company says the cost of running a token on its platform fell 80% since the start of the year, with domestic chips now part of its main inference fleet. It also says its growth over those months was not bought by cutting prices: “volume is rising, and price is rising.” Further on, the income statement shows the cost of sales, which rose mainly on compute service fees, at more than seven times the figure for the first half of 2025. Research spending rose by a third. Operating costs as a whole rose by about half.

A cheaper token and a smaller bill are different things. At Zhipu, use grew far faster than unit cost fell, and part of the saving went to margin. What a model firm needs in chips and in capital follows how many tokens it serves more closely than what each one costs. A big efficiency gain does not mean a lab will need fewer chips or less capital; on the serving side, it has come with more.

The question matters because Chinese labs cannot buy all the compute they want, and each efficiency result raises the hope, or the fear, that they will need less of it. For someone who holds a position in a chipmaker, funds a model firm or sells cloud capacity, the answer decides whether a cheaper model shrinks the market or enlarges it. The case rests on companies’ unaudited interim figures, Zhipu’s auditor-reviewed; DeepSeek, the third example, publishes no financials.

Where the 80% went

Start with what Zhipu sells. Most of its revenue now comes from its open platform and API, where developers and companies pay for model calls by use. That line brought in RMB 825.2m in the half-year, up from RMB 29.1m, and made up 86.5% of total revenue of RMB 953.9m. Total revenue was five times the first half of 2025. The company says token calls on the platform rose more than fortyfold from the start of the year to late August. Calls on its Coding Plan, a subscription for programmers, rose more than twenty-three times.

Now the bill. Cost of sales was RMB 702.3m, against RMB 95.5m a year earlier. The filing puts the increase down mainly to higher compute service fees and calls it “basically in line with the revenue growth trend.” Compute for serving tokens sits in that line, and it rose more than sevenfold.

Hold one thing still to see why. If calls had stayed flat, an 80% fall in unit cost would have cut the serving bill to a fifth. Calls did not stay flat. More than forty times the calls at a fifth of the cost each comes to about eight times the serving compute, assuming the mix and quality of the work stayed the same. The group’s cost of sales rose 7.4 times. The two numbers do not match exactly: the call figure runs to late August and compares levels, while the financials compare one first half with another and include local deployment. That does not weaken the argument, because both point the same way: the serving bill rose by a multiple, nowhere near falling to a fifth.

Two other figures could mislead a quick reader. Group gross margin fell from 50.0% to 26.4%, which looks like efficiency running in reverse. The whole fall comes from a change in what Zhipu sells. Local deployment, the business of installing models on a customer’s own hardware, shrank from RMB 161.8m to RMB 128.7m. That business had carried all of last year’s gross profit. The API business that replaced it lifted its own gross margin from −0.4% to 24.6%. Second, Zhipu says its average API selling price roughly doubled while unit cost fell to a fifth. At face value, cost per yuan of revenue fell to a tenth (0.2 ÷ 2): last year the segment spent 100.4 yuan to earn 100, so it would spend about 10, a margin near 90%, far above the 24.6% it earned. The filing does not explain the gap. Larger models, Coding Plan seats, the timing of the period, or what “cost” covers could each account for part of it, and the average price is itself an average across products. Treat the 80% as a direction the company reports, not a margin you can compute.

The bill the 80% never touched

The serving bill is the smaller one. Research and development expense was RMB 2,131.2m, up 33.6%, and the filing attributes the rise mainly to continued, larger investment in model training. It was more than twice the half-year’s revenue. It was most of the RMB 3,114.5m Zhipu spent on cost of sales, selling, administration and research combined, a total up 49.5% on the year. Selling and administrative costs fell. The company spent less on running itself and more on compute.

Inference efficiency lowers the cost of answering a question, not, by itself, the cost of building the model that answers it, and the research line rose. How much of the RMB 2.13bn is training, and how much is staff and everything else, the filing does not say. That is the open question for anyone estimating a Chinese model firm’s capital need. Whatever the split, the serving saving did not stop that line rising.

A second firm, in its own currency

MiniMax Group, also listed in Hong Kong, released its first-half results on 26 August, unaudited and in US dollars. Its figures are not directly comparable with Zhipu’s, but the direction is the same. Revenue rose about 3.8 times and cost of sales about 3.6 times. Research and development rose about 2.4 times, which the company attributes mainly to higher cloud-services spending for training. Its chief executive says token consumption reached 20 times its January level by July. The release gives no unit cost, average price or chip detail. Gross margin rose to 17.9% from 12.1%, which MiniMax credits to improving infrastructure efficiency. The loss for the period under IFRS fell to US$358.0m from US$402.2m; the adjusted net loss, which the company defines itself, went the opposite way, to US$293.0m from US$138.7m. Efficiency arrived as a margin, not a smaller bill.

DeepSeek’s different saving, and the same answer

DeepSeek offers a different technique leading to the same place. In its V4.1 Flash release note of 10 September, the saving is in memory rather than chips. When a model works through a long conversation, it keeps a running record of everything it has already processed so it does not have to recompute it at each step. That record is called the KV cache. In a long agent session, where the model writes code, reads results and tries again for hours, the cache keeps growing. It sits in HBM, the scarce fast memory packaged next to the chip, and in SSD storage. Weights are shared across sessions, but each session’s cache takes its own memory.

DeepSeek says the new generation cuts its HBM need to a quarter, and SSD need to an eighth, of the previous generation’s. The note does not say which model, workload or hardware that comparison uses. DeepSeek’s gloss is that in agent use the charge for cache hits, reusing stored records, is often a large share of cost, so compressing the cache cuts the cost of agent tasks sharply. It is moving all V4 Pro traffic onto Flash at Flash’s price and plans to retire V4 Pro.

DeepSeek states why it cut the price. Because of the architecture, V4.1 Flash “can serve more users at lower cost, so we have lowered its price.” Zhipu moved serving onto domestic chips and retuned its stack. DeepSeek shrank the memory each session occupies. Each lowered the cost of a token. DeepSeek, in its own words, priced the saving into serving more; Zhipu’s calls grew as its prices rose. The release ends by inviting firms that self-host the model to get in touch if they have “2k GPUs and a storage cluster.”

The rival reading

The best rival reading points to Zhipu’s ratios. In the first half of 2025, research spending was 8.35 times revenue. A year later it was 2.23 times. By the company’s account, its loss fell from about 12.4 times revenue to 2.2 times on IFRS, the accounting standard behind the headline loss, and from 9.2 to 2.1 times on its own adjusted measure. On this reading, efficiency is what lets revenue outrun costs. The capital needed per yuan of revenue is collapsing, and the absolute loss will follow.

The ratios are real, and they matter. A firm whose operating costs grow slower than revenue is moving the right way. But a backer funds the absolute gap, not a ratio, and that gap has not closed. The loss for the period under IFRS fell 12.1%, to RMB 2.07bn. More than all of that narrowing came from a non-operating line, the change in value of financial instruments it issued to investors: a charge of RMB 22.1m against RMB 429.3m a year earlier. The adjusted net loss, which the company defines to strip out items like that, rose 12.1%, to RMB 1.96bn. Most ratios improved, yet the adjusted net loss grew.

The next reporting period will bear on this. If Zhipu’s adjusted loss falls in absolute terms while revenue keeps growing, that would be evidence for the rival reading, not proof, and would leave open whether efficiency or price rises did it.

What it changes

If you hold a position in chips or cloud capacity, this half-year gives a narrow but useful result. At one listed Chinese model firm, an 80% cut in unit serving cost, with domestic chips in the fleet, came alongside cost of sales more than seven times larger. MiniMax’s half points the same way, but two firms’ accounts are not a measure of total chip demand. Still, it shows how a serving-side efficiency gain behaves inside an actual business: it looks less like a shock to demand and more like the thing that made more use affordable. If you back a model firm, the capital it needs is set by how fast use grows and by what training costs, and a falling unit cost does not lower it on its own. If you follow export controls, the evidence reaches serving on domestic chips at a scale Zhipu describes as on the order of 100,000 chips, with the makers unnamed. It says nothing about training frontier models on those chips, and it should not be stretched to.

I expect China’s model firms to keep cutting cost per token while their capital need stays set by how many tokens customers buy and by how much they train. That view would be wrong if a firm’s cost of sales stopped tracking the tokens it serves, for example if calls kept multiplying while the compute bill flattened.

Zhipu’s sentence, “volume is rising, and price is rising,” shows why the 80% never shrank the bill. Tokens cost a fifth as much. Zhipu says customers bought more than forty times as many calls. The bill followed the calls.

Sources

1. Zhipu, interim results announcement for the six months ended 30 June 2026 (HKEXnews, filed 31 August 2026), captured 2026-10-26: https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0831/2026083101540_c.pdf. Revenue, cost of sales, research and development, loss for the period and adjusted net loss, segment figures, the company’s account of its cost per token, call volumes and average price.

2. MiniMax Group, results announcement for the first half of 2026 (HKEXnews, 26 August 2026), captured 2026-10-30: https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0826/2026082600680.pdf. Revenue, cost of sales, gross margin, research and development, loss for the period and adjusted net loss, US$ thousand, unaudited.

3. MiniMax, “MiniMax announces first half 2026 financial results” (company release), captured 2026-10-30: https://www.minimax.io/news/minimax-announces-first-half-2026-financial-results-1787744160. The chief executive’s statement on token consumption.

4. DeepSeek, V4.1 Flash release note (10 September 2026), captured 2026-10-26: https://deepseek.com/news/deepseek-v4-1-flash/. The memory comparison, the reason given for the price cut and the invitation to large deployers.

Inside China’s Machine · Frontier, Episode 4

Disclaimer: This article is for informational and educational purposes only. It is not investment, financial, legal or tax advice, and it is not a recommendation or solicitation to buy or sell any security. Do your own research, and consult a licensed professional before making investment decisions.

Inside China’s Machine · Tokenomics, Episode 3 of 10

The Answer Price = price per token × tokens per attempt ÷ success rate
For agent work, model upgrades at a mostly unchanged token price, not token price cuts, now move the cost of a solved task down or up with success and tokens per attempt, and the maker that controls success is positioned to price that saving back in.

On Google’s price page for its Gemini models, read October 26, 2026, two Flash models sit one above the other. Gemini 3.8 Flash is “Our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows.” Gemini 3.6 Flash is “Our previous generation Flash model, balancing speed and multimodal capabilities across general agentic and everyday tasks.” Both carry the same three rates: $0.75 per million input tokens, $0.075 per million cached input tokens, $3.75 per million output tokens. Under each rate on both models there is a second sentence. For input it reads “$1.50 starting January 1, 2027.” Every rate doubles on that date.

DeepSWE v1.1 is a public coding leaderboard with 113 real software tasks, run under one agent harness, mini-swe-agent, and last updated September 22, 2026. It shows what the newer model does with those identical rates. At the vendor’s “high” effort setting, 3.8 Flash burns 72 percent more tokens per attempt than 3.6 Flash. Its attempt costs about the same, $2.36 against $2.21. It solves 74 percent of attempts against 47 percent. So a solved task costs $3.20 on the newer model and $4.73 on the older one, about a third less. The leaderboard’s cost figures for these rows reproduce Google’s list prices to the cent, so this is the price a buyer would actually have paid, not a model of it.

The token price did not move. The answer price fell by a third. And the maker has already written down the date the rates double.

That is the pattern. For agent work, the price per token has stopped doing the work. Model makers mostly hold the token price where it was and ship a better model. The buyer’s bill for a solved task then moves with three things: how much more often the new model succeeds, how many more tokens it burns getting there, and what those tokens cost on average. When success outruns the extra cost of appetite, the bill falls, often by a third or more. When the cost of appetite wins, a better model produces a dearer answer. The saving usually comes from success, and success is something the maker builds and the buyer can buy more of by paying for more effort. So the maker, not the buyer, is in the position to decide how much of that saving to keep.

