On September 7 in Guangzhou, Huawei launched a tri-fold phone called the Mate XT 2 and did something it had not done at a flagship event since 2020. It named a new Kirin processor on stage and explained how the chip was built.

Naming it was the smaller change. Huawei had already identified last year’s tri-fold chip as the Kirin 9020. What it added this year was the method. The Kirin 9050 Pro is, in Huawei’s description, the first mobile system-on-chip built with what it calls LogicFolding, its logic laid out across two stacked tiers of silicon. Huawei’s own image is a building: take a single-storey floorplan, make it two storeys, add lifts. The lifts are not free. There are about 50 million of them, and only 10% to 15% carry signals.

The phone starts at Rmb 19,999, about $2,900, and goes on sale on September 12.

A new manufacturing method has arrived in a phone category that ships in the hundreds of thousands a year. Huawei remains the only source of density and performance measurements for the chip. It also arrives with a silicon bill. A folded chip draws die area from two bonded wafers rather than one, and about 30% more silicon per transistor. That arithmetic is the main thing standing between LogicFolding and a phone that ships in the millions.

Reading a density figure against the silicon it is actually measured on 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.

What the density number counts

The case is made in a position paper by He Tingbo, who runs Huawei’s semiconductor business, published online in July by Science China Information Sciences. It went through peer review; its measurements have not been reproduced outside Huawei. Set against last year’s Kirin 9030 Pro, transistor density rose from 155 to 238 million transistors per square millimetre. Huawei calls that a 55% step; the two figures work out to 54%.

Both chips are built on the same mature process node. That clause matters more than the percentage. In this context “mature” appears to mean the foundry’s most advanced production line, not a legacy node. A density step that size usually means a better manufacturing generation, and Huawei says it did not get one. If that holds, the layout is what is left, and Huawei credits its three-dimensional reorganisation.

The 238 is measured against projected area. SemiAnalysis describes Huawei as measuring density per package footprint, with two tiers counted under each square millimetre. Counted against silicon consumed, assuming equal tiers, density falls by roughly a quarter.

Huawei has not named its foundry or its process. TechInsights, a chip teardown firm, identified last year’s Kirin 9030 as a product of SMIC’s most advanced line. No independent teardown of the new chip has been published. The paper never uses the retail name either, calling the chip Kirin 2026 throughout; matching that to the 9050 Pro named on stage is an inference.

Asked about the method in May, Jensen Huang told reporters in Taipei that it was “a breakthrough for Huawei, but it’s not a threat for TSMC,” noting that TSMC has stacked dies for close to a decade. Both halves hold. Stacking is old. Stacking two tiers of a phone processor’s logic and shipping it is not.

Huawei published the density gain and the engineering behind it. What it did not publish is the cost, and that number is what limits how far the method travels.

Read more

Two Documents, One Day

On September 2, Tencent opened WorkBuddy, its AI office agent, to outside developers and hardware makers. More than 100 partners joined, covering HR software, tax and accounting, business intelligence, marketing and securities data, plus glasses, recorders and earphones. English-language outlets covered it as an ecosystem story.

Tencent Cloud published a second document the same day. It sets out who may resell WorkBuddy.

The general bar for a reseller in Tencent Cloud’s partner system is registered capital of at least Rmb 1m, one year of trading, three full-time salespeople and a Rmb 10,000 deposit. WorkBuddy adds a layer of its own. At least three staff must complete a WorkBuddy product-sales exam, and at least three must pass a second certification named in the document as the Tencent WorkBuddy Efficiency Agent OPC Practitioner. The document does not require the two groups to be different. The partner posts another Rmb 50,000, taking the deposit to Rmb 60,000, about $8,800, and signs a confidentiality undertaking bearing the company seal.

No other product in that document carries its own bar. Exams for TDSQL, Tencent’s distributed database, and TencentOS, its server operating system, appear only as substitutes for the generic sales exam. WorkBuddy’s two are additional.

The consumer product went into closed testing in February. A separate enterprise edition launched on June 5. Three months later it has a professional certification.

The Machine Is Easier to See Than the Market

Tencent has gone furthest in making a product-specific channel visible. Rivals are moving toward enterprise distribution through different structures. Baidu launched a partner alliance for its DuMate agent in July, with tracks for resellers, software partners and delivery partners, promising industry-leading commissions without naming a rate. Alibaba sells QwenWork through Alibaba Cloud’s procurement structure. ByteDance moved the sales arm of Feishu, its collaboration suite, into its cloud unit Volcano Engine and folded the Feishu product team into Doubao.

