This continues the series on China’s two listed AI model companies. It began with “Running Out of Runway“ in December 2025 and ran through their first annual reports and their first-half results in August and September 2026.

Three AI model companies have now published accounts for the same six months. Zhipu, the Tsinghua University-backed lab that listed in Hong Kong in January, reported on August 31. MiniMax, the Shanghai-based lab that listed a day after it, reported on August 26. Anthropic, the maker of Claude, filed its S-1 with the US Securities and Exchange Commission on October 26, with interim accounts for the six months to June 30, 2026. Until this week a reading of the two Chinese companies had no such comparable.

Revenue first. Anthropic booked $16.3bn in the first half. Zhipu booked Rmb 953.9m, about $142m at the company’s own conversion. MiniMax booked $116.6m. The gap is more than a hundred times. All three grew fast: Anthropic from $1.2bn a year earlier, Zhipu by 399.7%, MiniMax by 283.1%.

Then the cost of serving that revenue. Anthropic’s filing presents its expenses by function in the notes. Cost of revenue, which it says consists almost entirely of compute for customer workloads, came to $2.4bn, or about 15% of revenue. Zhipu’s cost of sales took 74% of revenue, for a blended gross margin of 26.4%; the company attributes most of the increase in that cost to computing service fees. MiniMax’s gross margin was 17.9%.

Then research. Anthropic’s R&D expense, which carries training compute as well as salaries, was $9.7bn, or 60% of revenue. Zhipu’s R&D ran at 223% of revenue, MiniMax’s at 255%.

Then the bottom line. Anthropic’s operating loss for the half was $693m, about 4% of revenue, and its second quarter showed an operating profit of $974m. Most of the headline net loss of $80.7bn, about $72bn, is an accounting charge on convertible notes that rose in value with the company’s valuation. Both Chinese companies also carried financing-related charges, which fell after their pre-IPO instruments converted at listing. On the adjusted measures both prefer, Zhipu lost Rmb 1.96bn, about $292m at the same conversion as its revenue, and MiniMax lost $293m, two and two and a half times their revenue. The loss measures are not on one basis: Anthropic’s operating loss includes stock-based compensation, and the Chinese adjusted figures exclude it and include items below the operating line. The revenue and R&D ratios carry the comparison.

In December I argued that Zhipu and MiniMax were caught in a structural trap: profitable per API call, loss-making as businesses, with the required level of spending set by competitors rather than by their own plans. Anthropic’s gross profit, taken from its functional expense table, exceeds its R&D expense, a relationship neither Chinese company can show. The FT and Bloomberg frame the filing as a test of investor demand for AI stock. Bloomberg also examines its compute intensity. Neither goes on to set those costs against Zhipu and MiniMax, which is what this article does. I read the filing as proof that the trap closes at scale, and as a statement of the condition: a gross margin near 85%, which the second section gives my reasons for reading as mostly price. Zhipu and MiniMax run the same ratios at a hundredth of the revenue and between a fifth and a third of the margin. In September I wrote that Zhipu’s pricing power rests on release speed and workflow depth. The S-1 adds the test behind that, whether pricing power produces enough gross profit to fund research.

Hello China Tech publishes new analyses every week on China's AI, chips, robotics, and EVs, built from primary sources rather than second-hand coverage. If setting a US frontier lab's S-1 against the Hong Kong filings of China's two listed model companies is the kind of reading you want more of, subscribe free and get each one as it publishes.

The Same Ratio, Different Stages

The series has tracked one ratio since March: R&D spending against revenue. When it falls, revenue is growing faster than the cost of staying competitive. This series has called that the scissors closing.

Anthropic’s filing gives that ratio for three and a half years. In 2023 the company spent $422m on R&D against $41m of revenue, a ratio above 1,000%. In 2024 it was 589%. In 2025 it was 138%. In the first half of 2026 it was 60%. The same filing gives the quarterly path for compute and infrastructure expense as a share of revenue: 251% in the first quarter of 2025, 48% in the second quarter of 2026.

Zhipu and MiniMax report higher ratios. Zhipu’s R&D ratio fell from about 700% in 2024 to 440% in 2025 to 223% in the latest half. MiniMax’s fell from 619% to 320% to 255%. Anthropic’s earlier ratios give a comparison, not a timetable: it passed through the levels they report now at some point between 2024 and 2025, and both Chinese companies made the falling ratio a point in their own reports.

The filing also explains the mechanism, and the explanation matters more than the ratio. Anthropic says it procures compute 12 to 18 months before it becomes available, so its compute expense “is generally fixed in the near term” and rises as committed capacity arrives, not in response to demand in the period. Revenue then grew faster than the cost of compute capacity committed in advance. In the second quarter of 2026 revenue rose 1,368% from a year earlier while compute expense rose 300%.

Zhipu built a metric for a related relationship from the other side. Its interim report defines a “compute multiplier”: open-platform and API revenue per Rmb 1 spent on training and inference compute. The company says the multiplier rose about 14 times from a year earlier. Anthropic’s nearest number, total revenue divided by compute and infrastructure expense, went from 0.4 in the first quarter of 2025 to 2.1 in the second quarter of 2026. Both ratios relate revenue to compute spending, but the revenue covered and the cost definitions differ.

The absolute levels cannot be compared. Anthropic’s ratio crossed one. Zhipu does not disclose the absolute value of its multiplier, and its cost of sales alone takes 74% of revenue.

Where the Margin Comes From

The 85% figure needs an accounting caveat first. Anthropic no longer reports gross profit as a line item. The 15% comes from a supplemental table in the notes that regroups the same expenses by function. That breakdown puts the compute for free users and the fees paid to cloud partners in sales and marketing, and the compute for training and internal use in R&D. Cost of revenue carries customer-serving compute plus a small share of personnel and overhead. Zhipu attributes most of the increase in its cost of sales to computing service fees. The materials do not establish how closely that line matches Anthropic’s cost of revenue. After subtracting cost of revenue and all sales and marketing expenses, Anthropic has 62% of revenue remaining. Zhipu has 26% before selling costs, MiniMax 18%. Price is the most visible part of that gap. Zhipu’s own numbers show the limit of the price argument. It reported an 80% cut in unit inference cost in its first-half review; its API margin for the half was 24.6%; the 101% rise in average API price runs to the report date in September. What remains is utilization and scale, and neither Chinese report gives a unit cost that would settle how much.

The three companies’ list prices on October 27 sit against those margins. Claude Opus 5.5 charges $4 per million input tokens and $20 per million output tokens. Fable 5.1, the most expensive model listed, charges $10 and $50. Sonnet 5.5 charges $2 and $10. Zhipu’s flagship GLM-5.3 charges $1.40 and $4.40. MiniMax’s M3, on its standard tier for inputs up to 512,000 tokens, charges $0.30 and $1.20, after a 50% launch discount that the company made permanent.

On output tokens, GLM-5.3 sells at about a fifth of Opus 5.5. M3 sells at about a sixteenth. At the cheap end of each range the gap disappears: for prompts up to 100,000 tokens, Claude Haiku 5.5 charges $0.10 and $0.50, within a few cents of GLM-5.3-Flash at $0.15 and $0.50.

The September installment described the two price curves Zhipu’s management had drawn on its earnings call: falling prices for any given level of intelligence, and a sustained premium for frontier models that open new tasks. The list prices show who stands where. The filing does not split Anthropic’s revenue by model. It does say 80% of first-half revenue was consumption-based, and that as of September 30 about 250 customers had each paid at least $10m over the preceding twelve months. The aggregate cost ratio cannot say which models earn that revenue. The two Chinese companies’ blended margins carry more than text APIs: MiniMax earned 36.6% of first-half revenue from consumer products and 60.8% of its revenue outside mainland China, and Zhipu earned 13.5% from on-premise deployment. List prices cannot explain those margins on their own. They do show the top of each company’s range. Zhipu doubled its average API price this year and still lists its flagship at a fifth of Opus 5.5’s output price. MiniMax listed its flagship at double the old price, discounted it by half from day one, and made the discount permanent.

The S-1 names the segment where the two Chinese companies sit. Among its competitors it lists “open-weight or open-source” models “released by international and domestic organizations” that compete “for a growing range of enterprise workloads that prioritize cost efficiency over intelligence.” That is the lower curve.

Anthropic also describes how it cuts prices. It says Opus 5.5 reduced customer cost per task by about 40% compared with Opus 5, and that compute efficiency lets it deliver those reductions “while preserving our margins.” In China’s model market, price moved first: in the price war of late 2024 and early 2025 Zhipu cut its flagship by as much as 90%, as this series recorded in April. Zhipu’s price increases this year run the other way. Its interim report attributes the API margin of 24.6% to scale, higher pricing and engineering work that lowered inference cost.

The checkpoint for that attempt sits on the price pages as well as in the next results. The comparison to track is list prices and discounts against GLM-5.3’s $4.40 and M3’s $1.20. A higher launch price would test demand for a premium Chinese model, and the results after it would show whether customers paid it and whether the margin followed. A price cut alone would not settle either company’s long-term position, since cuts also come from cost declines and product tiers.

Both Formulas, One Company

Each Chinese company has stated a survival equation. Zhipu’s annual report opened with one: the commercial value of artificial general intelligence, or AGI, equals the intelligence ceiling multiplied by token consumption. MiniMax’s interim announcement gave the counter-formula: inference efficiency decides how much post-training and how many experiments a lab can run, which decides how fast its models improve, so cheap inference is a production function for intelligence.