One bill, three numbers

The Answer Price, the same measure as the cost per solved task in the table below, is what a buyer pays per solved task: price per token × tokens per attempt ÷ success rate. Read the Flash upgrade through it, one lever at a time.

The first lever is token price, and on the price page it is unchanged. The second is tokens per attempt, which went from 12.69 million to 21.88 million. The agent took 166 steps instead of 117. A buyer reading only those two numbers would expect the bill to rise by about 72 percent.

It did not, and the reason is that agents mostly re-read their own transcripts. On 3.6 Flash, 89.4 percent of input tokens per attempt were cache hits. On 3.8 Flash the figure was 98.7 percent. Google bills a cache hit at a tenth of the input rate, so almost all of the extra appetite was billed at the cheapest rate on the page. Averaged over every token the attempt used, the price fell from $0.174 per million to $0.108, a drop of 38 percent, while the price page stayed exactly the same. The token count rose while the average price per token fell. The effective price per token fell because the extra tokens were mostly cache hits, the cheapest kind of token. The cost of an attempt ended up 7 percent higher.

The third lever, success, then did all the work. The success rate rose from 0.467 to 0.738. $2.36 divided by 0.738 is $3.20.

There is a quiet point in that sequence. The usual story about agents says that cheaper tokens invite more thinking: the buyer gets a discount and spends it on longer runs. Here there was no discount to spend. The longer runs came with the model, by the maker’s design, at an unchanged list price. Effort rose on its own, and success decided whether it was worth paying for.

Read more

Inside China’s Machine · Tokenomics, Episode 2 of 10

The Answer Price = price per token × tokens per attempt ÷ success rate
A token count measures what the machine read, not what anyone wanted done, and most of what an agent reads is its own past.

In the last week of September 2026, the API keys that OpenRouter classes as agents used about seven times as many tokens as the keys it classes as human. The routing platform reported this in its data newsletter, “Inferred” edition 03, published October 8. The same newsletter reports what those agents were reading. Just under 90 percent of their prompt tokens in September were cached. That is text the model had already been sent, served again from stored work. In January the share had been closer to 60 percent.

Put another way, of every ten tokens an agent on OpenRouter sent as input in September, about one was not served from cache. The other nine were text the model had already been sent, served again from stored work.

The newsletter computes the share as cached prompt tokens divided by total prompt tokens, so output is left out, although one sentence calls it a share “of all agentic tokens.” It excludes reseller traffic. It describes OpenRouter’s network and nothing wider.

A public benchmark shows what one of those agent runs looks like from outside. DeepSWE v1.1 sets 113 real software tasks under the mini-swe-agent harness, with the page last updated September 22, 2026. On it, DeepSeek’s cheaper model, deepseek-v4-flash, at max effort, took 153 steps per attempt and wrote 108,000 output tokens. Each attempt cost $0.46, and the model solved 53 percent of the tasks (±4 points). The page publishes no input or cached tokens, so it cannot say how much of those 153 steps was re-reading. It does say how many tasks got solved, which no token headline does.

Every cached token is a token. It is billed, at a lower price, and it is served.

That is the problem with the token boom. A token count measures what the machine read, not what anyone wanted done, and as agents take over, the count can grow faster than the work, because most of what an agent reads is its own past. The headlines are a platform’s weekly tokens, a provider’s tokens per month or per day, and China’s national daily figure. If they count cached input, they overstate useful demand, and the more agentic the traffic, the more they can overstate it. The number that measures demand is solved tasks. Until someone publishes that, token growth should be discounted. Anyone who sizes capacity or a business on raw token counts is probably reading the boom too high, by an amount no outsider can yet measure.

Why an agent mostly re-reads

A chat question is one exchange. You send a question, the model reads it and answers. There is no transcript to re-read, so nearly all the input is new.

An agent works in steps. It reads the task, thinks, runs a tool, reads the result, thinks again. A model does not remember the previous step on its own, so at every step the agent sends the whole conversation so far back as input: the task, every earlier thought, every tool result. Each step adds a little new material. The history it carries keeps growing.

Assume each step adds roughly the same amount. Then the 153rd step re-reads 152 steps of history to add one. New input grows in step with the number of steps, and total input grows with the square of it. Double the length of a run and the new material doubles, but the re-reading roughly quadruples.

The published measurements vary with the harness. OpenRouter’s agent traffic was just under 90 percent cached in September. Requesty, an API gateway, reported cache-hit rates for coding agents in April 2026 ranging from 45.5 percent for Kilo Code to 91.9 percent for Claude Code. It computes them as cached tokens over input tokens and does not state its sample size. A public set of coding-agent traces on Hugging Face (netpreme, undated, Claude Code harness on SWE-bench tasks) reports cache rates of at least 97 percent on average and at least 99 percent at the median. Those are the dataset caption’s own floor figures: it says “at least,” and gives neither a range nor exact values. Apart from Kilo Code’s 45.5 percent, all of them show agents reading mostly what they have already sent.

A cache hit is billed far below a fresh input token: 50 times less on DeepSeek’s current Flash model and 10 to 40 times less across Anthropic’s current models, at list prices read October 29, 2026. “Cache hit” is the providers’ name for this re-reading. When the start of a request matches text they processed recently, they reuse the stored work instead of computing it again, and they bill that hit below the normal input price. DeepSeek’s official price page lists its current Flash model (deepseek-flash, V4.1-Flash-1017) at peak list prices of 0.04 yuan per million input tokens on a cache hit and 2 yuan on a miss. That makes a hit 50 times cheaper. Off-peak, both prices halve. The page shows no effective date, and the DeepSWE row’s $0.46 reflects whatever prices applied when it ran. Anthropic’s price page, read the same day, lists cache reads at a tenth of base input for Claude Fable 5 ($1 against $10 per million tokens), a fortieth for Fable 5.1, and a twentieth for Opus 5.5 and Sonnet 5.5.

The re-reading is real. Without it the agent would forget what it was doing. But it is the machine’s memory, not the customer’s demand, and it is priced as such.

Read more

Inside China’s Machine · Tokenomics, Episode 1 of 10

The Answer Price = price per token × tokens per attempt ÷ success rate
The Answer Price is the token cost of each solved task, failed attempts included.

On one public coding-agent leaderboard, DeepSWE v1.1, models work through the same 113 real software tasks under the same harness, mini-swe-agent. The tasks come from 91 repositories in five languages, and the page lists what an average attempt cost (leaderboard updated September 22, 2026). Take two models from one maker, both run at their maximum effort setting. Anthropic lists Claude Sonnet 5 at $2 per million input tokens and $10 per million output tokens. It lists Claude Opus 5 at $5 and $25, which is two and a half times as much on every kind of token. Read the price list and the choice looks obvious.

Then read the column for tasks solved. Opus 5 solved 74 percent of the tasks and Sonnet 5 solved 54 percent. Sonnet 5 also worked much longer at each task. It took 268 steps per attempt against 99, and wrote 214,000 output tokens against 118,000. An average Sonnet 5 attempt cost $26.40 and an average Opus 5 attempt cost $11.84. Divide the bill by the tasks that actually got solved, and an Opus 5 solved task costs $16.00 against $48.89 for Sonnet 5. The model that is 2.5 times dearer per token is about three times cheaper per answer.

That flip shows what the price list leaves out. You do not buy tokens. You buy solved tasks, and the price of a solved task depends on three things, of which the token price is only one. The other two are how much work each attempt takes and how often an attempt succeeds. The last is the one buyers watch least and the one that bites hardest. Our view follows from that. As machine intelligence gets cheap, the money goes to whoever controls the lever that is hardest to move. Today that is success on hard tasks, not price per token.

What an attempt costs, and what a success costs

Start with the simplest case. You send an agent at a task. It reads, thinks, writes and maybe runs some tools, and every token it reads or writes is billed. That is one attempt, and it costs the token price times the tokens it used.

Some attempts succeed and some don’t. You pay for the failures too, in full. So the cost of one success is the cost of an attempt divided by the share of attempts that succeed. If an attempt costs a dollar and half of them work, each working answer costs two dollars.

Put as one equation:

answer price = price per token × tokens per attempt ÷ success rate

We call the result the Answer Price. It is what a buyer spends per successfully solved task, on a stated workload, with a stated test of success, on a stated date; the division gives total spend over total successes across the whole task set, with every failed attempt paid for, so it is a cost per success for the set as a whole, not a price for any single task, which a model may never solve. It sits beside the token price, which is the number on the price page. The token price is real money, but it describes an input. The Answer Price describes what you came for.

Two properties of the formula do most of the work.

First, the levers multiply. A model whose tokens cost half as much but which uses twice as many per attempt has saved you nothing. Agents use a lot. Each step of a long task sends much of what came before back to the model as input, so tokens per attempt grow with the length of the job, not the length of the answer. The opening pair shows how far that goes. Sonnet 5 took 268 steps on an average attempt where Opus 5 took 99, and wrote nearly twice the output. That was enough to make the cheaper-per-token model 2.2 times dearer per attempt before success entered the sum. The leaderboard does not publish input or cached tokens. The page therefore cannot tell you the full token count per attempt, or how much of it was the agent re-reading its own transcript. What it does show is that tokens per attempt is a lever in its own right, and a big one.

Second, success sits in the denominator, and dividing is harsher than multiplying. Raising success from 20 percent to 40 percent halves the Answer Price, exactly as halving the token price would. But success does not stop there. A model that goes from 40 percent to 80 percent halves it again, and one that fails nineteen times in twenty makes every success cost twenty attempts. In the opening pair, success supplies the rest of the flip, a factor of about 1.4 (74 over 54). Opus 5 succeeds 74 percent of the time against 54 percent, which turns its 2.2-times advantage per attempt into 3.1 times per solved task. A spread that looks modest in percentage points becomes large once it sits under the line, and on hard work that is why this lever so often decides the bill.

Three levers, three different speeds

The three levers do not move together, and that is where the economics starts.

Token price falls fast and in public. Anthropic launched Claude Opus 4 on May 22, 2025 at $15 per million input tokens and $75 per million output tokens. Claude Opus 4.5, launched November 24, 2025, listed at $5 and $25, a cut of two thirds on both. Opus 5 kept that price at its launch on July 24, 2026, and Anthropic’s pricing page now lists Opus 5.5, launched September 22, 2026, at $4 and $20. Per-token cuts overstate per-text cuts, though. Anthropic’s documentation says the tokenizer in Claude 4.7 and later models produces about 30 percent more tokens for the same text, and its launch posts give a range of 1.0 to 1.35 times; that applies to those later models, not to the 2025 cut from Opus 4 to Opus 4.5. Token price is also the lever anyone can pull. Once a model’s weights are published, any cloud with the chips can serve the same model, and the price for those tokens is set by competition among hosts.

Sellers have even started pricing tokens by the hour. Since 16:00 UTC on August 16, 2026, DeepSeek has charged peak and off-peak rates for V4-Pro, with off-peak rates 50 percent below peak (DeepSeek notice of August 13, 2026). DeepSeek publishes the peak hours in Beijing time: 9:00 to 12:00 and 14:00 to 18:00 on weekdays, excluding Chinese public holidays. So the same token costs half as much at night, over lunch, at weekends and on public holidays as it does in the Beijing working day. The hourly price shows that the token price is a lever sellers adjust freely, which tells you it is not where their power sits.

Tokens per attempt is pulled by both sides. The model maker sets how much the model thinks and how many steps it takes before it stops. The buyer, or the application built on the model, sets how much context gets fed in, what gets cached, how many tools run and when to stop. Good engineering here cuts the bill without touching the model, and it is the part of the stack where an application company can save its customers money the model maker never sees.