The state has been building admission rules in parallel. Analysys, a Beijing research house, identified a cluster of second-quarter policy measures covering tiered governance of agents, anthropomorphic AI services, and national standards for agent interconnection.

All of this arrived quickly, and on a thin base of commercial evidence. Analysys counted about 60.6m interactions in June across the seventeen platforms it lists, with WorkBuddy at 20.97m.

Enterprise software usually earns its channel in the other order. A vendor proves that customers renew, then recruits resellers to scale what works. Here the infrastructure for selling WorkBuddy is more visible than the evidence needed to measure what it sells. Customers and usage exist. No company disclosure carries a reliable number that lets an outsider size paid demand or measure renewal.

There is a reason the channel comes first. A Tencent Research Institute survey of 2,012 online respondents, fielded in March and published in May, found 9.8% had paid for a large-model product. The sample skews young and better educated, which may flatter the number. Caijing, a Beijing financial magazine, polled 417 technology workers in late August in a directed sample. Among those who picked WorkBuddy as their one indispensable agent, 79.7% worked outside product and engineering, the people least able to expense a tool. A Chinese software founder put it plainly: no company here buys a Rmb 20,000 enterprise product without a salesperson in the room. Individual plans do sell. But a payment rate that low pushes Tencent toward employers, and enterprise customers in China are reached through direct sales and on-site work.

Tencent’s own second-quarter release shows the gap. It describes WorkBuddy as achieving rapid user growth and healthy retention, with strong willingness to pay through subscriptions and token top-ups. Three adjectives, and no number attached to the product anywhere. The one quantitative anchor is a footnote naming Analysys and defining its metric as monthly desktop visits in June.

Tencent has published a great deal about how WorkBuddy works and how it will be sold, and little that lets an investor measure demand. Its pricing page comes closest to showing the commercial design.

Hello China Tech publishes three analyses a week on China’s AI, chips, robotics, and EVs, built from primary sources rather than second-hand coverage. If reading a vendor’s public price list as its most complete financial disclosure is the kind of reading you want more of, subscribe free and get each one as it publishes.

Read more

By early September, the seven listed or listing AI compute chip designers tracked here had reported, or in Enflame’s case forecast, their first-half 2026 numbers. Combined revenue reached about Rmb 21.5bn, close to double a year earlier. Only two earned a profit from their core business. One is Hygon, a Shanghai-listed CPU and accelerator maker. The other is Cambricon, the AI chip designer whose market value passed Rmb 1tn last year. Two more, MetaX and Iluvatar CoreX, GPU designers that listed within the past year, reported profits that came almost entirely from paper gains on shares in another company. Core profit here means profit excluding non-recurring items for the mainland-listed companies, and reported profit less fair-value gains for the Hong Kong pair. The two measures agree in direction, not in detail.

The balance sheet points the other way. The six companies that disclose inventory held Rmb 23.9bn of it at the end of June, about $3.5bn at today’s rate. That is more than the seven companies’ combined revenue for the half. Inventory has to be financed before it becomes revenue. The income statement does not say by whom.

The five companies that publish operating cash flow burned a combined Rmb 6.4bn in the half. Only Cambricon reported a positive figure. The obvious reading is that Cambricon has reached self-funding and the rest have not. The filings support a narrower reading. None of the filings shows sales cash covering the inventory build. The filings point instead to four sources: supplier credit backed by banks, customer advances paid last year, bank loans, and fresh equity. Which source carries the load is the line to check first when the next disclosures arrive.

This piece extends “The Perpetual Raise,” which asked whether China’s AI chip companies could turn fresh capital into operating cash. If that is a useful lens, subscribe free and get every new analysis as it publishes.

Why the money sits upstream

Cambricon’s inventory note is the clearest map of where a fabless company’s cash goes. Foundry wafers, memory, substrates and packaging are paid for months before a finished card ships, and in a tight market the payment moves earlier still. Of Rmb 9.0bn in gross inventory, Rmb 5.7bn was raw materials and Rmb 2.5bn was materials held by contract manufacturers. Finished goods and goods shipped but not yet accepted came to Rmb 0.7bn, about 8% of the total. The money is locked in wafers and work in progress, not in products waiting for a customer to sign off. Cambricon’s advances to suppliers rose from Rmb 0.7bn to Rmb 2.9bn in six months, and two suppliers received 89% of them. The filing does not name them.