The S-1 contains both. Zhipu’s equation appears as net revenue retention of about 900% as of September 30. The filing calculates it by dividing current-period revenue by base-period revenue for the same qualified customers, and excludes Claude Pro and Max subscribers. It also appears as more than 11,900 customers that each paid at least $100,000 over the preceding twelve months. More capable models make more tasks possible, and the tasks are metered. MiniMax’s equation appears as what the filing calls “compute multipliers”: algorithmic gains in training and inference that Claude now helps discover. The company calls that recursion “closing the loop” and claims it serves more customers per unit of compute than other approaches.

An April installment of this series argued that China’s independent model companies could pursue only one commercial lane at a time, while platform companies could run several. The same budget logic applies inside a lane, to capability and efficiency. A frontier lab with $16bn of half-year revenue funds both from a research budget that takes 60% of revenue. Zhipu and MiniMax chose one half each because at 223% and 255% revenue cannot fund both. The likelier explanation for the choice between the two formulas is a budget constraint rather than a disagreement about how AI makes money. Anthropic’s gross profit now covers its R&D expense. Neither Chinese company’s does.

The constraint is also why the two Chinese companies have started to converge. Zhipu says GLM-5.3-Flash beats GLM-5.2 at a tenth of its price, a move onto the cost frontier. MiniMax is training M3 Pro, a model it says approaches 3 trillion parameters, to reach the capability frontier. Both companies are now working on the half they did not start with.

China in the S-1: Attacker, Risk, Student

The S-1 mentions China in three roles that matter for this comparison, and the third is the one that bears on the two companies.

The first is as an attacker. Anthropic says that in September 2025 it detected a campaign in which a Chinese state-sponsored group manipulated Claude Code into attempting intrusions at about 30 organizations, which it calls “the first documented case of a large-scale cyberattack executed without substantial human intervention.” The second is as a regulatory risk. The filing says Washington has increasingly scrutinized AI companies’ exposure to Chinese state actors, and that future rules could restrict its ability to serve users connected to China.

The third is as a student. The filing warns that third parties “have attempted in the past and will likely continue to attempt to replicate, distill, extract, or otherwise derive the functionality or capabilities of our models.” It does not name them there. Anthropic’s September threat report did: it said Zhipu rotated through 273 fraudulent accounts over ten days, attributed more than 3.4m exchanges to Zhipu across 17 days in June and July, and said MiniMax built a proxy network through a shell company. The filing describes the segment both companies sit in without naming them. The threat report names them, as users of the premium curve’s output.

The next sentences matter more for the two Chinese labs. If the ability to build highly capable models spreads, the filing says, “competition could increasingly depend on access to large-scale compute rather than on model development capability.” Anthropic lists that as a risk to its differentiation. For Zhipu and MiniMax, cheaper inference and open weights are the stated strategy, so the same sentence names their constraint: the contest then moves to compute access, and the filing also lists US export controls on advanced chips to China among the factors shaping compute supply.

Anthropic had raised $141.1bn of capital by June 30 and held $112.4bn in cash and $19.3bn in short-term investments. Bloomberg reported that compute agreements signed since the end of 2025 carry minimum obligations of about $417bn, some running 20 years; the filing leaves the total blank in this draft. Zhipu raised about Rmb 27bn in a July placement, then HK$15.66bn in a September placement and about $3.01bn net from a convertible bond the same month. Its interim report says it runs inference on about 100,000 domestic chips; on its earnings call, management said training supply remains heterogeneous. MiniMax reported a cash-balance measure of $1.32bn at June 30; its July placement and convertible bond raised HK$16bn gross, and management put reserves above $3bn. Both companies are in what the July installment called the perpetual raise, returning to the market before revenue can fund the next stage. Anthropic is raising too. The size of its raise is the contrast. Its revenue already covers its research budget, though not yet the whole cost base for the half, and its shrinking operating loss points to expansion being funded from revenue.

Capital First, or Model First

The December article emphasized balance-sheet depth. The March update emphasized distribution. The S-1 corrects both. At Anthropic the balance sheet and the model moved as one variable. Anthropic committed compute 12 to 18 months before the capacity became available, customer demand used the capacity, revenue then grew faster than compute expense, and the revenue funded the next model.

Zhipu and MiniMax have run the same loop in a different order. The model came first, because it is what the market would fund. Compute is bought as a service and prepaid, on terms the reports do not disclose, at a margin that leaves little to reinvest. Revenue arrives at prices set by a market with several comparable labs and platform companies able to sell below cost. Both companies have improved unit economics, and their gross profit still falls short of R&D spending. After the July and September raises, Zhipu holds capital on a scale closer to Anthropic’s order. My next test is whether Zhipu can use its larger capital base to secure compute that revenue can support. Prepayments and purchase commitments are two measures to watch in its next annual report.

On October 22, iFlytek released Spark X2.6. iFlytek is a Chinese speech-recognition company that has become one of the country’s main developers of large language models. Its chairman, Liu Qingfeng, told the audience that the new model was trained entirely on Chinese chips. He did not say which chips.

In its 2025 annual report, published in April, iFlytek said it planned to release a flagship model on the Ascend 950 at this same event. The Ascend line comes from Huawei, the Chinese telecoms equipment maker that also designs AI chips. Six months later the model arrived and the chip name did not.

Huawei’s roadmap gives chip specifications. It does not give training comparisons with Nvidia. iFlytek’s earnings briefings do. iFlytek has trained its main models on Huawei hardware since 2023. Since April 2025 it has given investors at least six readings against Nvidia. It has also said where the gap sits and how many months it adds to each model.

Read together, the readings show how much the answer depends on the task and the baseline. On the tasks iFlytek reported, tuning brought the 910B close to Nvidia’s older A100 and A800. Sampling on very long contexts stayed about five times slower than on the newer H200. iFlytek’s own explanations point most often to memory and training software. I read the record as the most detailed customer-side evidence that Huawei’s chips can carry a serious training program. Their weakest reported point is long-context sampling.

If reading a chip's training gap through its customer's own disclosures is new to you, this is a preview of what Hello China Tech does every week: reading China's AI, chip, robotics, and EV sectors from primary sources most English coverage never touches. Subscribe free to get every new analysis as it publishes.

The figures also need context. iFlytek has its own reasons to describe the gap the way it does, and those reasons shape how readers should assess each comparison.

A Flagship Without a Chip Name

Liu’s launch speech gave a training gain and a scheduling method, but no chip. iFlytek said it raised training efficiency on very long contexts about four times. It split long and short tasks and scheduled them across “existing domestic computing power.” That phrase does not identify the chips used for X2.6. “Domestic” is also wider than “Huawei”: iFlytek works with Hygon and Cambricon, two other Chinese AI chip designers. After the launch Liu told reporters that more progress would come in the first quarter of next year.

Huawei’s own calendar is the context for that plan. In September 2025, Huawei set out its roadmap. The Ascend 950DT is the version built for training. It carries 144GB of memory and 4TB/s of memory bandwidth, from high-bandwidth memory (HBM) of Huawei’s own design, and was due in the fourth quarter of 2026. Its sibling, the 950PR, targets the first stage of inference and recommendation systems and was due in the first quarter. In April 2026, iFlytek told investors that Huawei would mass-produce the 950 after the third quarter. At a media briefing during Huawei Connect in September 2026, Eric Xu, one of Huawei’s rotating chairmen, said 950DT supernodes were still under testing. Supernodes link many chips so they work as one system. He put large-scale supply at the end of 2026 or early 2027.

iFlytek’s August earnings briefing fits that timeline. The company said the 950DT “is mainly used for model training.” It said it “will complete” the replacement of its existing equipment, and that a joint team with Huawei had started work in advance. Those read as the words of a customer preparing for a chip.

iFlytek put the training time at two to three months, and Liu told reporters the work had run for two months. Neither account names the chips. X2.6 has the same parameter count as X2.5, released about 45 days earlier. iFlytek has not said whether X2.6 continued training from X2.5.

In April 2026, iFlytek said plainly that X2-flash, a smaller model, was trained on an Ascend 910B cluster. In October it named no chip at all. The launch did not show whether iFlytek had used the 950, so the hardware part of the April plan remains unconfirmed.

The paid section reads, line by line, the four comparisons that carry the argument, with the Nvidia baseline and the task behind each. It then asks why iFlytek publishes them, and what they leave out.

Read more

Every Sunday I send free readers a short note on what I wrote that week, and open one older paid analysis for a few days. Subscribe free

Two governments’ lines around Chinese AI came into view this week, from opposite sides. As I read them, Washington’s line follows where a model came from, and Beijing’s asks for registration before a model leaves. The buyers kept moving anyway. One overseas company committed to a $200 million floor for tokens served from China, while Zhipu’s Tmall store showed about 300 token units sold since it opened in September.

Washington’s DeepSeek order reaches third-party products.

A Homeland Security order tells US federal civilian agencies to remove DeepSeek products from their IT. It also covers “third-party products and services developed using DeepSeek implementations.” The order is dated July 24 and was filed with the Federal Register on October 19. Its risk summary says “DeepSeek servers are located in the PRC.” I see a data-flow risk written up as an origin rule, one that could reach DeepSeek models running on American servers.

In June I wrote: “The Chinese origin of the weights describes where they were developed. It does not, by itself, describe a live data connection.” Chinese AI Models Are Winning the Cost Layer is open to all readers until Tuesday, October 27, noon Beijing time.