Success rate is the hard one. On easy tasks every serious model succeeds, and the lever is flat. On hard tasks, the kind that agents are being sold for, success differs widely between models. On DeepSWE v1.1, the 21 models the page shows by default range from 74 percent, a level three models share (Opus 5, gpt-6-astra and gemini-3.8-flash), down to 36 percent. Success does not fall to competition the way token prices do. You cannot buy it from a second host. Effort settings and context management can move the same model’s success by some points, but the large differences between models come mostly from the model itself. The same maker at the same price shows where it moves. Claude Opus 4.8 [max] lists at Opus 5’s $5 and $25. It solves 59 percent of the DeepSWE tasks at $13.22 an attempt, or $22.41 per solved task. Opus 5, at $16.00, is 29 percent cheaper per solved task, mostly because it succeeds more often. That success came with a new model, not a new price.

Who gets paid

Put the three speeds side by side and the question of who gets paid has a shape.

A seller whose only edge is the token price is selling the abundant lever. Anyone serving the same weights can match the price, and the buyer can switch with a configuration change. Expect that margin to be competed thin. Cheap tokens are not a bad business because they are cheap. They are a hard business because they are interchangeable.

A seller who moves success on hard tasks is selling the scarce lever. Look at what the formula does for them. If their model succeeds twice as often, they can charge close to twice as much per token and still give the buyer the same Answer Price, or a better one. The higher token price is not a premium the buyer tolerates. It is a share of the saving that the model created and the seller keeps. That is how Opus 5 can be 2.5 times dearer per token than Sonnet 5 and still win the task. In that pair the shorter attempts did more of the work than the higher success rate: they made Opus 5 about 2.2 times cheaper per attempt despite its dearer tokens, and the higher success rate, 74 percent against 54, added about 1.4 times. Both are properties of the model, and neither shows up on a price list.

So our view is that the margin sits with whoever controls the lever buyers cannot route around. Today, on the work agents are being bought for, that is success. How much of the buyer’s saving the seller can keep, through a higher token price, is set by how much better the model is than the next best one, not by the price list. A buyer who shops on token price is comparing the one number sellers are happiest to cut.

The view has limits. It is a view about hard tasks. On easy, high-volume work, the alternative holds: success is near the ceiling for everyone, and the cheapest token is close to the cheapest answer. Vals.ai’s SWE-bench Verified board (mini-swe-agent, updated September 1, 2026) has been archived because, in Vals.ai’s words, performance on the benchmark has saturated. There Claude Opus 5 solved 97.0 percent at $1.29 per test and DeepSeek V4 Pro 96.4 percent at $0.10. Per solved test that comes to $1.33 for Opus 5 against about $0.10 for DeepSeek V4 Pro. That gap, about 13 times, is the same as the gap in price per test, because both models solve almost every test, so the cost gap passes through almost untouched. Vals.ai is a different operator with its own cost method, so its dollar figures do not line up with DeepSWE’s.

That easy-work market is large, and it will be priced like a commodity. The view also has a clock in it, though not a dated one. Success is scarce until the second-best model catches up on the tasks buyers care about. When it does, success flattens like token price did, and the margin moves to whatever is scarce next: capacity at the hours buyers need it, the engineering that cuts tokens per attempt, or the work of checking the answer.

China tests this sharply, because Chinese labs have pushed the token-price lever further than anyone. On October 29, 2026, DeepSeek listed V4-Pro at $0.66 per million cache-miss input tokens off-peak and $1.32 at peak, and at $1.98 and $3.96 for output. Anthropic listed Claude Opus 5, the model in the DeepSWE rows compared here, at $5 and $25 the same day. That puts DeepSeek 3.8 to 7.6 times cheaper on input and 6.3 to 12.6 times cheaper on output, depending on the hour. Where DeepSeek’s models match on success, as on the saturated SWE-bench Verified, that discount reaches the buyer almost one for one, and nothing in this framework softens it. Where they trail, the success gap shrinks the discount, but it takes a large gap to erase a large discount. On DeepSWE, DeepSeek-V4-Pro [max] solves 63 percent (±6) against Opus 5’s 74 and still comes to $2.65 per solved task against $16.00. DeepSeek-V4-Pro costs about 7 times less than Opus 5 per attempt ($1.67 against $11.84) and about 6 times less per solved task ($2.65 against $16.00), so its lower success rate gives back little of the discount. The same caveat about DeepSWE’s undisclosed cost method applies here, including which DeepSeek rate it charged. The question for any cheap route, in Beijing or anywhere else, is never how cheap the token is. It is whether the success rate holds on your work.

Back to the leaderboard

Go back to the two rows from the opening. On the price page, Opus 5 is 2.5 times dearer than Sonnet 5. On the tasks solved, it is about three times cheaper. Nothing on the price page was wrong. It was simply answering a question the buyer had not asked.

Ask what a solved task costs, not what a token costs. Then ask which of the three levers is moving it and who controls that lever. The token price will keep falling. Whether that makes anyone richer depends on the other two.

Sources

1. Datacurve, DeepSWE v1.1 leaderboard (page updated 22 September 2026), read 29 October 2026, https://deepswe.datacurve.ai/

2. Datacurve, DeepSWE v1.1 blog post, read 29 October 2026, https://deepswe.datacurve.ai/blog/deepswe-v1-1

3. arXiv, DeepSWE paper abstract (2607.07946), read 29 October 2026, https://arxiv.org/abs/2607.07946

4. GitHub, datacurve-ai/deep-swe Issue #21 “DeepSWE Audit: Pricing, provider routing, and effort-tuning concerns (solve-rate findings retracted)”, opened 1 June 2026, https://github.com/datacurve-ai/deep-swe/issues/21

5. Anthropic, Claude API pricing page, read 29 October 2026, https://platform.claude.com/docs/en/about-claude/pricing

6. Anthropic, Claude 4 launch post, 22 May 2025, https://www.anthropic.com/news/claude-4

7. Anthropic, Claude Opus 4.5 launch post, 24 November 2025, https://www.anthropic.com/news/claude-opus-4-5

8. Anthropic, Claude Opus 4.7 launch post, read 29 October 2026, https://www.anthropic.com/news/claude-opus-4-7

9. Anthropic, Claude Opus 5 launch post, 24 July 2026, https://www.anthropic.com/news/claude-opus-5

10. Anthropic, Claude Sonnet 5 launch post, read 29 October 2026, https://www.anthropic.com/news/claude-sonnet-5

11. Anthropic, Claude Opus 5.5 launch post, 22 September 2026, https://www.anthropic.com/claude-opus-5-5

12. DeepSeek, API pricing page, read 29 October 2026, https://api-docs.deepseek.com/quick%5Fstart/pricing/

13. DeepSeek, notice “DeepSeek-V4-Pro GA Release”, 13 August 2026, https://api-docs.deepseek.com/news/news260813

14. Vals.ai, SWE-bench Verified benchmark (archived board, updated 1 September 2026), read 29 October 2026, https://www.vals.ai/benchmarks/swebench

Inside China’s Machine · Tokenomics, Episode 1 of 10

Disclaimer: This article is for informational and educational purposes only. It is not investment, financial, legal or tax advice, and it is not a recommendation or solicitation to buy or sell any security. Do your own research, and consult a licensed professional before making investment decisions.

Estun’s half-year report for 2026 makes two statements that sit awkwardly together. It says Estun ranked first by shipments among all brands in China’s industrial-robot market, for the sixth quarter running. And it shows Estun’s revenue in China falling 8.82%, to RMB 1,641.0m, in a market whose unit sales grew 15.8%.

My view is that winning the volume in China’s industrial robots has not yet turned into money for China’s makers. This is a judgment about the industry, not a verdict on one company, and not a forecast. Its basis is in three numbers. Domestic brands held 58% of the 189,000 industrial robots sold in China in the half-year, about 110,000 units. Estun, the leader among them, reported net profit to the parent excluding non-recurring items equal to 2.72 yuan per 100 yuan of sales. And Fanuc, the Japanese maker, reported that its robot division’s sales to China in the April to June quarter rose 21.1% in yen on the same quarter a year earlier.

Share is the number everyone quotes, and it is a count of robots. What a maker keeps from each one decides who can fund the next generation of products and who cannot.

More robots in the market, less revenue at home

Start with Estun’s revenue. The half-year total was RMB 2,577.5m, up 1.14% from RMB 2,548.6m. Underneath that flat line, two things moved in opposite directions. Domestic revenue fell RMB 158.8m. Overseas revenue rose RMB 187.7m, up 25.06% to RMB 936.5m. The two add up to the RMB 28.9m by which the total rose.

So the market grew by about a sixth in units, domestic brands took 58% of it, and the domestic leader’s revenue at home shrank. Estun’s revenue covers the whole company, not robots alone, and its report gives no robot-only line or unit growth of its own. And part of the fall was chosen: Estun says it reduced low-gross-margin business in an orderly way, and the transfer of a controlled subsidiary also lowered domestic revenue. Its first place by shipments, as counted by MIR, the market-data source both Estun and Inovance cite, is not in doubt.

When a leader’s best use of its position is to sell less of some business at home, the unit count has stopped telling you what the business earns.

A foreign maker in the same market

Fanuc is the contrast, with a different calendar and a different currency. In the three months to 30 June 2026, its robot division’s sales to China rose 21.1% in yen on the same three months of 2025, to about ¥26.6bn from about ¥21.9bn. China was 27.7% of the division’s sales. The yen was weaker, ¥159.49 to the dollar against ¥144.59 a year before, so converted at those rates the rise is about 9.8% in dollars. The rate depends on the currency; the direction does not. Fanuc’s report puts steady growth in the Americas and China behind the higher sales.

Figure 1 sets the pieces side by side. They are not one measure, and the periods differ, so read it for direction, not as a league table.

Sources: Estun half-year report 2026 (domestic revenue RMB 1,641.0m, −8.82%; total revenue +1.14%; gross profit RMB 819.9m against RMB 704.4m); Inovance half-year report 2026, citing 睿工业 (MIR) (189,000 units, +15.8%); Fanuc results for the three months to 30 June 2026 (China robot sales +21.1% in yen) and its exchange rates of ¥144.59 and ¥159.49 to the dollar. Dollar growth of +9.8% computed from the figures shown.

Inovance, the other large domestic name, cannot be added to this table. It holds 9.2% of China’s industrial-robot shipments, third overall and second among domestic makers, but its report folds robots into one “emerging industry” line with energy storage: revenue RMB 1,467.1m, up 96.29%, at a gross margin of 28.38%, down 5.40 points as low-margin storage grew. Its robot margin is not published. It does describe the industry as one where competition is intensifying and profits are under pressure.

What 31.81% became

The strongest evidence against my view is Estun’s gross margin, which rose 4.17 points to 31.81%. At home it rose 7.04 points to 31.49%, because domestic cost of sales fell 17.32% while domestic revenue fell 8.82%. Gross profit was RMB 819.9m against RMB 704.4m, up 16.4% on revenue up 1.1%, and domestic gross profit rose by about the same proportion, to roughly RMB 517m. Estun booked less revenue at home and made more gross profit from it. If that margin holds while domestic revenue returns to growth, the view is wrong.

But gross profit is the start of the bill. Of every 100 yuan of sales, Estun keeps 31.8 as gross profit. Selling costs take 7.8, administration 8.0, R&D 8.5, finance costs 2.9 and taxes and surcharges 0.5. Together those five lines come to RMB 715.0m, 87.2% of gross profit, and leave about 4.1 yuan before impairments, other income, the fair-value gain and income tax.

The profit lines show where that ends up. Net profit to the parent was RMB 161.3m, up 2,314%. But RMB 99.5m of the company’s pre-tax profit, 50.69% of it, was a fair-value gain that the report says is not sustainable. It arose when Estun’s stake in Nanjing Gongyi was swapped for shares of listed Nanjing Chemical Fibre, not from selling robots. Net profit to the parent excluding non-recurring items was RMB 70.2m, up 498.64%: 2.72 yuan per 100 of sales, 8.6% of gross profit. It is a real improvement. It is also a small one.

What the share headline leaves out

“Domestic brands hold 58%” is a share of units. It does not say what the makers earn, whether the growth is theirs, or whether it was paid for in margin given up elsewhere. In this half-year it sits beside a leader whose home revenue fell, a foreign rival whose China sales rose, and a second domestic name whose robot margin is hidden inside an energy-storage line.