Paying ahead is common across the industry. Nvidia’s latest filing lists $279bn of supply commitments. The difference is scale relative to sales. Using average inventory and annualised cost of sales, Cambricon carries roughly 445 days of inventory and Moore Threads about 590, against roughly 108 at Nvidia. The Chinese figures mix chips with cluster projects, but the order of magnitude holds. The other difference is direction. Nvidia’s operating cash flow fell to 40% of net income in the quarter to late July after it extended payment terms of up to a year to large customers. Nvidia is financing its customers. China’s chip designers are relying on suppliers, banks and equity investors while they finance their own inventory.

Someone has to fund that gap, and the group found more than one answer. Among the five companies that disclose operating cash flow, the one positive figure is also the one most visibly built on somebody else’s money.

What follows traces where each company’s inventory money came from and which balance-sheet line moved to cover it. Each line has its own next disclosure date.

Read more

Editor’s note: This article extends Hello China Tech’s coverage of China’s EV supply chain. In September 2025, I found that CATL earned almost as much in net profit as China’s three most profitable listed carmakers combined. In July this year, I traced how CATL funds part of its cash cycle with credit from its own suppliers. Those pieces followed the company collecting. This one follows BYD, the carmaker paying, and where that cost shows up in its accounts.

On August 28, BYD published its half-year results. English coverage settled on two lines within hours. Overseas revenue passed half of the total for the first time. Quarterly profit rose for the first time in five quarters.

Both hold up. Revenue for the six months fell 7.1% to Rmb 344.8bn. Net profit attributable to shareholders fell 20.5% to Rmb 12.3bn. Overseas revenue reached Rmb 181.3bn, or 52.6% of the total, while domestic revenue fell 30.7% to Rmb 163.5bn. Second-quarter net profit rose 29.7% from a year earlier.

Three other lines in the same document drew no English coverage in the week after the release. R&D charged to the income statement fell 21.3%. Its three payables lines, a broad proxy for operating credit, slipped from 52.6% to 50.3% of total liabilities. And note 4, which splits revenue and cost by region and by business, shows BYD’s overseas automobile-related business earning a gross margin 11.4 points above the domestic one.

Read together, those lines point to a shift inside the accounts rather than an export story. The pressures behind BYD’s price war did not ease. They changed where and when they appear. Overseas auto-related sales are offsetting lost domestic gross profit. Supplier credit is moving from trade payables and Dilian towards bills and formal debt. More R&D is being capitalised, cutting the expense booked today. One is operating, one is financing, one is accounting timing. Together they change how the pressure is absorbed and where it lands in the accounts.

Three Lines Behind the Rebound

Subtracting the first quarter from the half-year isolates the second. Gross profit rose Rmb 4.06bn against the same quarter of 2025. R&D expense fell Rmb 3.41bn. Finance costs moved the other way, rising Rmb 4.34bn. Operating profit finished Rmb 2.35bn higher, and net profit attributable to shareholders Rmb 1.89bn higher.

The cut in R&D expense alone is larger than the entire increase in net profit. Hold expensed R&D at last year’s level, with every other line unchanged, and second-quarter operating profit would have fallen rather than risen.

The rebound is not hollow. That gross profit gain came as revenue fell, and the half-year gross margin improved to 18.85% from 18.01%. Strip out the R&D and finance-cost changes and operating profit still gained about Rmb 3.27bn, though that residual also carries grants, investment income and impairments.

Three lines shaped the quarter. Gross profit was the only one tied to sales and production. The other two came from expense recognition and finance costs. Management’s commentary covers the fall. It cites “the decrease in the new energy vehicle business and foreign exchange losses from changes in foreign exchange rates.” It explains the higher gross margin too, crediting growth in the overseas vehicle business. On the R&D line it says nothing.

For paid subscribers: BYD’s nine-year R&D capitalisation pattern, and where supplier credit went after the 60-day payment campaign. Also why a 28.4% gross margin in BYD’s overseas automobile-related segment still stops short of net profit.

If separating a quarterly profit rebound into the operating part and the accounting part 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.

Read more

I have followed China’s two listed AI model companies since their December 2025 prospectuses, beginning with “Running Out of Runway” and “The First AI Company to Show Its Cards.”