Beijing’s export registry has two names, and a date to watch.

Yuyuan Tantian, a social media account affiliated with state broadcaster CCTV, says that since the policy was published, DeepSeek and Shanghai AI lab MiniMax have completed the registration procedures and completed open-source release and licensed overseas use. The post names no model. Moonshot AI, the Beijing lab behind Kimi, said on October 23 that the full weights of its new K3.1 Pro will be released by November 4. Its launch post does not mention export registration.

My October 19 piece ended on two checks. The second was a disclosed registration at the next open release of a model in the usage top 10. This week answered half of it. November 4 may answer the rest, if K3.1 Pro ranks in that usage top 10.

A Nanning computing firm signed a $200 million token floor.

Wuxiang Cloud Valley, a wholly owned unit of Shenzhen-listed Runjian, has signed a one-year contract to sell AI token services across the border to an overseas company. 36Kr reported the filing, dated October 19. The buyer’s token spending for the year may not fall below $200 million. If usage comes in lower, it pays the difference. The filing withholds the buyer’s name, citing an internal confidentiality procedure, and does not give the per-token price. I read a floor in a contract as saying more about committed demand than a usage ranking does.

My February piece on Token Export argued that China can export AI as tokens, with the compute staying at home.

Shanghai AI lab MiniMax cut its prices before its exam.

MiniMax cut API prices with its new M3.1 Flash model. Input and output prices are each more than 50% below M3’s, The Paper, a Shanghai news site, reported. MiniMax says a Pro version will follow soon. On the morning of October 22, its shares traded below the HK$165 IPO price, with the Hang Seng Tech Index down more than 4% at the same time. One morning of trading says little about a price cut.

My August piece called M3 Pro “the exam.” I treat the coming Pro as that exam. The price cuts arrived first.

Tokens went on Singles’ Day shelves. One store page shows about 300 units since September.

Beijing AI lab Zhipu’s Tmall flagship store shows about 300 units sold in total across all token products since the store opened in early September, Yicai, a Shanghai financial outlet, reported. Tmall’s president said the platform will not put specific discounts into the category for now.

When Tmall opened its token shelf in September, I wrote on X that “the weaker link may have been the channel.” I gave the channel too much weight. My June piece on China’s carriers putting AI on the phone bill asked whether consumers would buy tokens. A store shelf has not answered that yet.

THIS WEEK

• A Top-10 Test for China’s AI Exports: Which Chinese models Beijing’s draft list could reach depends on which ranking it uses.

• BYD and the Captive Ceiling: Why BYD lost the most battery share in China among the top 15.

• China’s Robot Makers Fund Their Own Supply Chain: Startups preparing to list are investing in their suppliers. Their fund filings show three different numbers.

• China’s Payment Push Reaches CATL: How a Beijing-backed payment code shows up in CATL’s third-quarter cash flow.

Next week I publish my analysis of how Pony.ai pays for its robotaxi fleet. Partner-funded fleets are growing, yet Pony.ai is still spending more of its own capital.

The cost-layer piece closes Tuesday, October 27. Paid readers keep it, along with about 250 other analyses, and get my reports at $49 instead of $99.

$150/year, or $20/month.

Upgrade to Paid

Poe

Editor’s Note: FlashPoint is our paid quick-strike column on market-moving China tech events. Today: CATL reported Rmb 21.2bn in third-quarter profit attributable to shareholders. Operating cash flow was about Rmb 1.4bn. Management says part of the gap was its own choice.

CATL, the Chinese battery maker that held 39.4% of the global EV battery market in the first eight months of 2026, reported third-quarter results on October 20. Operating cash flow for the quarter was about Rmb 1.4bn, against Rmb 22.0bn a year earlier. Neither cash figure appears in the filing, which leaves the quarterly line blank. Both are derived from the nine-month and half-year reports. Gross margin fell to 20.69%, from 23.93% in the first half, and the shares fell more than 11% in Shenzhen during trading the next day. I argued in July that CATL collects from both ends of its supply chain, with supplier credit doing more of the work as customer advances fell. This quarter it began closing the supplier end, and it says it chose to.

This piece builds on the July reading of how CATL collects from both ends of its supply chain, now tested against China’s push for faster supplier payments. If that is a useful lens, subscribe free and get every new analysis as it publishes.

The choice has a policy setting. In June 2025, an official at the Ministry of Industry and Information Technology, China’s industry regulator, said tougher competition had stretched payment terms down the EV supply chain. A regulation in force since June 1 that year generally requires large companies to pay small and medium-sized suppliers within 60 days of delivery. Seventeen carmakers then pledged 60 days. On June 29 this year, two industry groups for EV batteries and energy storage issued a supplier-payment code, which the ministry backed in public. It asks battery makers to pay small and medium-sized suppliers in cash and to start the payment clock at delivery or acceptance. Eleven leading battery makers responded, and CATL says it was among the first. Only the first step is a binding rule. The others are public commitments, made with the ministry watching. The paid section measures CATL’s shift, reviews the three July tests, and names the signals outside its filings that will show whether the change holds.

Read more

On October 10, two small Chinese robot component companies announced funding rounds. Juxie Intelligence, a Wuxi maker of robot joint components, closed an A+ round of more than Rmb 100m. Moxian Technology, a Dongguan maker of flexible tactile sensors, closed an industrial round of tens of millions of yuan. X Square Robot, a Shenzhen humanoid robot startup, appeared on both investor lists.

X Square is two and a half years old. It has raised more than Rmb 5bn, is valued at about Rmb 20bn and has no mass-production revenue yet, according to Barron’s China. Reuters in September listed it among the humanoid makers preparing to go public. A company preparing for its own IPO has become an investor in its joint and tactile sensor suppliers.

I read China’s robot primary market in the second half of 2026 as a market that has begun to finance itself. That is a payer my report on the sector did not have on its list two weeks ago. Companies preparing to list are among the sector’s most active investors, putting money into suppliers, data vendors, model startups and each other. The financial funds already in the sector, Hillhouse and Sequoia among them, are still in the same rounds. DeepSeek, the AI lab, holds a small IPO allocation in a robot maker and has reportedly backed two robot model startups. The investing began before IPO scrutiny increased this autumn. I expect slower listings to raise the pressure on these companies to show what their investments bring to the business, and where an investment creates a related-party relationship, transactions with the investee will need closer review.

Hello China Tech publishes new analyses every week on China's AI, chips, robotics, and EVs, built from primary sources rather than second-hand coverage. If reading a robot startup's fund commitment as three separate numbers, what it promised, what it paid and what seat it got, is the kind of reading you want more of, subscribe free and get each one as it publishes.

The Gate Behind the Queue

Unitree Robotics, the Hangzhou company that is the public face of China’s humanoid boom, listed on Shanghai’s STAR Market, the technology board, on August 19. Its IPO price was Rmb 150.80 a share. By September 21 the stock had fallen about 55% from its peak. That day Reuters reported that regulators had used window guidance, informal instructions to investment banks and investment firms, to hold back some humanoid listings. One source called humanoid IPOs frozen for now. Another said there was no formal ban. The regulator did not comment, and no published directive was cited.

What has been published is the Hong Kong rulebook that most of the queue would have to pass. About 40 to 50 robot and embodied AI companies are among the more than 500 applicants waiting outside the Hong Kong exchange, by Barron’s China’s count. The route for a loss-making robot company is Chapter 18C, the exchange’s channel for specialist technology companies. A commercial company under 18C needs at least HK$250m of revenue in its latest audited year and a market value of HK$4bn at listing. Commercial here means revenue, not profit. A pre-commercial company needs HK$8bn. Those two floors are a temporary reduction from HK$6bn and HK$10bn. The reduction runs from September 1, 2024 to August 31, 2027 and covers applications submitted by that date. A commercial applicant must also show research and development spending of at least 15% of total operating expenditure, in two of the three years before listing and across the three combined.

The thresholds in the current rulebook are the ones set in September 2024. What changed is the secondary market’s reaction to the first big humanoid listing, and the regulator’s reading of that reaction. In China the published rule is one gate and the regulator’s reading of the secondary market is another, and the second has narrowed this autumn. Some private robot projects have already taken valuation cuts of 30% to 50%, according to Leo Wang of Qianchuang Capital, a venture investor. First-half 2026 fundraising in embodied AI topped Rmb 90bn (about $13bn), five times a year earlier, with more than twenty companies valued above Rmb 10bn, by Pedaily’s count. I read most of that money as priced against an IPO that has since moved further off.

In Who Pays While China’s Robots Learn, the report I published this month, I counted three payers for the wait before a general-purpose robot can pay for itself: private capital, public markets and the state. The deals of October 10 point to a fourth, which the report did not count because the documents did not yet exist. Robot companies are paying for each other’s wait. What follows is the record of that fourth payer, and it goes on the report’s Watchlist.

The paid section maps twelve disclosed deals by layer and compares three fund arrangements, separating commitments, payments and decision rights. It then reviews what is disclosed about DeepSeek’s robot investments.

Read more

Editor’s Note: FlashPoint is our paid quick-strike column on market-moving China tech events. Today: BYD’s share of China’s battery market in the first nine months was more than four points below a year earlier. In September 2025, I compared what CATL and China’s carmakers each earn. In September 2026, I traced where BYD’s price-war costs land in its accounts. Now the battery data adds a view from the supplier side.