Four things in the next half-year’s reports will say whether the view holds. Whether Estun’s domestic revenue moves with the market’s units, or keeps moving the other way. Whether its net profit to the parent excluding non-recurring items grows past 2.72% of revenue. Whether Fanuc’s China robot sales keep rising in both yen and dollars. And whether Inovance, or anyone else, publishes a robot margin by itself.

Share counts robots. Survival counts what is left after the bill. China’s makers have the first; the evidence for the second is still thin.

Sources

1. Nanjing Estun Automation Co., Ltd., 2026 half-year report (cninfo), read 2026-10-22: http://static.cninfo.com.cn/finalpage/2026-08-22/1225490661.PDF. Revenue by region, cost of sales and gross margin, profit excluding non-recurring items, the fair-value gain, the company’s account of the fall in domestic revenue, and its first place by shipments (citing MIR DATA BANK).

2. Shenzhen Inovance Technology Co., Ltd., 2026 half-year report (cninfo), read 2026-10-22: http://static.cninfo.com.cn/finalpage/2026-08-29/1225524610.PDF. China industrial-robot units and domestic share (citing 睿工业, MIR), its own share, and the “emerging industry” segment.

3. FANUC Corporation, Financial Results for the Three Months Ended June 30, 2026 (presentation of 31 July 2026), read 2026-10-22: https://www.fanuc.co.jp/en/ir/announce/pdf/2026/reference202606_e.pdf. Robot division sales to China, the division’s regional share and the exchange rates.

Inside China’s Machine · Frontier, Episode 3

Disclaimer: This article is for informational and educational purposes only. It is not investment, financial, legal or tax advice, and it is not a recommendation or solicitation to buy or sell any security. Do your own research, and consult a licensed professional before making investment decisions.

On 20 August 2026, Alibaba reported its June quarter and, for the first time, split its AI business in two.

The first box, AI Cloud and Compute Services (compute, for short), rents out servers and holds T-Head, Alibaba’s own chip designer. The second, AI Labs and Applications (the labs), holds the labs behind Qwen, Alibaba’s family of models, along with the Qwen consumer app.

Compute earned RMB 5,628m, measured in adjusted EBITA, Alibaba’s own operating-profit figure. The labs lost RMB 13,861m on sales of RMB 3,338m, about four times what they sold, and Alibaba puts part of that loss down to the running cost of its consumer Qwen app. The loss is about two and a half times the profit beside it. A year earlier the two together lost RMB 805m.

Who pays for China’s AI models? It depends on the kind of company. Inside a large group, the group’s other businesses pay: at Alibaba, the shopping business, whose profit of RMB 39.7bn is nearly three times the labs’ loss. A model maker outside the group pays with its investors’ money. MiniMax is one: Alibaba holds an indirect stake of about 11% in it, and MiniMax bought about US$75.6m of cloud services from Alibaba Cloud in the first half of this year. Its research spending in that half was two and a half times its revenue, and it paid for its cash needs mainly from financing. In both cases spending on models ran ahead of what the models sold, and someone other than the models’ buyers covered the gap.

The one part of Alibaba’s AI business that made money is the part that owns the machines. The models are the cost.

Who is left with the profit from China’s AI models decides who can afford to build a frontier model and who has to be carried by someone else.

Is it just a consumer app?

Inference is running a trained model to answer requests, and the Qwen app is a consumer product reported inside the labs. Alibaba says the labs’ loss came mainly from increased investment in AI capabilities and “higher inference cost related to Qwen app”. So the RMB 13.9bn is not simply the cost of serving paying customers.

That is fair, and it does not change the answer. The question is who carries the spending, whatever it buys. The labs’ loss grew from RMB 3,224m to RMB 13,861m, about RMB 10.6bn worse, while their sales grew RMB 456m. Research and app costs are mixed inside that figure, and the AI side did not earn either back.

The shopping business carries it

Set against compute’s profit, the labs’ loss falls from RMB 13.9bn to RMB 8.2bn. E-commerce profit exceeds what remains, and the whole labs’ loss equals about 35% of it. This is an offset between segments in an accounting table, not a transfer of cash.

Compute’s margin is 11.6% on revenue of RMB 48.4bn, up 45%, with profit up 133%. On each added 100 yuan of revenue compute kept about 21 yuan of profit, which is the change in its adjusted EBITA divided by the change in its revenue.

Compute’s bar grows by RMB 3.2bn and the labs’ falls by RMB 10.6bn. E-commerce, the tallest bar, barely moves. It was large and steady enough this quarter to absorb the labs’ loss.

A free app with a quota

The shopping business is not an accident of the accounts. Alibaba says the Qwen app is built into Taobao, Tmall and Taobao Instant Commerce to extend the reach of its e-commerce business, that 250 million users have had their first AI-driven shopping experience through the app’s agentic features, and that the link makes the AI and e-commerce businesses feed each other, which it calls a flywheel.

For someone building on the app, or backing a lab whose consumer app runs at a loss, the stake is whether “free” is a durable offer. CLS reported on 10 August that on 7 August the app launched an office assistant and other features free to all users, with extra usage for professional users to be paid for. CLS also reported that the app’s App Store page in August showed the office assistant in three paid tiers: RMB 200 a year, RMB 568 a year (RMB 49 a month) and RMB 1,499 a year (RMB 128 a month). Alibaba had not responded when that report was published. DoNews reported on 12 August that the tiers give 2, 5 and 20 times the usage quota on some features, and that basic office functions stay free with limited quota. Alibaba’s results came ten days after the CLS report and put part of the labs’ loss on the app’s inference cost.

Alibaba is not alone in this. DoNews reports that the rival app Doubao launched three paid tiers in June at RMB 688 to 5,088 a year, and that Baidu’s Wenxin tried charging in November 2023, ended the membership on 1 April 2025 and made the product free.

What the reports show is that the basics stay free while heavier use is rationed by quota and sold in tiers. They do not say which features cost most to run. Alibaba’s results carry no count of paying users, so how much the tiers recover is unknown, and Wenxin’s reversal shows that a paid tier can be abandoned.

The model company with a cloud bill

MiniMax shows the case without that tall bar. In the six months to 30 June it had revenue of US$116.6m and a gross profit of US$20.8m, a margin of 17.9% against 12.1% a year earlier, which it puts down mainly to improving infrastructure efficiency. Serving customers is getting better.

Training is the bill. Research and development cost US$296.9m, about fourteen times its gross profit, mainly from higher cloud services expenses for training. The loss for the period was US$358.0m.

So MiniMax pays for cloud services to train its models, and financing pays MiniMax. It says it funded its cash requirements in the half mainly from financing activities, and held US$1,322.8m at 30 June. Alibaba is both an indirect holder of about 11% of MiniMax and a supplier to it. Financing is supporting its training expansion, though MiniMax’s disclosures do not trace each dollar.

Owning the compute has its own bill. Alibaba’s capital expenditure was RMB 67.7bn in the quarter, 75% more than a year earlier, and its free cash flow was an outflow of RMB 44.7bn. These are group-wide figures. Alibaba ties the cash outflow mainly to higher cloud infrastructure spending, says the capital expenditure reflects continued investment in AI infrastructure to meet strong and growing customer demand, and lists higher prices for chip components among the reasons. For scale, group capital expenditure is 1.4 times compute’s quarterly revenue. The profit lands with the owner of the machines, and so does a large capital bill.

Whose bill comes round

If you buy or build on a Chinese model, the seller’s staying power depends on who pays for its compute and its losses. A lab inside a group with a large, profitable customer business, as at Alibaba, is carried by that business. A model company outside the group, as MiniMax is, is carried by the investors who financed it, and Alibaba, its cloud supplier, holds about 11% of it indirectly.

Two companies do not make an industry pattern. They show two ways the spending can be funded, and at both, in this period, spending on models ran well ahead of what the models sold.

A group lab depends on the size and steadiness of the business beside it. A model maker outside a group depends on its financing. The accounts show the bill and who has paid it so far: a shopping business in one case, financiers in the other.

Sources

1. Alibaba Group, announcement of results for the June quarter 2026, 20 August 2026 (HKEXnews), read 2026-10-21: https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0820/2026082000745.pdf. Segment revenue and adjusted EBITA, the definition of adjusted EBITA, the Qwen app, capital expenditure and cash flow.

2. Alibaba Group, Form 6-K, Exhibit 99.1, the same results, 20 August 2026, read 2026-10-21 through a web reader: https://www.sec.gov/Archives/edgar/data/0001577552/000110465926099220/tm2623667d1_ex99-1.htm

3. MiniMax Group Inc., interim results announcement for the six months ended 30 June 2026 (HKEXnews announcement, as hosted by the company), read 2026-10-21: https://file.cdn.minimax.io/public/586c7995-cab1-41b5-9178-c06685b32c88.pdf. Revenue, gross profit, research and development expenses, loss for the period, cash requirements and cash balance.

4. MiniMax Group Inc., announcement of 26 August 2026 on continuing connected transactions (HKEXnews), read 2026-10-21: https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0826/2026082601949.pdf. Alibaba’s indirect interest in MiniMax and MiniMax’s purchases of Alibaba Cloud services.

5. IBM Research, “What is AI inferencing?” (the definition of inference), read 2026-10-21: https://research.ibm.com/blog/AI-inference-explained

6. 财联社 (CLS), “千问办公助理收费 会员套餐分旗舰、精英、高级三档”, 10 August 2026, read 2026-10-21: https://www.cls.cn/detail/2450364

7. DoNews, “阿里千问App上线付费会员服务，最高1499元/年”, 12 August 2026, read 2026-10-21: https://www.donews.com/news/detail/4/6667389.html

Inside China’s Machine · Frontier, Episode 2

Disclaimer: This article is for informational and educational purposes only. It is not investment, financial, legal or tax advice, and it is not a recommendation or solicitation to buy or sell any security. Do your own research, and consult a licensed professional before making investment decisions.

On 9 August 2026, Qwen put a model on Hugging Face, the site where labs post models for download, with no licence file. Its model card says: “For the first time, Qwen3.8 brings a Qwen-Max-class model to open release.” Three days later, on 12 August, Qwen added the licence. Any company that sells access to the model, usually through an API, and whose group revenue passes US$50 million over any twelve months, must first get a separate licence from Qwen. The count is the revenue of the company and its affiliates, not what the model earns.

Moonshot’s Kimi K3 draws a line of the same kind at US$20 million. Z.AI’s GLM-5.3 draws it at US$10 billion, 500 times Moonshot’s.

Why would Chinese labs let anyone download their models and still keep a say over who sells them? Was giving the weights away a strategy the labs chose, or something the market forced on them? The question matters to a company that wants to sell these models: how much say a lab keeps over who sells them, and which sellers have to ask.

The files show what each lab kept: at the top of their range, Moonshot, Qwen and Z.AI publish the weights and keep the decision over whether a seller above a revenue line may sell, and the terms differ by lab and by release. Each licence draws a line, and the line shows which sellers the lab wants to hear from first.

Picture the one who has to ask: a cloud company of Kingsoft Cloud’s size (US$1.37 billion of revenue in 2025) that wants to sell access to Kimi K3. It is over Moonshot’s line and Qwen’s, and under Z.AI’s. And Moonshot’s clause and Z.AI’s turn out to be the same sentence with two edits.

What the cloud must do first

Moonshot’s clause targets a business it calls Model as a Service: giving a third party access to a model’s inference or fine-tuning, say through an API, with meaningful control over inputs, parameters or training data. Relaying requests to a model someone else hosts is outside it. A cloud that sells customers the model through an API is inside.

Now take a retailer that does not exist. Its engineers download Kimi K3 and build it into a feature its customers see, an advice box on the shop pages. Elsewhere in the same group, a separate company sells an inference API on other labs’ models. It has nothing to do with K3.