Zhipu, the Tsinghua University-backed lab that listed in Hong Kong in January, reported first-half results on August 31. Revenue rose 399.7% to Rmb 953.9m ($142m). Bloomberg reported that the figure came in about 30% below the average analyst estimate. Hours later, on the earnings call, management said its revenue run rate, August revenue multiplied by twelve, had reached $1.6bn at the end of August, against $1bn in July.

Both statements are accurate. One reports six months of recognized revenue. The other annualizes a single month.

The accounts close on June 30. The operating disclosures run to August 31. Most of them use the start of the year as their baseline, so the 101% price rise and the 40-fold token growth span both periods. What belongs only to the later window is the run rate and two model launches. Investors were asked to price June accounts against August operating metrics. The fresher numbers look faster. They are also less comparable.

In April I wrote that Zhipu had reduced its business logic to one equation: AGI commercial value equals the intelligence ceiling times token consumption scale. I also wrote that cloud gross margins could approach 30% within a year. They came in at 24.6%, up from -0.4% a year earlier.

The equation survived its first test. On the same call, management added a term the April version did not contain.

The Numbers

The revenue mix inverted inside a single reporting period. Cloud-based open-platform and API revenue reached Rmb 825.2m, up 2,735.7%, moving from 15.2% of the total a year ago to 86.5%. On-premise deployment fell 20.5% to Rmb 128.7m. The filing describes part of that decline as deliberate, reflecting a narrower approach to software licensing and a shift toward higher-value industries.

Two margin numbers moved in opposite directions. Blended gross margin fell from 50.0% to 26.4%, because the faster-growing segment is the lower-margin one. The API segment itself went from -0.4% to 24.6%. The company credits scale, higher pricing, and engineering work that lowered inference cost.

The loss line needs the two-column reading I used for MiniMax last week. Reported loss narrowed 12.1% to Rmb 2.07bn, largely because last year carried a Rmb 429.3m charge on instruments that converted at listing. Adjusted net loss, closer to operating reality, widened 12.1% to Rmb 1.96bn. Operating loss widened 13.0%.

R&D reached Rmb 2.13bn and grew 33.6%. Revenue grew nearly twelve times as fast as R&D. R&D has fallen from 7 times revenue in 2024 to 4.4 times in 2025 and 2.2 times now. Sales and marketing fell 14.8% and administrative expenses 44.2%. Growth did not require higher reported selling costs, though thin margins leave room for subsidy inside the product itself. The revenue line missed. The adjusted loss came in below a Rmb 2.75bn consensus estimate cited by Smartkarma, an analyst research platform. The stock fell 1.34% on September 1 and a further 5.26% on September 2.

The first-half numbers support part of the case I made in April. They also show its limit. Model quality may explain why customers pay more today. It does not explain how Zhipu keeps them paying when the next model catches up.

Reading a set of half-year accounts against operating metrics that close two months later 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.

Read more

This is the first installment of Origins, a Hello China Tech series. Every China tech headline rests on a historical premise that has never been fully told in English. Over the past year, I have written about why WeChat became the gatekeeper for China’s AI agents, about the structural power that makes platform blocking possible, about why ByteDance launched an agent no platform would approve. They rest on the same foundation. This is that story.

If the connection between a 2010 desktop software war and today’s AI agent gatekeeping is new to you, this is what Hello China Tech covers three times a week: China’s AI, chip, robotics, and EV sectors, read from primary sources most English coverage never touches. Subscribe free to get every new analysis as it publishes.

An Extremely Difficult Decision

Shortly after 6 pm on November 3, 2010, QQ users across China began seeing a pop-up window on their screens. The message, titled “A Letter to All QQ Users,” opened with a line that would become one of the most quoted sentences in Chinese internet history:

“When you read this letter, we will have just made a very difficult decision.”

The decision: until Qihoo 360, the security software company, stopped what Tencent called unauthorized interference with QQ and malicious attacks on its reputation, Tencent would cease running QQ on any computer with 360 installed. Users had two options: uninstall 360 and keep QQ, or keep 360 and lose the messaging service that connected them to nearly everyone they knew.

By the end of 2010, Tencent reported 647.6 million active QQ accounts, with a peak of 127.5 million simultaneous users online. The word “accounts” matters; many people held multiple. But roughly 20 million users initially lost normal access, and about 5 million were still affected a week later.