China’s power battery installations grew 16.7% in the first nine months of 2026, to 576.3 GWh. The figures come from the China Automotive Battery Innovation Alliance, the industry body that counts installations from vehicle production data. BYD, which makes both electric cars and the batteries inside them, went the other way. Its installations fell 4.3% to 106.44 GWh. Its share dropped 4.04 percentage points to 18.53%, the largest fall among the top 15 suppliers. CATL, the market leader, gained 1.41 points to 44.16%.

BYD still ranks second. I read the fall as a car problem that shows up in battery data. Most BYD batteries go into BYD cars. Call the limit this creates the Captive Ceiling. A parts business that sells mainly to its parent is bounded by the parent’s products: how many it sells, and how much of the part each one uses. Outside buyers are the way past that limit.

If reading a battery maker’s market share as a measure of its parent’s car sales is new to you, this is a preview of what Hello China Tech does every week: reading China’s AI, chip, robotics, and EV sectors from primary sources most English coverage never touches. Subscribe free to get every new analysis as it publishes.

The limit mattered less while BYD’s car sales rose. In a price war, making its own cells can keep the supplier’s margin inside the group. It also ties battery volume to one order book. BYD sold 3.13m new-energy vehicles, or NEVs, in the first nine months, down 3.94%. China’s NEV category covers battery-electric, plug-in hybrid and fuel-cell vehicles. The home market shrank. Figures from the China Association of Automobile Manufacturers, published by the industry ministry, put national NEV sales at 12.47m and exports at 4.01m, more than double a year earlier. Subtracting exports puts sales inside China down about 10.6%. CATL sells to many carmakers, which gives it more sources of demand as the market grows. The paid section sets out what BYD’s production and battery figures can and cannot show, then compares its overseas growth with CATL and other Chinese battery makers.

Read more

On October 16, China's Ministry of Commerce and Ministry of Science and Technology published Announcement No. 48. It proposes adding 40 technologies to the Catalogue of Technologies Prohibited or Restricted from Export, one of China's three main export-control lists and the one that covers technology transfer. Twenty-three of the 40 are AI technologies. Public comments close on November 15. Until the final list takes effect, companies must register before they export any listed technology.

Reuters reported that models, agents and computing clusters are on the list. Its report does not give the thresholds in Attachment 1, which sets them out. For large models, the restricted entry covers weights, checkpoints, source code, architecture, and training and inference techniques. Its second item reaches that core technology if a model meets any one of three tests. Its first item covers models trained on China's "important data" and then used to provide services, and it lists exclusions. A prohibited entry uses the same wording for "core data." Both are categories under Chinese data-security law. This article focuses on the three tests.

The first test is training compute of at least 10²⁶ floating-point operations. The second is at least 2 trillion parameters. The third is "globally leading capability." A footnote defines that phrase. It means a top-10 global rank, at the time of the export, in public assessments by mainstream evaluators or large industry platforms. The assessments can cover overall performance, specialized capability, global usage volume, safety or training efficiency.

The footnote does not say which ranking counts.

I read this as the line that decides how far the rule reaches below the largest models. The parameter test already covers models above 2 trillion parameters. The third test depends on rankings at the time of export. On the capability boards I checked, no Chinese model is in the top 10. On OpenRouter's usage board, six are. These counts show how far apart the boards are. They do not settle how many models the draft would restrict. Each later version would need its own assessment.

If reading an export rule through the leaderboard it depends on is new to you, this is a preview of what Hello China Tech does every week: reading China's AI, chip, robotics, and EV sectors from primary sources most English coverage never touches. Subscribe free to get every new analysis as it publishes.

The compute and parameter tests

The compute test matches an American one. The same number, 10²⁶ operations, was the threshold in ECCN 4E091, the model-weights control in the AI Diffusion Rule the US Commerce Department published in January 2025. The US rule also excluded weights that had already been published, as Beijing's draft does. Washington announced it would rescind that rule on May 12, 2025. Beijing's draft sets the same compute threshold.

I found no public record of a Chinese model above it. Epoch AI, a research group that tracks training runs, estimates Kimi K3, from Moonshot AI, at 2 × 10²⁵ floating-point operations. That is a fifth of the line.

The parameter test has already been crossed. Kimi K3 has 2.8 trillion parameters. Alibaba's Qwen3.8 has 2.4 trillion. Both labs published the weights. The draft excludes technology that has entered the public domain through open release or academic publication. It does not define that term. Kimi K3's weights are released under Moonshot's own licence, and whether a licensed release counts is open. If unpublished checkpoints or training methods meet the entry's conditions, their export would remain subject to the proposed controls.

At least one large model meets the parameter test without public weights. Baidu's ERNIE 5.0 has 2.4 trillion parameters. I found no published weights for it.

The ministry's question-and-answer note sets out what is not an export. Once the list takes effect, using a restricted model to offer API access, app and web use, SaaS or lightweight on-device deployment is not a large-model technology export. Data-export rules remain separate. The cyberspace regulator says overseas access to data stored in China can count as a data export.

That leaves the ranking test, and a different question: how far the draft reaches below the largest models.

The paid section compares three types of ranking and gives a three-part guide to the transactions the draft covers.

Read more

Every Sunday I send free readers a short note on what I wrote that week, and open one older paid analysis for a few days. Subscribe free

Most of what I followed in China tech this week was about who gets to say no. Lenders said yes to Tencent while its shares kept sliding. China’s biggest apps sit on ByteDance’s own list of those declining its agent. A robot maker pushed its own prospectus into next year.

Tencent borrowed $5.7bn. Its shares are down 29% this year.

Tencent raised about $5.7 billion in US dollar and offshore yuan bonds on October 13, Reuters reported, with orders more than four times the final size. The shares are down about 29% so far this year. The way I read the two prices, lenders are paying for the business that exists, and shareholders are still discounting a capital expenditure bill that more than doubled in the first half. In August I wrote that analysts confirmed the core business was healthy, then kept charging the AI spending against the group. The bond market just did the first half.

On September 30 I set a test for October 15. It came back a day early.

ByteDance’s Doubao phone assistant said on October 14 that it would gradually open screen automation in third-party apps, The Paper, a Shanghai news site, reported. Its list of apps that, it says, have declared refusal was updated that day. WeChat, Taobao and Alipay are on it, with banks, the three carriers, DeepSeek and Kimi. By my count, more than 150 apps.

In September I wrote that the app, not the phone, owns the agent’s work, and that I expected WeChat, Taobao and Alipay to refuse or be blocked. The list runs far wider than the two super apps I built that piece around.

AgiBot is raising again, and its IPO slips past 2026, an investor says.

AgiBot, the Shanghai humanoid maker, has started a new round of about Rmb 1bn to Rmb 1.5bn at a valuation of about Rmb 50bn, Caijing, a Chinese business magazine, reported on October 13, citing an investor. Hong Kong rules, as Caijing describes them, add a 120-day wait after new pre-listing financing, absent a waiver. The investor told Caijing the round means AgiBot cannot complete its IPO in 2026. AgiBot told Caijing it had no comment for now.

In this month’s Founding Letter I wrote that if the quarter ends without AgiBot’s filing, I will read China’s robot listing queue as slowing. It is heading that way.

In August I wrote that only mandatory disclosure makes a robot company give up the right to pick its numbers. An investor now says the listing that would do that will not finish in 2026. The Folklore Ledger of China’s Robot Boom is open to all readers until Tuesday, October 20, noon Beijing time.

Beijing named the problem my Ulanqab piece is about. It did not define it.

On October 16, China’s State Council, chaired by Premier Li Qiang, heard a progress report on its “AI Plus” campaign. Xinhua’s readout lists “coordinating computing power and electricity” as one step to strengthen AI’s foundation. It does not say what the term covers, and it gives no figures. In my October 15 piece on Ulanqab, the term has an address: a finished campus, and a 500kV substation and line that, as of a September 3 report, were still being built.

“Half of China’s compute makes short dramas.” So went the headline.

The claim went around Chinese social media this week. The source is a June roadshow record from COL Group, a Shenzhen-listed short drama company, reported by Caiwen News at the time. It told investors that AI dramas and video take 55% of China’s token consumption, e-commerce and marketing 24%, software development 15%. The record names no source.

In April I wrote that tokens are not fungible. A video token is not a coding token. A share of tokens is not a share of spending, and it is not a share of compute.

THIS WEEK

• China’s Parallel AI Stack: My second report. What training on domestic chips costs, chapter by chapter.

• In Ulanqab, Operators Pay for the Wait: A finished campus in Inner Mongolia, and the costs that start before the meter does.

• Manus and the Surviving Shape: What survived Beijing’s block of the Meta sale, and who may copy it.

• DeepSeek Fills Its Cap Table at July’s Price: The round doubled, the price did not. Who got in, at what fee.

• The Second Seat in China’s Consumer AI: Free to read. How China’s chatbot rankings moved in a year, and what Alibaba spent.

Next week I publish my analysis of Beijing’s proposal to put large AI models on its export-restriction list. One test is a top-10 global rank. The draft does not say which ranking counts, and that choice decides which Chinese models it reaches.

The Folklore Ledger closes Tuesday, October 20. Paid readers keep it, along with more than 240 other analyses, and get my reports at $49 instead of $99.

$150/year, or $20/month.

Upgrade to Paid

Poe

Doubao, ByteDance’s AI assistant, led China’s AI-native apps in August 2025 and leads them still. Qwen replaced DeepSeek in second place.