Read the clause against that group. It applies when the licensee or any of its affiliates runs a Model as a Service business, and it does not say that business has to serve Kimi K3. The sister company’s API business is such a business, and its revenue counts toward the group’s. Once the group’s revenue passes US$20 million over any twelve months, the retailer must enter into a separate agreement with Moonshot before using K3 for any commercial purpose.

The retailer might reach for the exemption for internal use, which covers use that gives no third party the software, its outputs or its capabilities. An advice box that answers customers hands them K3’s outputs. That is not internal use, and the exemption does not cover it. So a company that sells no access to anything stands in front of Moonshot because of what its sister company sells.

GLM-5.3 has the same affiliate trigger and the same revenue base, and its file has no internal-use exemption.

Back to the cloud company. Above US$20 million of group revenue, the licence requires a separate agreement with Moonshot before any commercial use. Any twelve months can qualify, so a year of Kingsoft Cloud’s US$1.37 billion counts. That is 68 times Moonshot’s line and 27 times Qwen’s, so the same cloud also needs a separate licence from Qwen. Under GLM-5.3 it is below the line, and Z.AI’s clause does not reach it.

Same sentence, two edits

Put Z.AI’s clause for the same business beside Moonshot’s. Of Kimi K3’s 132 words, 122 appear in GLM-5.3’s, in the same order. Two places differ.

The first is the line: US$20 million against US$10 billion, 500 times apart. The second is what a seller above the line must do. Moonshot’s seller must “enter into a separate agreement with Moonshot AI”. Z.AI’s must “pass Z.AI’s security review”, and Z.AI sets the scope and method of that review. Qwen’s line is 2.5 times Moonshot’s, and its remedy is a separate licence from Qwen.

So the widest gap among the three licences sits in a clause whose wording is otherwise shared.

Moonshot’s line reaches every company on the chart, and Z.AI’s reaches only the three largest.

What each lab keeps

Each lab leaves a different way past its line.

Besides internal use, Kimi K3’s licence exempts any use accessed through Moonshot’s official products or certified inference partners. For a seller above the line, the way past the agreement runs through Moonshot’s own channel.

Qwen’s model card sends anyone who wants managed inference to the official Qwen API service from Qwen Cloud. Alibaba’s own results describe a platform, Model Studio, that offers Qwen models to users from individual developers to large enterprises. A seller above Qwen’s line asks permission of a lab whose own cloud sells the same service.

Qwen’s line also reaches products that are not APIs. A company above US$50 million of group revenue that runs an independent coding or office product of the kind the licence calls an AI Work Assistant needs a separate Qwen licence before commercial use, even if it hosts the model for nobody. Internal use is exempt. The category is an independent AI product made mainly for coding or office productivity, with Qoder and QwenWork as the licence’s examples. Caixin reports that Alibaba created QwenWork. The category excludes an assistant made mainly for another domain, and the licence’s example of that is Taobao’s shopping assistant. Taobao sits in Alibaba’s China e-commerce group.

GLM-5.3’s licence file names no exemption for Z.AI’s own products or partners. For a seller above Z.AI’s line, the file offers the review and nothing else.

Strategy, or forced?

The strategy case says a lab gives the weights away because it earns elsewhere. Alibaba’s chairman put it most plainly. At the World Government Summit 2026, as Alibaba’s site reported on 6 February, Joe Tsai said: “Our Qwen model is open source, but we also run our cloud business”. Alibaba earns, he said, because people train, develop and run inference for their models on its infrastructure. Kyle Chan of the Brookings Institution says the same of Moonshot in Scientific American (23 July 2026): it can keep selling access through its API and subscription products even after other companies host K3.

The forced case says the lab had no choice, and Moonshot supplies its sharpest evidence. Scientific American reported that three days after Moonshot unveiled Kimi K3 on 17 July, it stopped accepting new subscriptions because demand had overwhelmed its computing capacity, and that Moonshot said it planned to release the full weights by 27 July. The first commit is dated 27 July. Chan gives the general mechanism: US export controls do not fully explain Chinese developers’ embrace of open weights, but limited compute makes the strategy more attractive, because releasing weights puts to work compute that others have built.

A Takshashila Institution discussion paper dated 7 August 2026 answers with Alibaba. It has the largest cloud and GPU footprint in China and is the country’s most prolific open-weight releaser, so the paper argues that inference constraints are at most a secondary factor.

The licences can tell the two readings apart on one point. If the weights feed the lab’s own sales, the licences should guard those sales. If pressure put the weights out, the pressure still does not write the terms that ship with them: neither the capacity shortage nor Chan’s mechanism says what a licence contains. The licences fit the first reading: Kimi K3’s exemption for Moonshot’s own products and Qwen’s pointer to Qwen Cloud both send a seller to the lab’s own channel. And the forced case’s best example ships with the lowest of the three lines. Whatever pressure put K3’s weights out, the licence that came with them draws a line at US$20 million.

So I read it this way. The timing of K3’s release may well have been pushed by Moonshot’s short capacity. The terms are the labs’ own, written release by release. Kimi K2.6 shipped with no hosting clause and K3 with the gate; GLM-5.2 shipped as plain MIT and GLM-5.3 with the gate.

What the three gated labs kept is the same in each. The weights go out, which in MIT Technology Review’s account wins developers’ goodwill and free advertising. The decision over selling access to them stays with the lab. Moonshot’s and Z.AI’s gates fall on one business, selling access to a model, wherever in the licensee’s group it sits, and Qwen’s adds independent coding and office products. Selling access is also the business a lab’s own API or cloud runs, as Tsai describes for Alibaba and Chan for Moonshot. Z.AI’s gate falls on the same defined business, and its file names no exemption for its own products. For these labs, then, open weights are a strategy, and the gate protects what the lab itself sells.

What the twelve licences show

Figure 2 sets the repositories side by side. Of the six labs in it, four (Moonshot, Z.AI, Qwen and MiniMax) have a licence that draws a revenue line, and two, DeepSeek and Xiaomi, have none. The lines that exist sit 500 times apart.

The answer is a reading of the four. DeepSeek and Xiaomi make the opposite choice: they ship top-tier weights with no line at all. DeepSeek published V4-Pro-0813 on 13 August under the plain MIT licence, and its own benchmark table sets the model against Kimi K3 and GLM-5.2. Xiaomi’s MiMo-V2.6-Pro-RL, published on 21 September, is MIT too. For DeepSeek, its founder Liang Wenfeng gave a reason in a July 2024 interview with Waves, a 36Kr outlet, that is neither revenue nor necessity: open source “is more of a cultural behavior than a commercial one” (ChinaTalk’s English translation).

What it changes

For a company that wants to sell access to these models, “open” no longer answers the question. The same download means three different things to a seller the size of that cloud: an agreement to make with Moonshot, a licence to get from Qwen, and nothing yet to ask of Z.AI.

A business growing toward a line is growing toward a conversation with the lab.

The weights are open to everyone. The right to sell them at scale is still the lab’s to give.

Sources

1. Hugging Face repository licence files, read 2026-10-20, at https://huggingface.co/REPOSITORY/raw/main/LICENSE: moonshotai/Kimi-K2.6, moonshotai/Kimi-K3, zai-org/GLM-5.2, zai-org/GLM-5.3, zai-org/GLM-5.3-Flash, Qwen/Qwen3.8-2.4T-A95B, Qwen/Qwen3.8-27B, Qwen/Qwen3.8-Flash-Next, MiniMaxAI/MiniMax-M3, deepseek-ai/DeepSeek-V4-Pro-0813, deepseek-ai/DeepSeek-V4.1-Flash-1017.

2. Hugging Face repository commit histories, read 2026-10-20, at https://huggingface.co/api/models/REPOSITORY/commits/main: the eleven repositories in line 1 and XiaomiMiMo/MiMo-V2.6-Pro-RL.

3. Hugging Face repository records, read 2026-10-20, at https://huggingface.co/api/models/REPOSITORY: the same twelve repositories.

4. Hugging Face, moonshotai/Kimi-K3, LICENSE at its first commit, read 2026-10-20: https://huggingface.co/moonshotai/Kimi-K3/raw/c5d1dd4c428bd1ce8b88c5044f3b6ccde9e3b721/LICENSE

5. Hugging Face, zai-org/GLM-5.3, LICENSE at its first commit, read 2026-10-20: https://huggingface.co/zai-org/GLM-5.3/raw/7cda81930d6e4cef42f48555de830aa32ecdde28/LICENSE

6. Hugging Face, Qwen/Qwen3.8-2.4T-A95B, LICENSE path at its first commit (no file at that commit), read 2026-10-20: https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/f84a3dbf08869f3a91d9d514c573cdecbd7e03bd/LICENSE

7. Hugging Face, Qwen/Qwen3.8-2.4T-A95B, model card (README.md), read 2026-10-20: https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/README.md

8. Hugging Face, XiaomiMiMo/MiMo-V2.6-Pro-RL, model card (README.md), read 2026-10-20: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL/raw/main/README.md

9. Hugging Face, deepseek-ai/DeepSeek-V4-Pro-0813, model card (README.md), read 2026-10-20: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813/raw/main/README.md

10. Alibaba Group, “Alibaba Group Announces March Quarter 2026 and Fiscal Year 2026 Results” (company press release as distributed on Seeking Alpha), read 2026-10-20: https://seekingalpha.com/pr/20512712-alibaba-group-announces-march-quarter-2026-and-fiscal-year-2026-results

11. Tencent Holdings, 2025 results announcement (PDF), read 2026-10-20: https://static.www.tencent.com/uploads/2026/03/18/bd32d8ee320b72f0d72d545fd2d65851.pdf

12. Baidu, “Baidu Announces Fourth Quarter and Fiscal Year 2025 Results”, as distributed on Seeking Alpha, read 2026-10-20: https://seekingalpha.com/pr/20414723-baidu-announces-fourth-quarter-and-fiscal-year-2025-results

13. Kingsoft Cloud, “Kingsoft Cloud Announces Unaudited Fourth Quarter and Fiscal Year 2025 Financial Results”, as distributed on Seeking Alpha, read 2026-10-20: https://seekingalpha.com/pr/20450479-kingsoft-cloud-announces-unaudited-fourth-quarter-and-fiscal-year-2025-financial-results

14. Alibaba Group, “Joe Tsai on the Future of Open-Source AI: Why Full-Stack Companies Will Excel”, 2026-02-06, read 2026-10-21: https://www.alibabagroup.com/document-1956438710056648704

15. Waves (36Kr), interview with Liang Wenfeng, July 2024, in ChinaTalk’s English translation, “DeepSeek: The Quiet Giant Leading China’s AI Race”, 2024-11-27, read 2026-10-21: https://www.chinatalk.media/p/deepseek-ceo-interview-with-chinas

16. Scientific American, “China’s Kimi K3 and the rise of open-weight AI models”, 2026-07-23, read 2026-10-21: https://www.scientificamerican.com/article/china-kimi-k3-and-the-rise-of-open-weight-ai-models/

17. Takshashila Institution, “Why Is China Giving Away Its AI Models?”, discussion paper, 2026-08-07, read 2026-10-21: https://takshashila.org.in/content/publications/20260807-why-is-china-giving-away-its-ai-models.html

18. MIT Technology Review, article on China’s open-source models, 2026-04-21, read 2026-10-21: https://www.technologyreview.com/2026/04/21/1135658/china-open-source-models-ai-artificial-intelligence/

Inside China’s Machine · ICM Frontier, Episode 1

Disclaimer: This article is for informational and educational purposes only. It is not investment, financial, legal or tax advice, and it is not a recommendation or solicitation to buy or sell any security. Do your own research, and consult a licensed professional before making investment decisions.

This is the first installment of a quarterly series mapping China’s AI model makers.

Xiaomi, which sells phones and cars, puts nearly 30 percent of a ¥18.2 billion half-year research budget into AI, by its own account. That is up to about ¥5.5 billion ($0.8 billion) by our estimate, covering all its AI work, language model included, so the comparison that follows is not like for like. It is more than twice the ¥2.13 billion ($316 million) that Zhipu, a lab listed in Hong Kong, spent on all of its research in the same six months.