According to Wu Xiaobo’s company history, CEO Ma Huateng made the final call that morning at Tencent’s Shenzhen headquarters. Dowson Tong recalled to Wu that the original pop-up showed two uninstall buttons of different sizes. Ma ordered them made equal, insisting on a “fair choice.” Fair or not, the choice was unprecedented. In effect, a corporate dispute had become a loyalty test imposed on hundreds of millions of users.

That night set a rule that still governs China’s internet. The full genealogy remains largely absent from English-language coverage.

China’s Two Walls

Why can’t you open a Taobao link inside WeChat? Why does WeChat decide which AI agents may reach its users?

The instinctive Western answer involves the Great Firewall, China’s state-run system for blocking foreign websites and services. But the Great Firewall does not directly explain this. The wall that shapes the daily experience of Chinese internet users is a different structure, built by China’s own technology companies against one another. Its blueprint can be traced to November 3, 2010. It was normalized in 2013, partially demolished in 2021, and evolved into a controlled access system for AI agents in 2025 and 2026.

Three Escalations

The conflict became known as the 3Q War, a shorthand combining 360 and QQ. It was not a story of villain and victim. It unfolded in three escalations, each crossing a new line.

The backdrop was an industry already angry at Tencent. Two months before the conflict, Computer World, a major Chinese IT publication, had run a cover story headlined “Fucking Tencent,” a title whose profanity was deliberate. The article’s core accusation: Tencent “followed at every step, then carefully copied, then ruthlessly surpassed” competitors by using QQ’s massive user base to enter one market after another. Published online on July 24 and in print on July 26, the piece captured the industry’s mood. It preceded the conflict by two months.

The first escalation was Tencent’s. The company had been expanding into security software for months. On September 21, it announced that QQ Doctor and QQ Software Manager would automatically merge into QQ Computer Manager. Qihoo 360 was still an unlisted company with annual revenue of $57.7 million in 2010, against Tencent’s Rmb 19.6bn (about $2.9bn at 2010 rates) and a market capitalization of roughly $42bn by late November. But by January 2011, 360’s security products had reached a penetration rate of approximately 83.9% among Chinese internet users, according to iResearch data cited in the company’s pre-IPO SEC filing. What collided was not market capitalization but two kinds of reach: QQ’s social graph and 360’s desktop footprint.

The second escalation was Qihoo’s. On September 27, 360 released Privacy Protector, accusing QQ of scanning users’ personal files. On October 29, the company went further, launching Koukou Bodyguard, software that modified QQ’s interface, blocked its advertisements, and disabled revenue-generating features. 360 claimed over 10 million downloads within 72 hours. Qihoo had crossed from competition into direct interference with another company’s product.

The third escalation was Tencent’s response, and it was the most consequential. Rather than waiting for the legal process it had already initiated, Tencent extended the corporate dispute directly to ordinary users. The practice became known in Chinese as er xuan yi, “choose one of two.” The phrase does not appear in the original letter, but the action it described was unambiguous.

The concentrated enforcement of the forced choice lasted roughly one night. By November 4, after emergency intervention by the Ministry of Industry and Information Technology, 360 withdrew Koukou Bodyguard. But Reuters reported on November 10that 5 million users remained affected. The MIIT formally reprimanded both companies on November 21, ordering public apologies, a halt to mutual attacks, and full software compatibility.

The Price List: What Disconnection Cost Tencent

The MIIT’s reprimand criticized both companies for “unfair competitive practices.” But the language was asymmetrical. The phrase “even unilaterally cutting off service to users, causing adverse social impact” described conduct attributable to Tencent, though the reprimand addressed both companies together.

A regulation issued in December 2011 explicitly prohibited “malicious incompatibility” among internet service providers. The 3Q War was a significant backdrop for its rapid passage. No publicly documented penalty for disconnection under this rule has surfaced in the years since.

Two lawsuits reached the Supreme People’s Court. In the first, Tencent sued Qihoo over Koukou Bodyguard. The court found 360’s software constituted unfair competition by directly modifying QQ’s product and disrupting its business model. 360 was ordered to pay Rmb 5m in damages and issue a public apology.