In August 2025, 50.1% of Chinese users of AI-native apps opened DeepSeek at least once that month. In August 2026, 22.8% did. Over the same twelve months, Qwen, Alibaba’s AI assistant app (formerly Tongyi), went from 1.1% to 29.4%. Doubao rose from 54.9% to 77.2%.

The figures come from QuestMobile’s 2026 report on AI users and marketing, carried by 36Kr. QuestMobile is a Chinese mobile-internet data firm. It calls the measure active penetration: the share of all AI-native app users who opened a given app in the month.

The base grew fast. QuestMobile counts 537m monthly users of AI-native apps in August 2026, up from 287m a year earlier. Multiplying each app’s share by that base gives a rough user count. The arithmetic is mine, not QuestMobile’s. By this calculation, DeepSeek fell from about 144m monthly users to about 122m while the market nearly doubled. Qwen rose from about 3m to about 158m.

I read this as a ranking set by distribution and paid reach, with model reputation less decisive. Assistant markets elsewhere face the same question: does the best model or the widest distribution win the daily habit? This report measures neither directly. The ranking is consistent with a distribution advantage.

Reading an AI app's rank by where its users settle after a promotion ends, and by its share of ad spending, is one example of how Hello China Tech reads China's tech sector: primary sources first, new analyses every week, across AI, chips, robotics, and EVs. Subscribe free to get the next one in your inbox.

The seats are not exclusive. Users who open only one AI app fell from 36.7% of the base to 25.4%. Those who use three or more rose from 32.5% to 44.0%. Second place means reaching more users than most rivals. It does not mean those users rely on Qwen alone.

Over the past year I made six calls about this market. The report lets me revisit all six. Four can be checked against the data. Two remain open or untested. The clearest miss is about Qwen.

How the Seat Changed Hands

DeepSeek, whose app made AI a mass habit in China in February 2025 by QuestMobile’s account, is a model lab. Unlike ByteDance, Alibaba and Tencent, its owner runs no shopping app, short-video feed or messaging network that could send users into its chatbot. When the market nearly doubled, its share was diluted. By my calculation, its user count fell as well. The report does not measure how much traffic any app received from a parent platform. The platform gap here is about ownership. The traffic itself is not measured.

Qwen had two things working for it. The first was an owned transaction layer. In February 2026, Alibaba announced a Rmb 3bn Lunar New Year promotion through Qwen. The mechanism was a Rmb 25 voucher with no minimum spend. A user updated the app, then spent the voucher on milk tea, takeout or holiday shopping through Taobao’s instant-delivery service. Qwen’s daily active users went from under 10m in January to a peak of 73.5m in early February, and the timing ties the jump to the promotion. The February piece cited an earlier daily count of 58.48m. The full January-February chart shows the peak was higher. At the end of February, Qwen was still at about 31m.

Tencent’s Yuanbao ran a similar push with a different tool. Tencent announced Rmb 1bn in digital cash gifts, known as red packets, and users shared them through WeChat and QQ, Tencent’s messaging apps. Yuanbao’s daily users went from under 10m to a peak of 40.5m in mid-February. By the end of the month they were back near where they started, at about 7m.

The second was paid reach that kept running after the holiday. QuestMobile estimates ad spending for three assistants it tracks as typical AI advertisers: Qwen, Doubao and Yuanbao. These are third-party estimates and do not cover all actual spending. The influencer figures cover Douyin, Kuaishou and Xiaohongshu, and exclude commissions and placement fees for livestream deals. In June 2026, Qwen accounted for 77.7% of the three apps’ media ad spending and 79.4% of their influencer spending. June was also the month of university applications after China’s national college entrance exam, and all three apps offered tools for that process. In the last week of June, Qwen had 101m weekly users, against 318m for Doubao and 26m for Yuanbao.

The spending figures do not explain Qwen’s growth on their own. In the week to February 1, Yuanbao outspent Qwen on media ads by more than two to one. Its daily user count later returned close to its January level. In the two weeks to January 25, Qwen outspent both rivals and stayed under 10m daily users. I read Qwen’s rise as a combination of paid reach and a subsidized shopping use case.

The data does not separate the two causes. Some users may have stayed because ordering through Qwen was useful, others because Alibaba kept paying to remind them; by June it was still paying. The peak says little about where a Chinese AI app stands. The level after the push says more. By the end of February, Doubao’s daily user count was about 39m above its January level and Qwen’s roughly 25m higher. Yuanbao was close to its January level. These are changes in daily counts, not retention rates for campaign users.

What the Payment Chart Cannot Say

QuestMobile also reports a payment rate: monthly users who paid for anything inside an app, divided by that app’s monthly users. Among the ten apps it charts, Ant Afu, Ant Group’s health assistant, leads at 16.5%. Doubao is second at 12.4%. DeepSeek shows 7.1%, Qwen 5.9%, Yuanbao 5.7% and Kimi, Moonshot AI’s chatbot, 4.0%. Among users who spend more time than average in AI apps, Doubao’s rate is 19.1%.

The definition matters more than the ranking. QuestMobile counts users with any payment inside an app during the month. It does not say which payments it counts for DeepSeek, whose chatbot is free to use. Either way, the figure cannot be read as a subscription rate. It is also not comparable with the subscription conversion figures in my May piece.

If most of these users paid for transactions, the rate would say little about subscription adoption. ByteDance has opened both routes. In May 2026, Doubao’s App Store listing disclosed paid tiers under test at Rmb 68, Rmb 200 and Rmb 500 a month. From August 10, ByteDance charged hotels a 12% all-in fee on bookings routed through Doubao, according to a company disclosure cited by Yicai. The chart cannot show which of the two the 12.4% reflects.

Six Calls, Graded

Distribution owners win: holds. The November 2025 piece said “the winners will be distribution giants that bolt AI onto existing moats, not standalone innovators.” Both leaders belong to platform companies, and DeepSeek, the independent lab, lost ground. That result is consistent with the call. The report does not measure how much each parent platform contributed, so it supports the call without proving the mechanism. One part of that piece goes untested. It argued that the gap between AI features inside super apps and standalone AI apps would last. This report counts standalone apps only, so it cannot settle that half.

Tencent’s wall: holds. The February 2026 piece said “distribution power without product gravity produces downloads that do not convert to habits.” Yuanbao’s February curve fits that reading: a spike to 40.5m daily users, then a return to about 7m, near its starting level. Its penetration fell from 11.5% to 8.2% over the year.

Yuanbao at 100m: target open, reading withdrawn. The November 2025 piece asked whether Yuanbao could reach 100m monthly users by the end of 2026. By my calculation, it grew from about 33m in August 2025 to about 44m in August 2026. The December deadline is still ahead, but the reading I attached to a miss was wrong. I wrote that failure “would confirm that AI platforms require technical leadership, not just distribution muscle.” The data points the other way. DeepSeek, the lab best known for its model rather than its platform, lost share. The two apps that gained most had platform owners behind them.

User acquisition is over: wrong. In November 2025 I wrote that Qwen’s rebrand “marks the end of user acquisition and the start of execution hell.” I was wrong to call the acquisition phase over. Alibaba kept funding promotions and advertising as Qwen grew. The continued spending suggests paid acquisition is still part of the competition in China’s consumer AI.

Alibaba’s wall is reliability: untested. The February piece named reliability as Alibaba’s constraint, citing an agent that could take 10 million orders in nine hours and still deliver the wrong coffee. This report does not test that claim. Qwen’s user growth does not show whether its orders became more reliable.

ByteDance stays out: too strong. The February piece said ByteDance’s Spring Festival strategy “had nothing to do with” the consumer race between Yuanbao and Qwen. It treated the Gala mainly as a hardware and enterprise play. It also noted that Doubao handed out cash red packets without disclosing a total. Doubao joined late, on February 10. For the night of the Spring Festival Gala, China’s Lunar New Year television show on CCTV, the state broadcaster, it announced tech gifts and cash red packets. QuestMobile also estimates Doubao’s media ad spending at about Rmb 47m in the week to February 1, more than Qwen’s that week; most other weeks were close to zero. Doubao’s daily users were about 75m in January, peaked at 144.5m on Gala night and stood at about 114m at the end of February, the largest rise of the three. The ad spending and the rise in daily users point to a consumer acquisition effort as well. “Nothing to do with” was too strong.

Doubao had more weekly users despite a smaller share of the tracked ad spending. Before any holiday push began, it already had about 75m daily users. In June it accounted for 12.0% of media ad spending and 18.2% of influencer spending among the three tracked apps, and had more than three times Qwen’s weekly users.

The largest revision in this scorecard is to Qwen. Two earlier pieces treated its position as a test of execution. This year it looked more like a test of spending. What I am watching is whether Qwen’s user base holds if Alibaba reduces its ad spending.

The Financial Times reported on October 8 that about 15 gigawatts (GW) of data center capacity had been committed in Ulanqab, a city in Inner Mongolia about 350 kilometers northwest of Beijing. It cited an unnamed executive at a leading operator in the region. Goldman Sachs counted 12.5GW of operating and planned capacity in June, up from 3.3GW in July 2025, according to summaries of its report. The two estimates come from different sources and dates.

The FT framed the limit as chips. It quoted a Jefferies analyst: “The question is whether the data centres can get enough chips to put into the buildings.” X.PIN, a China tech newsletter, argued in September that the grid matters more. Neither piece asks who pays while a finished building waits for its power line.