Their language models sit one point apart on the Artificial Analysis Intelligence Index, an independent benchmark. Xiaomi’s MiMo-V2.6-Pro scores 46 and Zhipu’s GLM-5.3 scores 45. Running the full index costs a buyer $207 on MiMo and $2,503 on GLM. That is the buyer’s bill. What the models cost their makers is a different number, and this piece asks who pays it. For Xiaomi the answer is its phone and car business, and the piece sets that profit against the AI spend.

A model pays for itself when its revenue exceeds what it costs to build and run, research included. Zhipu and MiniMax, the two listed labs that sell models as their business, both report a positive gross margin, which is revenue less the direct cost of serving customers, so the gap lies outside the cost of revenue they report.

Zhipu’s accounts show the loss. It lost ¥2.07 billion ($308 million) on ¥954 million ($142 million) of revenue in the half, and raised or agreed to raise about ¥61 billion ($9 billion) between July and mid-September, by our sum of company filings and reports, some of it not yet shown as closed.

Our answer is that no account we read shows a model paying for itself. At Zhipu and MiniMax, customers’ gross profit covers about one-eighth and one-fourteenth of research. Where a parent can be measured, its profit covers its AI spend about 2 to 7 times. The labs fill the rest from investors, at share prices that move. The parents that file report model costs inside larger segments, so no model profit and loss exists to read. The private labs and ByteDance do not file.

That matters to an investor sizing exposure to Chinese AI and to an operator choosing a Chinese model to build on. Neither can read a maker’s staying power from its score or its price.

Most of the Field Discloses Little About What Its Models Cost

Five of the 19 makers on the map in Figure 1 file accounts that say anything about models: Zhipu, MiniMax and SenseTime, plus Alibaba's and Baidu's AI lines. Tencent, Xiaomi and Kuaishou give model figures in results releases and calls, and the rest is read through company statements and press reports.

Tap the figure to enlarge. How firm: Filed means in the maker’s own filed accounts. Company means stated by the company on a results call or release. Reported means a named press outlet, insiders or an unconfirmed filing, and should be read as unverified. None means nothing disclosed. H1 is the first half of 2026; ARR is annual recurring revenue. Run rates are not comparable across makers. Zhipu’s stated $1.6 billion compares with about $283 million from doubled H1 revenue. Ant Group and Huawei disclose at group level only. Not tabulated: the telcos, which host and distribute models and publish no model revenue; ModelBest; Shanghai AI Laboratory, a non-profit.

With so little filed, a buyer starts from what every model publishes, its score and its price. What do they say about the maker?

One Score Band, Bills Up to 24 Times Apart

Five Chinese models score between 44 and 46 on the index, as the Artificial Analysis leaderboard rounds its scores to whole numbers: Xiaomi’s MiMo-V2.6-Pro, Alibaba’s Qwen3.8 Max, Zhipu’s GLM-5.3, Moonshot’s Kimi K3 and StepFun’s Step 5 Preview. These are the five highest Chinese scores on Artificial Analysis’s board, read on October 1. The best US model on the same index scores 58, 12 points above the top Chinese score. Several other Chinese makers sit lower or are not placed.

A score says what a model can do. Artificial Analysis also reports what it cost to run the whole index on each model, which is what a buyer would pay to run the same tests. Among the five, that bill runs from $207 for MiMo to $4,935 for Qwen, nearly 24 times as much for the same band of score. GLM’s $2,503 is twelve times MiMo’s. Figure 2 shows the five bills.

Token counts can outweigh list price. A token is a fragment of a word, and models charge separately for the text they read and the text they write. Qwen’s list price for what it writes is lower than Kimi’s, yet Qwen cost $4,935 against Kimi’s $3,658 because it read 9.9 billion input tokens on the tests to Kimi’s 2.6 billion. An operator should price a model against its own workload.

Grey rows are US reference runs. Claude Opus 5.5 at its high setting scored 54 and cost $2,172, less than GLM’s $2,503 at 45.

MiMo’s list output price is $0.87 per million tokens. A price that low can reflect low cost or a decision to win customers, and nothing on the buyer’s side tells which. For what a model costs its maker, and who pays it, the evidence is in the makers’ accounts.

At the Two Listed Labs, Customers Do Not Close the Gap

Zhipu and MiniMax file accounts that are mostly the model, and have no parent business to pay for them. Zhipu’s revenue grew 399.7 percent in the half and its net loss was ¥2.07 billion. Its gross margin was 26.4 percent, down from 50.0 percent a year earlier, so revenue is growing fast at a thinner margin. MiniMax’s revenue grew 283 percent to $116.6 million. Its net loss was $358.0 million and its research spend $296.9 million, at a gross margin of 17.9 percent.

Both labs earn a positive margin on what they sell. Where a lab books compute, in cost of revenue or in research, is its own choice, so the ratios that follow are a guide. By our calculation, Zhipu’s gross profit was about ¥252 million against ¥2.13 billion of research, and MiniMax’s about $21 million against $296.9 million: research ran at about 8 and 14 times what customers left after serving costs. At both labs research spend is close to the size of the net loss, so research accounts for most of the gap.

Figure 3 sets gross profit beside research at the two labs. SenseTime, also listed in Hong Kong and maker of the SenseNova models, reports generative AI as 79.9 percent of revenue and vision AI as 17.1 percent, but we found no model-level cost in its results, so its accounts cannot show a model paying for itself. Alibaba’s AI segment is covered in the next section.

No account we read shows customers covering what a model costs to build and run, research included. Someone else is paying the difference.

A Parent Pays Only as Long as Its Core Business Holds

Three parents disclose enough to size the cover, which is the profit left after the AI spend set against that spend. The rest of the field cannot be measured. Figure 4 sets the lab ratios beside the parent ones.

In the second quarter, Tencent’s operating profit before certain items, its own adjusted measure, rose 9 percent to ¥75.6 billion, already net of about ¥10.5 billion ($1.56 billion) of cost from new AI products, which include its Hy models, the Yuanbao assistant and coding tools. The ¥10.5 billion is a cost line with some revenue inside it, and Tencent reports no AI revenue line. Cash is where the strain shows: its free cash flow, a cash measure the company defines for itself, was negative ¥13.8 billion ($2.0 billion) because of prepayments for computing, and would have been positive ¥37.6 billion without them. Its Hy3 model scores 25 on the index, against 44 to 46 for the five leaders, and its newer Hy4 preview is not placed, so the parent with the highest cover here does not hold a placed leading model.

Alibaba’s new AI Labs and Applications segment holds its model labs together with the Qwen consumer app and its QwenWork product, so its revenue covers more than model sales. In the June quarter the segment lost ¥13.86 billion ($2.06 billion) at adjusted EBITA, Alibaba’s operating profit measure, on revenue of ¥3.34 billion ($0.5 billion). Its e-commerce group earned ¥39.7 billion ($5.9 billion) of adjusted EBITA, about three times that loss, while Alibaba’s total adjusted EBITA, after the AI loss, fell 30 percent to ¥27.3 billion ($4.1 billion), about twice it.

Xiaomi, the opening’s test case, reports its profit after research. Adjusted net profit for the half was ¥12.3 billion ($1.83 billion), down 42.8 percent. Against up to ¥5.5 billion of AI research by our estimate, that is at least about 2.2 times, close to Alibaba’s group figure and well below Tencent’s. Xiaomi told analysts that revenue from selling access to its model had begun, and that it is not treating monetization as the main goal yet. A $0.87 output price is one a profitable parent can choose to hold, whatever it costs to serve.

By our calculation, on different bases and periods (quarters for Tencent and Alibaba, the half for Xiaomi), the cover is about 7 times at Tencent, about 2 times at Alibaba’s group level and at least 2.2 times at Xiaomi. On the same inputs, cover would fall to one only if profit after the AI spend dropped a further 50 percent from today’s level at Alibaba’s group level, at least 55 percent at Xiaomi and 86 percent at Tencent, with the spend unchanged. Xiaomi’s adjusted net profit has already fallen 42.8 percent, and the 2.2 times is measured after that fall. Figure 4 puts unlike measures on one scale: the lab bars use gross profit before research and the parent bars profit after the AI spend, so it shows which side of one each maker sits, not a like-for-like gap.

Baidu and ByteDance, the two largest remaining parents, cannot be sized. Baidu’s release gives revenue lines for its AI business but no AI cost figure. Its online marketing revenue fell 19 percent in the second quarter while its AI-powered line, mostly AI cloud infrastructure, grew 25 percent. ByteDance does not file. It says its Doubao models process 180 trillion tokens a day, a measure of use, and The Information reported, citing insiders, that its profit is slipping.

A parent’s cover is a moving number. Xiaomi’s adjusted net profit fell 42.8 percent in the half, Baidu’s online marketing fell 19 percent in the quarter, Alibaba’s group profit fell 30 percent with the AI loss inside it, Tencent’s rose 9 percent, and ByteDance’s profit is reported slipping. No parent’s filing says whether its model earns its keep. A lab has no parent profit to lean on, so who covers it there?

A Lab Pays With Shares, on Figures a Reader Cannot Check

A lab has no other business to draw on, so it pays with shares and bonds. Zhipu’s path ran like this. Its shares closed at a peak on June 22. In July it completed a share sale of about ¥27 billion ($4 billion), according to press reports. By September 9 the shares closed 62 percent below that peak close. On September 12 it agreed a placing of about ¥13.5 billion ($2 billion), a sale of new shares to investors, and a ¥20.14 billion ($3 billion) convertible bond, a loan that can turn into shares. That is ¥33.6 billion agreed after the fall, not yet shown as closed. The notice puts 60 percent of the placing’s proceeds to research and infrastructure, 15 percent to expansion and strategic investments and 25 percent to working capital. By September 30 the shares were 73 percent below the peak close.

Zhipu’s raises total about ¥61 billion ($9 billion) by our sum, nearly 30 times its half-year loss, so if the September deals close it is not short of cash on first-half spending. At the first-half loss rate that cash would last for years, so first-half spending does not force another sale. The loss is not standing still, though: Zhipu’s adjusted loss, which leaves out items such as share-based pay, widened by 12.1 percent from the first half of 2025. What stays open is the price of the next sale. At the September 30 close a new sale would price off shares down 73 percent. MiniMax raised about $2.0 billion in July through a placing and a convertible bond, about 5.6 times its half-year loss.

Alongside the shares, the figures that come with the funding cannot be checked. The most checkable is Zhipu’s. It stated an annualized $1.6 billion on its model-access platform in August, nearly 6 times the roughly $283 million implied by doubling its ¥954 million ($142 million) of first-half revenue, all products included. A run rate annualizes a recent period, so it runs above a half-year average at a fast-growing business, and Zhipu’s full-year accounts will test it.

Private labs publish no accounts at all. The Information reported on September 24, as relayed by other outlets, that DeepSeek’s API, the paid channel through which developers call its models, earned an 82.9 percent gross margin in the first seven months of the year. That is a margin on selling tokens, with research and training outside it, on a different base from the filed 26.4 percent at Zhipu and 17.9 percent at MiniMax, and DeepSeek has not confirmed it. A margin that high would still leave the question open, because DeepSeek does not report what its research costs. Its first funding round, according to Caixin, valued it at ¥350 billion ($52 billion) after about ¥50 billion ($7.4 billion) came in. Moonshot reportedly closed $3.5 billion at a $35 billion value in July, per Bloomberg. Neither round came with a filed account. None of these figures, filed or reported, shows a model covering its cost. What would?