In the second, Qihoo sued Tencent under the Anti-Monopoly Law. The judgment, dated October 8, 2014, and publicly pronounced on October 16, found for Tencent. The first-instance court described the forced choice as a restriction-of-trade practice that exceeded what was necessary. The Supreme People’s Court reproduced that reasoning in its judgment but resolved the appeal on narrower grounds. It concluded that Tencent had not been proven dominant and that the incompatibility had produced no clear exclusionary effect. The conduct therefore did not violate the Anti-Monopoly Law. The court added: “This does not mean the respondent’s ‘product incompatibility’ conduct is beyond reproach.”

The company that modified a competitor’s product paid Rmb 5m. The company that forced hundreds of millions of users to choose between their messaging service and their antivirus software received no financial penalty for the disconnection itself.

The conflict produced a seemingly contradictory outcome. Within weeks, Ma Huateng issued an internal letter linking the crisis to corporate reflection and later announced that Tencent would reorganize around “openness and sharing.” Litigation, regulatory pressure, and the shift to mobile all contributed. But Tencent itself treated the 3Q crisis as a catalyst. The company became more open to developers inside its ecosystem, not to rival ecosystems connecting to it. Every platform in China could read the price list.

Taobao, WeChat, and the Human Router

The tactic Tencent demonstrated in 2010 became the industry’s default. The mechanism changed: 3Q was desktop software detecting local installations; the mobile era brought server-side URL blocking. But the power dynamic held: the platform controlling an indispensable user entrance could impose the cost of disconnection on others. In the next major episode, Alibaba moved first.

On July 31, 2013, Taobao, Alibaba’s e-commerce marketplace, suspended WeChat-linked applications from its seller service platform. Beginning August 9, Taobao banned newly posted external QR codes, primarily affecting WeChat-based commerce. On November 22, Taobao redirected WeChat users trying to access product and store pages to a download page for the Taobao app. Zhang Yu, then president of Taobao, told reporters: “As long as WeChat can guarantee security and integrity, we will reopen.”

WeChat responded by blocking Taobao links entirely, directing users to copy the URL into a browser. The precise date of WeChat’s countermeasure cannot be confirmed from public records; contemporary reporting says only “shortly after.”

The connection to 3Q was drawn immediately. As early as August 2013, state media analysis linked the Taobao-WeChat standoff to the 3Q precedent. In November, Alibaba’s competing messaging app Laiwang explicitly invoked Tencent’s blocking history.

The pattern replicated across years and companies. WeChat blocked Alipay’s red-envelope links in 2015. Douyin links vanished from WeChat Moments in 2018. ByteDance’s workplace tool Feishu was flagged in 2020. By January 2021, a single WeChat enforcement action caught QQ Music and QQ Browser alongside Zhihu and Xiaohongshu. Even Tencent’s own products were not exempt. Blocking had become a routine instrument of platform governance.

Direct links being blocked, platforms built workarounds that trained hundreds of millions of users to function as human routers. Want to share a Taobao product on WeChat? You could not paste a link. Instead, Taobao generated a string of characters that looked like gibberish, a platform-generated sharing code known in Chinese as a kouling. You copied it into a WeChat message. Your friend copied the noise, opened Taobao, and waited. The app read the clipboard, decoded the product identifier hidden inside, and displayed the item. When WeChat began filtering these passcodes, platforms added special characters and garbled text, what Chinese internet users called “Martian script,” to evade detection. By November 2021, after regulators had ordered platforms to open their links, a Pengpai News test found the workarounds had become more complex, not less.

Some of the security concerns were genuine. Phishing stores proliferated when Taobao traffic flowed through WeChat’s browser. But security enforcement and commercial protection were not fully separable. During Chinese New Year 2026, WeChat blocked AI-powered red-envelope links. The banned list included Tencent’s own Yuanbao AI product. The enforcement did not map neatly onto corporate rivalry.

The Half-Demolition

On September 9, 2021, the MIIT summoned nine major platforms to an administrative guidance session and ordered them to stop blocking one another’s links without justification by September 17. Reported standards required that legitimate URLs open within messaging apps as web pages.

WeChat moved in two phases: one-to-one chats opened to external links on September 17; group chats began a trial for e-commerce links on November 29. WeChat’s Moments, its social feed, was never included in either phase.

The intervention removed the most visible layer: the inability to open a URL. Everything beneath remained untouched. Platforms retained control over native sharing buttons, distribution through WeChat Moments, login and identity APIs, external payment acceptance, and app-level interoperability. “Accessible” did not mean “interoperable.”