One operator shows what that wait looks like. On August 18, VNET, one of China’s largest data center operators, told investors it plans to deliver 585 megawatts (MW) of new capacity over the next 12 months. Most of it will come from its Ulanqab campus.

On September 3, Xinhua Finance, the financial news service of China’s state news agency, published remarks by the campus’s chief engineer. The first phase was running, he said, and the second was under construction. “We look forward to the Dongying 500kV project being completed and put into use soon.”

Dongying is a 500kV substation and line project. As of that report, it was still being built.

Data centers in Ulanqab used 3.27 billion kilowatt-hours (kWh) in the first half of 2026, up 89.5% from a year earlier, according to the city’s economic planning agency. That is an average load of about 750MW. It does not measure connected capacity.

These figures do not show how much capacity is waiting for power. Grid notices show that supply infrastructure is still being built. I read the risk in Ulanqab as a cost operators carry before billing starts: GDS’s disclosures show the operator invests before service begins and usually collects no fee until then.

If reading a data center build-out by the date tenants start paying is new to you, this is a preview of what Hello China Tech does every week: reading China’s AI, chip, robotics, and EV sectors from primary sources most English coverage never touches. Subscribe free to get every new analysis as it publishes.

Cheap Power, Slow Connection

Data centers in Ulanqab pay about Rmb 0.358 per kWh, the FT reported. Alibaba Cloud has cited Rmb 0.32 to 0.35. In Beijing, industrial power costs about Rmb 0.80. Inner Mongolia had 117GW of wind capacity by June and roughly 130GW of mostly coal-fired plants, the FT reported. David Fishman of The Lantau Group, an energy consultancy, told the FT the region likely had the “largest electricity local oversupply” of any administrative region in the world.

A large data center needs a substation that steps high-voltage power down for its halls, plus lines to its gate. A cloud industry insider told a Chinese business news site that a conventional data center hall takes about 12 months to build. If new substations and other power works are needed, the same person said, that work can take three to five years.

Cheap power does not mean a fast connection. A substation is a project with land, permits, transformers and line routes. Grid works follow their own schedule.

Ulanqab now runs on two clocks: the dates operators promised customers, and the dates the grid can deliver. Who pays when they do not match depends on when billing starts.

Read more

Since May 2025 I have written about China’s AI chips one listing, one teardown and one export rule at a time. Read one by one, those pieces answer an old question: can Chinese chips train large AI models? By mid-2026 the public record says yes, at least in company claims. This report takes up the harder question that replaced it. What does it cost to run AI on China’s own chips, who is paying, and what decides when the gap closes?

My core argument: China’s parallel AI stack is furthest along in chip design and system assembly, and weakest in memory and advanced wafers. The cost gap is real but unmeasured. What the accounts do show is how the build is financed.

Five findings from the report:

• The cost comparison does not exist yet. I collected the 12 public readings that come closest. None compares the total cost of the same training job on Chinese and Nvidia hardware. The report sets out the four readings a fair comparison would need.

• Sales do not yet pay for the build. In the first half of 2026, six listed Chinese AI chip designers used about Rmb 8.4bn of operating cash and held Rmb 23.93bn of inventory. Cambricon, a Shanghai-listed chip designer, paid suppliers with bank acceptance bills that rose from Rmb 351m to Rmb 3.1bn in six months.

• Two gates shape demand. Nvidia’s latest quarterly filing says its licensed sales to China “were restricted by the PRC government” and that US controls have “effectively foreclosed” it from the market. In September a state security evaluation added Huawei’s Ascend 950 to its list of AI chips that have passed.

• Memory is the scarcest input. Teardowns of two samples of Huawei’s Ascend 910C accelerator found Samsung and SK Hynix memory, according to Bloomberg. In September Reuters reported higher prices for Huawei’s newest training card as memory ran short.

• Some layers need no Chinese version. In optical modules, Chinese firms already supply the global AI build. Dependence there runs both ways.

Where my earlier calls were wrong or are still open, I say so in the chapter where it matters, and name the assumption that failed. The report closes with six directions to watch and a short guide to reading the next chip headline.

Buy the report

Read the free sample (Executive Summary and Chapter 1)

Paid subscribers pay $49 for this report instead of $99. The report draws on the analysis they read each week. If you want the next one as it develops, subscribe below.

Subscribe now

DeepSeek, the Hangzhou AI lab, is nearing completion of its second funding round at about Rmb 100bn (about $15bn), according to deal participants cited by Caijing, the Beijing business magazine. That is double the Rmb 50bn target the round opened with in July. The target doubled in under half a month. The pre-money valuation did not move. It stays at about Rmb 500bn, the figure set when the round opened. DeepSeek did not respond to Caijing’s request for comment. Bloomberg had already reported commitments of at least Rmb 80bn, with a final tally that could approach Rmb 100bn. Reuters reported more than Rmb 80bn and an expected October close. Reuters had earlier reported that DeepSeek hired CITIC Securities to prepare a listing on Shanghai’s STAR Market, the exchange’s board for technology companies. Caijing and Bloomberg both put the listing in 2027. The wrapper on the money, and the fee behind it, have not appeared in English coverage so far.

A round that doubles without a price increase can be read two ways. One is demand: Bloomberg’s sources said investor interest exceeded the plan after DeepSeek’s latest model release. The other is allocation: the company is filling a shareholder list before an IPO filing, at a price it settled in July. I read this as allocation. The evidence sits in three lines of the deal. The first is the price, and it is flat. The paid section sets the other two beside it, names the listing rule behind the investor list, and ends with the line in the IPO filing that will test the shareholder arrangement.

Hello China Tech publishes new analyses every week on China’s AI, chips, robotics, and EVs, built from primary sources rather than second-hand coverage. If reading a pre-IPO round by its fees and shareholder count rather than its headline size is the kind of reading you want more of, subscribe free and get each one as it publishes.

Read more

Two Notices, No Explanation

On April 27, 2026, an office inside the National Development and Reform Commission (NDRC), China’s state planning agency, published a notice one sentence long. The Foreign Investment Security Review Office prohibited the foreign acquisition of the Manus project and ordered the parties to reverse the transaction. The notice left out Meta’s name, any legal provision and any reason.

On October 8, Butterfly Effect, the parent of the AI agent startup Manus, said it had raised more than $500 million. Boyu Capital, a Chinese private equity firm, and IDG Capital, an investment firm known for its US dollar funds, co-led the round. Tencent, the Chinese internet group behind WeChat, joined again, with HSG (formerly Sequoia China) and ZhenFund, a Beijing seed investor. Butterfly Effect did not disclose a valuation. A $4 billion target had been reported in September.

Neither the April notice nor the later documents cited in this piece explain what ownership structure the review required. The April notice made one demand: reverse the deal. It said nothing about penalties or about Manus’s structure. Manus’s own September 1 notice said it had “formally resumed independent operations.”

Over the following months, the company’s owners and plans changed.

What Changed Between the Notices

In 2025, Manus moved offshore. Butterfly Effect registered in Beijing in April 2022. In 2023 it set up entities in Hong Kong, the Cayman Islands and Singapore, Caijing, a Chinese business magazine, reported. The Singapore entity was wholly owned by a Cayman entity. Manus launched in March 2025, won backing from Benchmark, the Silicon Valley venture firm, and moved its headquarters to Singapore by mid-2025. In July 2025, more than 40 of its roughly 120 China staff were expected to move with it, and the rest faced layoffs. In December 2025, Meta agreed to buy the company for more than $2 billion. A Meta spokesperson said there would be “no continuing Chinese ownership interests in Manus AI” and that Manus would stop its services and operations in China. In August, I wrote that among Chinese-founded AI startups rebuilding themselves offshore, Manus “went furthest.”

By October 2026, the picture had changed. Manus’s overseas headquarters is still in Singapore, a person familiar with the company told Lianhe Zaobao, Singapore’s main Chinese-language daily. Benchmark is gone. After the buyback and before the October round, Tencent held about 20% and was the largest single shareholder, Caijing reported in August, citing two people familiar with the deal. They said the founders likely held control. The ownership split after the new round has not been disclosed. The October money, all of it new, came from the two lead investors and three earlier backers. On the same day, Manus listed 17 job openings with Beijing as the work location, the paper added. The company had said in late September that it was building a team for a China product. Reuters reported, citing a source, that a Hong Kong IPO process would not start before 2027.

Manus has now gone through relocation, sale, review, reversal and refinancing. It still operates. Beijing blocked the sale to Meta. None of the public documents cited here says whether regulators approved each part of what followed. I read this as the template that other Chinese-founded AI companies abroad are likely to be measured against.

This piece builds on the Identity Squeeze, a four-layer reading of how Chinese-founded AI companies build an offshore identity. If that is a useful lens, subscribe free and get every new analysis as it publishes.

The Question Already Asked

Two English-language reports read the round as investors staying with Manus. CNBC wrote that the raise suggests investors “aren’t deterred” by Beijing’s order. Bloomberg wrote in September that the deal signaled “growing confidence that Manus is shaking off the aftermath” of the Meta purchase. Both reports also note the change in ownership. Neither sets the new structure against the old one, part by part. That comparison is what this piece adds.

Stefanie Kam of the S. Rajaratnam School of International Studies in Singapore said one question is still open. The model is Chinese capital, a Singapore overseas headquarters and a possible Hong Kong listing. Whether it becomes a template for others, she said, remains to be seen. The paid section uses the case as a checklist.