Only a Model-Level Account Would Settle It

One filing would show it: a segment in which model revenue exceeds model build-and-run cost, research included. Alibaba’s AI Cloud and Compute segment sells compute and cloud services, so it does not qualify. Three nearer readings would show whether the gap is closing. Zhipu’s and MiniMax’s full-year accounts, due next year, will show whether gross profit is rising against research. Gross profit covered about one-eighth of research at Zhipu and one-fourteenth at MiniMax in the first half. Full-year gross profit reaching research, the point where gross profit alone pays for it, or a model-level account showing revenue above cost, would end the thesis. A share of at least double the first-half figures, about one-quarter at Zhipu and one-seventh at MiniMax, but short of one would show the gap narrowing with the thesis still standing. A share at or below the first-half figures would show no progress. A listing document from a private lab, with audited gross margin and research spending, would test claims like DeepSeek’s. A parent that reports a line where model revenue and model cost sit together would show it from the other side. Only the labs’ full-year accounts have a due date.

Until these arrive, staying power can be judged from two things: how well a parent’s other business covers its model, which Tencent’s, Alibaba’s and Xiaomi’s reports let a reader check, and how a lab’s share price stands when it next sells shares, which cannot be read in advance. An operator building on a Chinese model is exposed to the same two, and can ask who funds its vendor and keep a second model ready. On July 19 Moonshot said it had paused new consumer subscriptions because requests to Kimi K3 were approaching the limit of its computing clusters. Its statement did not mention the API, so an operator on Kimi should ask whether API capacity is limited too. An investor can size exposure by asking how well a maker’s other business covers its AI spending and when it next needs to sell shares.

Xiaomi, the maker with the lowest bill, covers its AI research at least 2.2 times from adjusted net profit that fell 42.8 percent in the half. Zhipu has raised or agreed to raise nearly 30 times its half-year loss while customers’ gross profit covers one-eighth of its research, so its staying power rests on those raises closing and on the price of its next sale, with its shares 73 percent below their peak.

Data covers developments through September 30, 2026; Artificial Analysis’s board and model pages were read on October 1, 2026 (Intelligence Index v4.3.2, effort settings as Artificial Analysis labels them, no update date shown on the board). Share prices are frozen at the September 30 close. Currency conversions use the People’s Bank of China central parity rates for September 30: $1 = ¥6.7351 and HK$1 = ¥0.85842. Private valuations, run rates and insider figures are reported, not filed. Xiaomi’s AI research amount, the Zhipu raise total (including the ¥33.6 billion agreed in September and the HK$ placing converted to yuan), gross profit against research at Zhipu and MiniMax, the Tencent, Alibaba and Xiaomi cover multiples and breakeven declines, the run-rate and raise multiples (nearly 6, 5.6 and nearly 30 times), the 62 and 73 percent share price declines, and the bill ratios are ICM calculations from the inputs shown.

Inside China’s Machine is independent research, not investment advice.

Sources and Data Attribution

Scores, costs and token counts: Artificial Analysis leaderboard and model pages (artificialanalysis.ai), read October 1, 2026; effective prices and ratios are ICM calculations from those fields. List prices: Artificial Analysis for Step 5 Preview and GLM-5.3 output prices; Moonshot (platform.kimi.ai); Alibaba Cloud Model Studio (Singapore tier).

Filings and company statements: Zhipu and MiniMax interim reports and HKEX announcements (2026), including Zhipu’s September 12 to 18 placing and convertible announcements; SenseTime interim results (HKEX, August 26, 2026); Alibaba June-quarter results (Form 6-K, August 2026); Tencent second-quarter results release and call (August 12, 2026); Baidu second-quarter results (August 18, 2026); Xiaomi results announcement and second-quarter call (August 2026); Kuaishou second-quarter results release (for Figure 1’s Kling row). Zhipu closing prices: Yahoo Finance.

Reported figures: The Information (September 15 and 24, 2026), as relayed by other outlets; South China Morning Post, Bamboo Works and Bloomberg for Zhipu’s July share sale; Caixin (May and July 2026); Bloomberg as relayed by Silicon Republic and TechWire Asia; Yicai (July 19, 2026) for Moonshot’s statement; LatePost and 36Kr via Sina for Kling’s 2025 loss (Figure 1); ByteDance’s token figure as relayed by Sohu, IT Home and Tencent News from its June 23 event; iFlytek and Meituan figures as reported by Chinese financial media; Wang Xiaochuan to 36Kr (May 23, 2026); Bloomberg via TNW for 01.AI.

Zhipu’s August annualized figure and Figure 1’s remaining rows: company statements as relayed by Chinese financial media.

Currency: People’s Bank of China central parity, September 30, 2026, as published by Xinhua.

This is the inaugural article in a quarterly series mapping China’s humanoid robot landscape.

Three numbers describe China’s humanoid boom, and none of them counts a robot doing paid work: a robot performing a task its owner pays for, as opposed to a demonstration, a research platform or a data-collection rig.

The first is units. IDC puts China’s humanoid shipments in the first half of 2026 at about 19,000. The second is price. When Unitree debuted on August 19 on Shanghai’s STAR Market, the exchange’s technology board, the market briefly valued it at ¥445 billion ($63 billion), 262 times its 2025 revenue. By September 24 the stock had lost 56 percent from that peak. The third is scenarios. In June, the Ministry of Industry and Information Technology (MIIT) and the state asset regulator SASAC launched a “real-scene training” action for humanoids. Its year-end targets are written in scenarios, more than 100 high-value ones, with no count of robots at work.

Each is real. Each measures something other than work, and the sections below take them in turn.

The First Number: Units Are Not One Unit

Volume is large by every tracker: Omdia estimated about 13,000 humanoids shipped worldwide in all of 2025, and Counterpoint Research estimates more than 22,000 worldwide for the first half of 2026 alone. The two come from different trackers, so read the jump as direction, not a growth rate. IDC’s 19,000 for China alone, above, is a different scope again. No source used here says what the robots are used for or in what mix.

A unit is not a fixed quantity of anything. Unitree’s compact R1 Air starts at ¥29,900, about $4,200, and its G1 starts at ¥85,000. UBTECH’s FY2025 annual report shows ¥821 million of full-size humanoid revenue on 1,079 units, about ¥761,000 each, or $107,000. Unitree’s prospectus shows ¥868 million of humanoid revenue in 2025 on 5,215 units sold (5,511 shipped), about ¥166,400 each, or $23,400. Those are segment revenue divided by units sold, not list prices. Unitree’s figure is also falling: the prospectus puts its average humanoid revenue per unit at ¥260,400 in 2024 and ¥593,400 in 2023, so 2025 was a 36 percent drop in a year.

Set side by side, the two companies booked humanoid revenue within 6 percent of each other, and one sold nearly five times as many robots. A thousand of one and a thousand of the other are different industries sharing a word.

That matters for the market shares the trackers publish. In the first half of 2026, two trackers, Counterpoint and Smart Analytics Global, put AgiBot at 43 to 44 percent of global units and Unitree at roughly 31 percent. Those are shares of units. A share of units says nothing about a share of revenue, and none of these trackers publishes the second.

AgiBot’s own counter is wider still. Its milestone of 15,000 robots by June 28 counts embodied-AI robots across all its product lines, including wheeled and industrial units, so it cannot be checked against tracker counts of humanoids.

The field, company by company

Stage definitions: pre-commercial means announced or unveiled with no external sales; pilot means paid pilots or a first production line; commercial means sales at volume from a running production line. Commercial does not mean verified paid work, which no company here discloses. Three companies sit in that top tier, and they are not alike: Unitree and AgiBot ship thousands of units a year, while UBTECH ships about a thousand full-size robots at roughly five times the price.

The Second Number: The Price of a Whole Company

Unitree: four prices in five weeks

Unitree priced its STAR Market listing at ¥150.80 per share on August 6, 2026. It reported 2025 revenue of ¥1.699 billion and net profit excluding non-recurring items of ¥591 million, a margin of about 35 percent on that basis. On debut, August 19, the stock rose 629 percent intraday to ¥1,100 and closed up 460 percent at ¥845. The company’s first-half 2026 results, published August 17, showed revenue of ¥1.152 billion, up 48.5 percent, and net profit excluding non-recurring items of ¥244 million, down 19.3 percent. That is a margin of about 21 percent.

Every market value below is share price times about 404.5 million shares, converted at ¥7.1 per dollar, so the four points share one method.

Unitree prices are frozen at the September 24 close.

By September 24 the stock was 56 percent below its peak and 42 percent below the day-one close. It still stood at 3.2 times the IPO price, about 116 times 2025 revenue and about 334 times 2025 net profit excluding non-recurring items. The multiple applies to the whole company. The prospectus puts humanoids at ¥868 million of the ¥1.699 billion, about half, with quadrupeds at about 42 percent, so it is not a humanoid multiple.

A price of 116 times revenue is a bet on growth far beyond current sales. The dedicated Unitree piece covers the company’s full arc.

What ¥197 billion needs

At 30 times mature earnings, the September 24 value needs about 95,000 humanoids a year at Unitree’s first-half margin, or about 58,000 at its 2025 margin. Unitree’s first-half 2026 shipments of 5,900 to 7,000 imply 12,000 to 14,000 a year. So the value needs seven to eight times the current run rate at the lower margin and four to five times at the higher one, with revenue per robot held at the 2025 level. “Mature earnings” means the profit Unitree might earn once growth slows, and 30 times is a mid-range multiple chosen for illustration. This is built from Unitree’s disclosed numbers, not company guidance, a forecast or a fair-value model.

Each step is division: profit is the market value over the multiple, revenue is profit over the margin, humanoids are about 51 percent of revenue, and each humanoid brings in about ¥166,400. Two caveats move the answer most. The table keeps today’s product mix, so quadrupeds grow alongside humanoids. If quadrupeds and other lines stay flat, the humanoid volume needed is about 1.9 times higher, roughly 180,000 at 30x and the 21 percent margin. Mix can also cut either way: a shift toward cheaper models such as the R1 Air, at ¥29,900 against a ¥166,400 average, would raise the volume needed, and a shift toward full-size robots near UBTECH’s ¥761,000 would lower it. Smaller caveats: both margins come from a small book and are not proven rates at scale, and the table ignores the years it takes to get there.

Unitree: who buys the robots

The prospectus does not split humanoid sales by buyer type, so no buyer-mix percentage exists to quote. What it does say bears on the question.

The company lists five uses for its humanoids: scientific research, application development, education, cultural performance and smart services. Paid work in a factory or warehouse is not on the list. In the same filing it also says that most humanoids on the market remain at the stage of application-scenario testing, with no large-scale commercial deployment in daily-life or industrial settings.

Sales are spread thin. The five largest customers were 12.1 percent of 2025 revenue, and the largest, a European buyer, was 3.1 percent; JD.com was second at 2.9 percent. About 51 percent of revenue came from direct offline sales, 38 percent from distributors and 11 percent from online channels. That concentration is low, which means no single buyer explains the volume, and it also means the filing gives no view of what the many buyers do with the robots.

One buyer is visible: Galbot, itself a humanoid maker and a related party, bought ¥18.2 million of Unitree products in 2025, up from ¥0.4 million the year before, for its own research and for resale after secondary development. That is a humanoid company buying humanoid parts and platforms, which fits research and development demand more than paid work.

The hinge is repeat demand. Revenue from a lab, a data-collection program or a demonstration is real revenue, but it need not recur: those buyers purchase to learn or to show. Sustaining 95,000 a year needs demand that comes back, and a reorder is the first sign that a buyer found the earlier robots worth their price. It is still not proof of paid work, since a lab can reorder too, so it counts only alongside what the buyer uses the robots for. No company disclosure found so far reports repeat customers or their use.

UBTECH: similar humanoid revenue, different robots

UBTECH booked ¥821 million of full-size humanoid revenue in FY2025 and lost ¥790 million. At about HK$56 billion (¥51 billion) in April, the latest market value found, it was priced close to Unitree’s ¥61 billion IPO value, on FY2025 humanoid revenue within 6 percent of Unitree’s. Unitree’s stock then rose to ¥445 billion at its August peak and stood at ¥197 billion on September 24. Humanoid revenue does not explain that gap. Unitree earned ¥591 million excluding non-recurring items where UBTECH lost money. The requirement table above prices growth in volume and profit, not today’s humanoid sales. The dates differ by months, so read the comparison as an order of magnitude. The dedicated piece on UBTECH’s Walker S2 covers what factory autonomy saves in labor cost, and what it does not yet.