The paid section compares five parts of Manus’s structure: legal entity, capital, people, product and exit. For each, it sets the 2025 position against the October 2026 position, and separates what is reported from what is planned or not public.

Read more

Every Sunday I send free readers a short note on what I wrote that week, and open one older paid analysis for a few days. Subscribe free

Most of what I published this week asks who pays while something waits. A Beijing AI cloud company signs years of fixed leases, while its customers sign shorter contracts. China’s robot makers are still learning to do paid work, and someone is funding them in the meantime.

What I cannot tell yet is whose patience runs out first: the tenant’s, the landlord’s, or the investor’s.

OPEN UNTIL TUESDAY

In March I wrote that in China’s robot rental market, “every deployed robot effectively arrives with a human engineer.” Rental prices had fallen fast. The people had not gone away.

Before China’s October holiday, Chimelong, a Chinese theme park group, deployed more than 300 robots at its park in Hengqin, next to Macau. About one staff member still assists every three, Huxiu, a Chinese tech news site, reported on October 6. AgiBot, a Shanghai humanoid robot maker, hopes to raise that ratio to 10 to 1, or even 20 to 1.

Robots for Rent, Humans Included is open to all readers until Tuesday, October 13, noon Beijing time.

THIS WEEK

• Who Pays While China’s Robots Learn: My first research report, on China’s embodied AI industry.

• Chinese Carmakers Push Back on Name Rent: Huawei and CATL gave up ground at two Chinese carmakers. What were buyers paying for?

• Who Pays When China’s Compute Sits Idle: A Beijing AI cloud’s biggest customer is also one of its landlords.

• What I got half right this quarter: My first Founding Letter. I graded six of my China tech calls in public.

The South China Morning Post quoted me twice this week, on model fatigue in China’s AI race and on Claude’s new Chinese-language interface.

Next week I publish my second research report. Chinese companies now report AI training runs on domestic chips. The report asks what that costs against Nvidia, and who pays the difference.

I also publish my analysis of Ulanqab, a city in Inner Mongolia where about 15 gigawatts of data center capacity has been committed, the Financial Times reported. Power there is cheap. The question is who pays while a finished building waits for its power line.

The robot-rental piece closes Tuesday, October 13. Paid readers keep it, along with more than 240 other analyses, and get my reports at $49 instead of $99.

Founding members also get the full quarterly letter, including the parts I do not publish, and one report a quarter. This quarter’s is the robotics report.

$150/year, or $20/month. Founding membership is $500/year.

Upgrade to Paid

The Customer Is Also the Landlord

On September 29, 2026, DataCanvas (九章云极), a Beijing AI cloud company, filed for a listing in Hong Kong. In 2025, its largest customer, which the filing calls Customer I, accounted for 48.0% of revenue. DataCanvas had delivered AI computing infrastructure to it. The same filing lists Customer I as one of DataCanvas’s landlords: firms that lease it part of a computing center, servers included.

Customer I was already a leasing supplier in 2024, with Rmb 178.3m of leasing purchases, before DataCanvas recognized Rmb 527.4m of project revenue from it in 2025. The filing shows the overlap, but does not match those amounts to a single facility. Customer group J did both in one year: Rmb 230.9m of project revenue and Rmb 234.9m of leasing purchases in 2025. Together, I and J supplied 69% of 2025 revenue. The leasing figures are purchase amounts, not rent paid: cash paid on all of DataCanvas’s leases in 2024 was Rmb 24.4m.

In a July piece on who owns China’s Token Factories, I described the cleanest version. An asset owner holds the hardware and absorbs depreciation, and an operator runs it for fees. That piece found no disclosed contract confirming that split. DataCanvas’s filing describes a different arrangement. DataCanvas books the project revenue. The customer owns the delivered infrastructure and leases part of it back. For that part, DataCanvas carries the demand risk for the life of the lease, typically five years. I read this as the clean split with the demand risk moved to the operator.

Reading a project customer and a landlord as one counterparty is one example of how Hello China Tech reads China’s tech sector: primary sources first, three analyses a week, across AI, chips, robotics, and EVs. Subscribe free to get the next one in your inbox.

DataCanvas began in 2013 as an enterprise AI software firm and moved into computing infrastructure in 2023. Revenue grew from Rmb 106m in 2023 to Rmb 1.1bn in 2025. The filing describes the loop. DataCanvas delivers AI computing infrastructure to a customer. Its customers include local government bodies and state firms. It then leases a designated part of the site back and runs it as cloud capacity. In the filing’s words (my translation), this secures computing power “without funding the construction of the facilities.” DataCanvas says the sales and leases are negotiated case by case, do not depend on each other, and involve no bundled or back-to-back deals.

A Glut Looking for a Tenant

In July 2025, Reuters reported a glut of computing centers backed by local governments. Four sources put utilization at about 20% to 30%. A project manager said builders expected government bodies and state firms to buy.

An owner with a half-empty center needs a tenant. A cloud operator willing to sign a five-year lease appears to be that tenant. The filing does not show whether the sites DataCanvas leases sat idle before.

By 2026, prices were rising. The filing cites CIC, a consultancy DataCanvas hired. It estimates that average monthly cloud fees per AI-accelerated server rose by about 100% to more than 150% from 2025 to the first half of 2026. The two figures measure different things: center utilization in 2025, and server cloud prices into mid-2026. In a tight market, a long lease looks cheap. A lease can also outlast the prices that made it look cheap. The landlord has a claim to fixed rent, as long as the tenant can pay. The tenant holds the empty racks if demand slips.

The lease terms show who keeps paying when demand weakens, who shares the upside, and whether project profits turn into lasting cloud returns.

Read more

The answer is my core argument. China’s embodied AI industry earns its revenue from businesses that do not need general intelligence: mainly research hardware, safety work and paid performance. Its valuations price the one business that does. Private investors, public markets and the state are paying for the wait. There are two waits: one for robots to earn repeat revenue from productive work, and one for general-purpose capability. The first can end well before the second.

Five findings from the report:

• Research buyers carry the clearest disclosed revenue. Research and education buyers produced 73.6% of Unitree’s humanoid revenue in the first nine months of 2025. Industry applications produced 9%, and less than a third of that was manufacturing, inspection or logistics work.

• The businesses that make money need the least intelligence. DEEP Robotics, a Hangzhou maker of industrial quadruped robots, earns most of its revenue from robots sent into dangerous sites. Even that business runs on the budget cycle of state-owned buyers.

• The state now says it will share the risk.China’s 2026 Government Work Report calls for mechanisms to share the risks of future industries, embodied AI among them. A June 2026 ministerial notice encourages leasing and payment linked to use or utility for state-owned users. Contracts will show how much risk actually moves.

• The wait is getting longer. Unitree’s founder, Wang Xingxing, has moved the earliest year implied by his own estimate for a robotics “ChatGPT moment” out by about two years.

• The end of the first wait will show first in who pays. Watch productive buyers’ share of revenue, repeat orders and cash collection. I expect it to show there before it shows in funding totals or order announcements.

The report also includes a Scorecard. It grades eleven of my earlier calls on this sector against public documents, including forecasts I got wrong and framings I have withdrawn. For each one I write down the assumption that failed. The biggest miss, my 2026 shipment forecast, turned out to be the clearest evidence for this report’s argument.

A Watchlist of twelve checkpoints follows, starting with Unitree’s third-quarter report on October 29.

Buy the report

Read the free sample (Executive Summary and one chapter)

Paid subscribers pay $49 for this report instead of $99. Founding members get it at no cost. If you want the next one as it develops, subscribe below.

Subscribe now

All reports, past and coming, are listed on the Reports page.

Available now

China’s Parallel AI Stack: Built on Chips, Waiting on Memory, 2026 Edition

Chinese companies now report training runs on domestic AI chips at every level of difficulty. This report shows what is known about the cost, who finances the build, and why memory and advanced wafers are likely to set the pace.

Three questions it answers:

1. China can now train large AI models on its own chips. What does the public record show about the cost against Nvidia, and what would a fair comparison need?

2. Who is financing China’s domestic AI chip build: buyers, suppliers and banks, or the stock market?

3. Which inputs decide when the gap narrows, and which layers of the stack do not need a Chinese version at all?

Inside: five chapters, from the training runs on record to the layers where China already supplies the global build. Where my earlier calls were wrong or are still open, the chapter says so. The report closes with six directions to watch and a short guide to reading the next chip headline. The data appendix holds six tables, from chip designers’ first-half accounts to memory and foundry indicators, each figure with its source. 32 pages, data as of October 7, 2026.

It is written for investors who follow Chinese AI chip, memory and foundry stocks or Nvidia’s China exposure, policy researchers who track export controls and China’s purchasing rules for AI compute, and industry readers who need a map of China’s AI compute stack by layer rather than by company.

Buy the report ($99) · Read the free sample: the Executive Summary and Chapter 1

Who Pays While China’s Robots Learn: China’s Embodied AI Industry, 2026 Edition

China’s robot companies earn their money from research labs, safety work and stage shows. Their valuations assume something else. This report shows who pays for the gap, and what evidence would change that picture.

Three questions it answers:

1. Which of China’s robot businesses make money today, and who are their buyers?

2. Who is paying for the general-purpose robot that valuations assume: private investors, capital markets or the state?

3. Which evidence would show that productive demand has arrived, and which would show progress toward general-purpose capability?

Inside: five chapters (One Label, Five Businesses; Who Pays Today; Who Pays for the Wait; How Long the Wait; How You Will Know). A Scorecard grades eleven earlier calls. A Watchlist lists twelve checkpoints, starting with Unitree’s third-quarter report on October 29, 2026. The data appendix holds shipments, funding and two prospectuses side by side. 29 pages, data as of October 6, 2026.

It is written for investors who follow Chinese robotics or Hong Kong and STAR Market listings, policy researchers who track how the Chinese state funds and measures a new industry, and industry readers who need a map of the sector by business rather than by company.

This is the Founding report for the October to December 2026 quarter.

Buy the report ($99) · Read the free sample

What every report includes

• My earlier calls, checked. Where I got something wrong before, the report says so and names the assumption that failed.

• What to watch. Most reports close with three to five things worth following, and why. Annual editions, and this first robotics report, close with a full Scorecard of my earlier calls and a dated Watchlist.

• Data appendix. Each figure with its source and a source level. Where sources disagree, both are shown.

Each report also has a free sample, the Executive Summary and one full chapter, open to everyone.

Pricing

• Anyone: $99 per report.

• Paid subscribers, monthly or annual, and team subscription members: $49 per report. The code is in each report’s launch post, visible when you are signed in.

• Founding members: one report each quarter at no cost, chosen by me.

• Institutional clients: every report, delivered on launch day.

Each PDF carries the buyer’s email address on every page. Payment and delivery run through Payhip. The download page opens as soon as payment clears, and the link also arrives by email.

Coming next

• How China Funds Its Tech Race. Where the money comes from: state funds, listings before profit, and the mechanisms that keep companies raising.

• The token economy. How China prices and sells AI compute, from telecom plans to enterprise quotas. Title to be confirmed.

New reports are announced here when they are scheduled. Subscribers hear first by email.

Hello China Tech is an analysis newsletter on China’s technology industry, written in Beijing for investors, operators and policy researchers. I read the Chinese-language sources every day: IPO prospectuses, government filings, procurement notices, earnings calls and industry press. I publish new deep dives every week. The analysis has been cited by Bloomberg, the Financial Times, Reuters, The Economist, CNN, AFP, WIRED and the South China Morning Post.

This page is the shortest route to deciding whether the newsletter is for you.

Read these first

All five are free in full.

• The China Tech Field Guide. Six principles from a year of essays, each with the evidence that would prove it wrong. Re-scored and re-issued every July. There is a PDF edition for sharing.

• DeepSeek’s Theory of the AI Gap. A four-hour investor meeting with DeepSeek’s founder, published only in Chinese. I reorganized it by theme and checked it against my earlier calls. This is how I work from a primary source that has no English version.

• China’s AI Model Companies Can Only Afford One Bet. Why platform giants can pursue every AI monetization route at once while independent labs can afford exactly one. This is how I build a framework for a whole sector.

• What China’s 470 AI Dramas a Day Reveals About AI Entertainment at Industrial Scale. Production economics, labor, regulation and culture in one story, and why China’s market is a preview of everyone else’s.

• What I Got Wrong About China Tech. One year of paid essays scored against what happened. This is the standard every future essay is held to.

More than 70 full articles are open to everyone. They are listed under Free Reads.

To read by sector instead, use Topics: AI models, platforms, chips, EVs, robotics, and policy and capital.

What I publish

• Deep dives. New ones every week. Long-form analysis built on Chinese primary sources, each with a framework you can reuse. Paid subscribers read every piece in full. Free subscribers get the preview of each, plus selected deep dives unlocked in full.

• FlashPoint. Short, financial-first analysis of a market-moving event, published the same day or the next morning. Paid only. Browse FlashPoint.

• Weekly Recap. A Sunday post that ties the week’s essays into one thread and flags what to watch next. Free to every subscriber.

• Reports. Long reports that rebuild a year of coverage on one sector into a single document, with a data appendix and a stated data cutoff date. Sold separately; paid subscribers pay $49 instead of $99. Two are out now, on China’s AI chip stack and China’s embodied AI industry. See all reports.

What is free, and what paid adds

Free subscribers get

• Previews of every new deep dive

• Selected deep dives unlocked in full

• The Weekly Recap every Sunday

• The China Tech Field Guide and the annual calibration letter, in full

Paid subscribers get

• Every deep dive in full, with new ones every week

• FlashPoint, published while the event is still moving

• The complete archive: more than 200 analyses and every framework

• Research reports at $49 instead of $99

Founding members get

• Everything in paid

• The Framework Index, the reference that collects the named tools in this archive

• One report each quarter, chosen by me, at no cost

• The Founding Letter, four times a year, where my calls get settled in public. Past letters

Teams and institutions. Group subscriptions and institutional access, including every report, are described on the Institutional Access page.

Subscribe now

How I work

Every essay carries a date, so every call can be checked. The Field Guide re-scores the core principles each July. The calibration letter grades individual calls each August. When the analysis is wrong, you read it here first.

I think and outline in Chinese. AI tools help me draft and edit the English. Every topic, argument, number and conclusion is mine, and I check each one against the original Chinese source. The full description is on the About page.

Hi,

Four times a year I open the ledger. What I called right, what I called wrong, what I am watching next. This is the first one. The full letter goes to founding members. The first part, the grades, is open to every paid reader.

Most of the grading this quarter happened inside the articles themselves. Six pieces went back to an earlier call and scored it in print. You have read them. I will not re-grade them here. What I can do, and what a single article cannot, is lay the six grades side by side. They have one shape. The direction held. A second variable was missing.

If scoring a quarter of your own calls against the numbers they predicted is new to you, this is a preview of what Hello China Tech does three times a week: reading China’s AI, chip, robotics, and EV sectors from primary sources most English coverage never touches. Subscribe free to get every new analysis as it publishes.

In April I wrote that Zhipu’s cloud margins could approach 30% within a year. They came in at 24.6%, and management added a term the April equation did not have, which was time. April’s formula had a level. August’s has a slope. In March I set four conditions for MiniMax’s escape from the cost trap. By August one split, two held, and one needed a new label, because I had treated enterprise customers and enterprise delivery economics as one thing. For Alibaba, three of the four indicators I set in March now read positive, and the five markers from July are half settled: internal AI spending got a line, cloud orders from portfolio companies did not. In January I wrote that Baidu’s best assets unlock value only after leaving. Kunlunxin’s reported target went up and Baidu’s market value went down. The subsidiary priced, and the parent did not collect. In September I wrote that the real test for Huawei’s folded Kirin would begin when it reached a mainstream Mate. On October 1 it did, in the Pro Max and the RS only. The base Mate 90 kept last year’s chip. The test has begun and has not passed.

One call was plainly wrong, and I have already said so in print. In November 2025 I wrote that for China’s AI assistants the one viable path to revenue was a cut of the transaction. WeChat’s agent showed that claim was too narrow. An agent can be paid for through the payment and advertising businesses the platform already runs.

Read more

Editor’s note: In December 2025, I described how Huawei turned from auto supplier into platform. In July, I examined how CATL collects from both ends of its supply chain and named the test that would show its pricing power slipping. This piece settles part of that test and qualifies part of the December argument.

On September 7, Li Auto, a Beijing-based maker of premium family SUVs, posted a cut-off time on its WeChat account. Buyers of its flagship MEGA minivan who locked in an order from 3pm that day would get battery cells designed by Li Auto. Earlier buyers would get cells from CATL, the Chinese group that supplies about 40% of the world’s EV batteries.

Eight days later, Huawei and Seres, a Chongqing carmaker, changed how they run Aito, the best-selling brand in Huawei’s car alliance. Seres now leads product definition, design, marketing, retail and service. Huawei’s role is to “participate and enable.” On September 30, the two signed a new five-year agreement on those terms.

Huawei and CATL are among China’s most powerful auto suppliers. Each spent years making sure car buyers knew its name. Huawei sold Aito cars in its own stores. CATL ran airport ads telling shoppers to pick an EV by its battery. In the same month, each gave up ground at one of its largest customers, though not the same ground.

Huawei remains Aito’s supplier for its Qiankun driver-assistance system and Harmony cockpit software. It handed Seres the lead on retail and product decisions. CATL lost more at Li Auto. On several new models, Li Auto now specifies the cells and orders them from Sunwoda and CALB, two Chinese battery makers.

I read this as Seres and Li Auto taking control of what the buyer sees and what goes into the car. That control may also let them challenge the premium a supplier earns from its name. They still buy the parts that are hard to make, though not always from the same supplier. Call that premium Name Rent: the extra a supplier can charge because buyers know its brand, stores or logo. Neither supplier discloses how large it is. When carmaker margins fall, it becomes an early target for the largest customers.

If separating the premium a supplier earns from its name from the price of the part itself is new to you, this is a preview of what Hello China Tech does three times a week: reading China’s AI, chip, robotics, and EV sectors from primary sources most English coverage never touches. Subscribe free to get every new analysis as it publishes.

Western coverage has mostly asked whether CATL will lose market share. I do not accept that framing, because share alone misses what is at stake. CATL still held 41.45% of China’s power battery installations in August, down from 42.33% in July. In China’s passenger-car battery installations, its first-half share rose 5.6 percentage points from a year earlier to 46.7%, according to its half-year report. At Li Auto, CATL is losing volume. Across the market, the open question is pricing power, and the first-half accounts do not yet isolate it.

The two suppliers put their names in front of buyers in different ways. What follows looks at how each arrangement works, why the pressure falls on them now, and why buyers may decide the outcome.

Read more