AgiBot: the next price

AgiBot is preparing a Hong Kong listing, with no public prospectus and no price yet. Reporting in July, citing an AgiBot investor quoted by Caijing, put its expected valuation at HK$40 to 50 billion, about $5 to 6 billion, a reported target and not a price. AgiBot’s Deng Taihua put 2025 revenue at ¥1.05 billion in April, a management figure, not a filed one. ¥4 billion is a market projection for AgiBot’s 2026 revenue. At HK$1 = ¥0.91, the reported target is about ¥36 to 46 billion. AgiBot is also reported to offer robots on rental terms, so its unit and revenue counts may not map cleanly to sales.

On trailing revenue, the reported target would list AgiBot near Unitree’s IPO multiple and far below its September 24 one. The bases differ, management-disclosed against prospectus and trailing against projected, so the table is a range, not a verdict. A filing would supply audited revenue and units, and the offer price would then show where the multiple lands. Whether the filing also shows what share of its units do paid work is the open question. The dedicated AgiBot piece covers the company in full.

The Third Number: The State Counts Scenarios

The June MIIT and SASAC notice sets a year-end goal: application verification and regular deployment in representative scenarios, what the notice calls entering “operating mode.” Its figures are all scenarios. It calls for more than 100 high-value application scenarios, at least 20 key scenarios from each of ten provinces and municipalities, and at least 10 from each named central state enterprise, the large firms the central government owns. It also aims to help form deployment capability at “ten-thousand-unit scale.” That is a capability aim, not a count of robots deployed, and the notice sets no count.

The scenario count cannot supply one, and the notice’s figures do not say who pays for the robots in those scenarios, so they do not show whether a scenario becomes an order or stays a demonstration. As an illustration, take the smallest headline figure, 100 scenarios. At 10 robots each that is 1,000 robots. At 100 each it is 10,000, the notice’s aim. The same target fits an industry of a thousand working robots and one of ten thousand.

The notice does add something the trackers and the market lack: verification. Results are to be checked by the user units or by entrusted third parties under scenario-specific test procedures, with MIIT and SASAC tracking and evaluating. If those results are published with robot counts, they would be the first count of robots deployed in verified scenarios, a step toward measuring work.

What to Watch

Each of the three numbers has a nearest step toward measuring work. For units, a tracker that splits shipments by use: paid work, data collection, research, performance. For scenarios, a published MIIT and SASAC verification result that gives the number of robots behind each scenario. For price, any disclosure of repeat customers, and Unitree’s third-quarter report, due by the end of October, its second set of figures since listing.

One set of Unitree results cannot show a sevenfold rise in volume. It can show whether humanoid revenue per unit stops falling from ¥166,400 while units climb, whether the margin recovers from 21 percent, and whether customers reorder and what they use the robots for. Without that, the ¥197 billion is pricing work that no disclosure yet shows.

Inside China’s Machine is independent research, not investment advice.

Sources and Data Attribution

Shipment estimates: 2025 (AgiBot, Unitree, global total): Omdia “General-purpose Embodied Intelligent Robots” report, January 2026, plus Unitree’s prospectus for its own 2025 figure. Independent, unaudited. H1 2026: Counterpoint Research (global, above 22,000), Smart Analytics Global (global, 19,100), IDC via TechNode (September 29, 2026; China above 19,000 of nearly 25,000 worldwide).

MIIT/SASAC notice on the 2026 humanoid and embodied-intelligence real-scene training action (June 2026): official text as reposted by Xinhua, Sina Finance and Eastmoney.

UBTECH financials: 2025 annual report (HKEX: 9880), with FY2025 results also reported by Sina Finance and Huxiu. Dobot: HKEX FY2024 results (HKEX: 2432).

AgiBot funding and IPO plans: Bloomberg, Reuters, Caixin Global, TechNode, and Caijing/Sina (July 24 to 25, 2026) for the HK$40 to 50 billion figure, attributed there to an AgiBot investor; 2025 revenue per Deng Taihua’s April 2026 disclosure (Tencent News, Sina Finance, ITHome), not a filed figure; the ¥4 billion 2026 figure is a market projection. AgiBot production milestones: company announcements (June 28, 2026), as reported by Sina Finance.

Unitree financials and customers: STAR Market prospectus (2025 revenue ¥1.699 billion; net profit excluding non-recurring items ¥591 million, 扣非净利润; humanoid units sold and shipped; average revenue per humanoid; top-five customers; channel split; Galbot related-party sales; IPO placement investors; R1 Air and G1 starting prices; filing language on the humanoid application stage). First-half 2026 results (revenue ¥1.152 billion, net profit excluding non-recurring items ¥244 million): Unitree’s August 17 update as reported by 21st Century Business Herald and Sina Finance. IPO pricing and post-listing trading: CNBC, Bloomberg, Nikkei Asia, The Robot Report, Outlook Business, Sina Finance, company disclosures; listing-day peak and close reconciled against Chinese-language coverage; September 24 close (¥488.00) from Baidu Finance, Futu and Ifeng quote pages. Market caps are share price times about 404.5 million shares, converted at ¥7.1/$. The “What ¥197 billion needs” tables and charts are ICM’s own illustrative arithmetic on those disclosed figures. Galbot funding: Caixin Global, Caproasia, 36Kr (36氪), company press releases. Fourier: TechCrunch, Xinhua, company announcements; 2026 update via Caixin/Cailianshe coverage of the Runyang Technology investment.

Tech giants: BYD’s Xiao Di unveiling (Sina Finance, Guancha, 21st Century Business Herald, Interesting Engineering); XPeng IRON production line (company announcement and trade press, Sept 2026). Matrix Robotics: company disclosures and WAIC/BEYOND Expo Macao coverage.

All private valuations are Estimated from the most recent disclosed round unless marked Confirmed. Multiples use ¥7.1/$, HK$1 = ¥0.91, and are pinned to the dates shown. Where sources conflict, the piece notes the discrepancy.

The week in one idea

Unitree’s post-IPO slide looked like a correction of an irrational debut print, and this week’s coverage largely told it that way. Its disclosure, filed before the stock ever traded, describes something more specific. First-half revenue grew 48.5 percent. The profit line that strips out one-time items fell 19.3 percent, because research and marketing spending grew faster than sales did. Buyers who pushed the stock 629 percent above its IPO price were pricing in years of unusually high growth and strong profitability. The filing instead shows growth decelerating and its defense turning expensive, not a business that stopped working. That gap will matter to more than one stock: Unitree’s repriced multiple is becoming a new mainland reference point for Chinese humanoid and embodied-intelligence listings.

The number that matters

48.5% revenue growth, -19.3% adjusted profit. Unitree’s first-half 2026 results, disclosed August 17.
Why it matters: The market’s opening print valued Unitree 629 percent above its IPO price, leaving little room for growth deceleration and heavier reinvestment. The interim filing shows both: decelerating growth and shrinking adjusted profit, as R&D and marketing costs outrun sales.

What changed

The retreat resumed after a brief pause. Unitree closed at 514.98 yuan on September 18, a month after its August 19 listing, up 2.33 percent that day and briefly described in press coverage as the stock finding a floor. That description did not last: Nikkei Asia’s September 21 report, with retail investors reassessing humanoid robots, put the decline at more than 50 percent from the opening price. By September 23 the stock had slipped further, to around 491 yuan. Against the 1,100-yuan opening print on debut day, that is a decline of about 55 percent and roughly 246 billion yuan of paper value gone, on a share count of 404.5 million unchanged since listing.

The swing to profit is entirely a base effect. Unitree’s net profit attributable to shareholders was 274 million yuan in the first half of 2026, against a loss of 32 million yuan a year earlier. That reversal is more than explained by one item: a 349 million yuan share-based compensation charge, larger than the entire reported swing, tied to an employee-incentive capital injection and booked as non-recurring. Strip non-recurring items from both periods and the picture reverses: adjusted net profit fell to 244 million yuan this year from 302 million yuan, down 19.34 percent. The business did not swing from loss to profit; on the comparable measure, it went from a larger profit to a smaller one.

Growth and margin were already moving before this quarter. Revenue growth has decelerated sharply from 2025’s pace: up 332.64 percent in the first quarter of 2025, down to 68.49 percent in the first quarter of 2026, and to 48.54 percent for the full first half. Unitree’s own filing attributes the shrinking adjusted profit to research spending up 82 million yuan year on year and to marketing spending built around a televised Lunar New Year robot-dance campaign. Gross margin on the humanoid product line, not the company overall, fell from 87.67 percent in 2023, when volumes were still tiny, to 63.18 percent in 2025, as the business scaled into higher-volume, lower-priced models. None of this is new this week. It is the base rate the market is now pricing against.

The repricing sets the next company’s number, not just Unitree’s. An embodied-intelligence investor quoted by Beijing News’s financial desk made the point directly: Unitree’s opening-print valuation, briefly above 400 billion yuan, was never sustainable. As the first humanoid maker to list on a mainland exchange, Unitree gave still-private Chinese humanoid and embodied-intelligence companies a mainland reference multiple that did not exist a month ago.

What it changes

Unitree’s slide is not evidence the humanoid opportunity failed. Revenue is still growing rapidly, and the humanoid line remains highly profitable on gross margin despite the decline from its early, low-volume economics. What changed is the assumption embedded in the opening print: that growth would stay exceptional while margins held. The interim filing instead shows a company spending heavily on research and brand to defend share, in a market where a widely cited industry report puts Chinese humanoid shipments above 40,000 units in the first half alone, a figure other trackers count meaningfully lower, with new entrants arriving either way. That is an ordinary trade-off between margin and market position, not a demand failure, though it prices differently from a company assumed to compound for free.

Model update: Read Unitree’s post-listing volatility as the market recalibrating a growth-and-margin assumption rather than withdrawing conviction in the sector, and expect the next Chinese embodied-intelligence listing to be priced against that recalibrated multiple, not against August’s opening print.

The strongest countercase

A well-informed skeptic would say most of this decline is a genuinely irrational opening print correcting itself, not a reassessment of the business. Unitree priced its IPO around 61 billion yuan; the market’s opening print, on thin float with no daily price limit for five trading days and only intraday halts as a backstop, briefly valued it near 445 billion yuan before settling closer to 200 billion now, making the opening print itself the unusual data point, not this month’s retreat from it. Unitree’s filing describes no material adverse development in the first half, and at around 491 yuan the stock still trades more than three times its offer price. Chinese commentary has called the long-term robotics thesis intact while calling the opening print unsustainable.

Medium confidence: the earnings mechanism is disclosed and consistent across independent outlets, but how much of the price move reflects thin float and sentiment rather than the earnings read cannot be cleanly separated with this week’s evidence.

Since last week

Two forward tests from last ICM weekly remain open with no new evidence. CXMT’s board has not yet disclosed how it will use the roughly 36.8 billion yuan of net IPO proceeds above its original project plan, freed up after the greenshoe’s full exercise, and its third-quarter report is not due until late October. Huawei’s fourth quarter, the test for whether 950DT shipments catch up to price, has not started. Separately, CXMT announced this week that its fifth-generation DRAM platform has entered mass production, narrowing its process gap with the leading global memory makers on metrics like active-area half-pitch, though the announcement names no HBM product built on it.

What we’re watching next

Unitree’s next disclosure. Research and marketing spending growing in line with revenue, rather than faster, would say the reinvestment phase is ending. Another double-digit adjusted-profit decline alongside slower revenue growth would strengthen the case that growth is becoming more expensive.

The next embodied-intelligence listing. A Chinese humanoid or component maker pricing its next raise or IPO below Unitree’s post-retreat multiple would confirm the market has reset the sector’s reference valuation. A comparable multiple would suggest this month’s repricing has not yet reached the private market.

Inside China’s Machine Weekly is research, not investment advice.