Paid connections often opened faster than free ones. Advertisers paying for Moments placement could link directly to Taobao long before ordinary users could share a product link. Taobao’s full integration of WeChat Pay in September 2024represented genuine progress. A regulation that took effect the same month explicitly prohibited the targeted blocking of lawful information, but retained exceptions for security risks and legitimate justification.

As Caixin Global, an investigative business publication, reported, different regulators pursued different objectives. The MIIT focused on market order; other agencies prioritized content security. Full openness risked complicating their work. The contrast with the EU’s Digital Markets Act is instructive: the DMA imposes structured interoperability obligations on designated gatekeepers. China moved faster but left enforcement standards to evolve informally. The walls retreated. They did not fall.

WeChat Agents: From Blocking to Controlled Access

AI agents transplanted the 3Q War’s question to a new interaction layer.

In December 2025, when ByteDance launched its Doubao Phone Assistant, a system-level AI agent that read screens and simulated taps across apps, WeChat blocked Doubao from operating inside the app within days of the launch. The parallel with 2010 is structural rather than legal. Koukou Bodyguard modified QQ’s interface; Doubao simulated user actions from the handset layer. What they shared was the absence of the application owner’s permission. In both cases, Tencent responded by asserting control over its product boundary.

Six months later, Tencent confirmed a reversal. WeChat would provide authorized A2A (Agent-to-Agent) access to phone AI assistants from Huawei, Honor, Xiaomi, OPPO, and Vivo. But access runs through Tencent’s protocol: WeChat executes the commands and defines the scope. The power to disconnect appears to have evolved into the power to authorize connection.

Data Compliance China, a legal analysis site, independently drew the 3Q comparison but argued that the legal frameworks differ. Samm Sacks of Lawfare has explained how walled gardens obstruct agents without connecting the history. I found no previous English-language account that assembled this full sequence.

The underlying question remains: who decides whether software connects or disconnects.

The Garden Was Never Legalized

No Chinese regulator ever authorized platforms to block one another’s links. No court ruled that disconnection was a right. The first-instance court described the forced choice as exceeding what was necessary. The Supreme People’s Court ultimately rejected the antitrust claim while stressing that Tencent’s conduct was not beyond reproach.

What happened was simpler. The price for forcing hundreds of millions of users to choose turned out to be a public apology and zero financial penalty for the disconnection itself. When the cost of building a wall is that low, platforms build them.

The wall that blocks foreign websites from China has a name. The wall that determines what you can and cannot open inside China’s own apps does not. It has only a birthday: the evening of November 3, 2010.

The next time a platform blocks an AI agent, a payment link, or a competitor’s content, the template will be the same one. The public record still shows no financial penalty for disconnection.

Hello China Tech covers how AI, semiconductors, and platform power are reshaping China’s technology sector. Origins is a free series tracing the historical foundations behind today’s headlines. For deep analysis as these stories unfold, subscribe to Hello China Tech.

This is the fourth in a series examining the economics of China’s token supply chain. The first article tested how much of China’s 140-trillion daily token consumption represents real commercial demand. The second used SiliconFlow’s IPO filing to show what it costs to sell tokens for a living. The third examined whether reorganizing hardware ownership changes the economics. This article follows the supply chain to its newest participants: local government investment companies and their banks.

The Full Stack of State Capital

Guangzhou’s Huangpu district cut its land-sale revenue forecast by Rmb 3.7bn last year, from Rmb 9bn to Rmb 5.3bn. In the same district, Guangzhou Science City Group (科学城集团) spent Rmb 15m to build a computing cluster and began selling AI inference tokens to local businesses. The group, a state-owned investment holding company, carries Rmb 134bn in total debt. Local governments in the city subsidize token purchases through separate programs. Three banks have started using token-consumption data in credit decisions.

Four layers of state capital are now visible in the same city: a state enterprise owns the asset, retails the output, local government subsidizes the demand, and banks use token-consumption data as one input in credit decisions. Public disclosures do not show that the layers share clients, data, or funds. But the pattern they form is recognizable. Internationally, Japan’s ABCI charges for compute time, the EU’s EuroHPC opens up to 20% of capacity for commercial use, and Abu Dhabi’s Core42 sells multi-model APIs through a sovereign-backed company. Each covers one of these layers. Among documented cases, none combines all four.

Through a technology subsidiary, Science City Group deployed open-source AI models on its servers. The result is a platform called Bay Area Smart City Model Supermarket (湾区智城模型超市), which offers token-based API access to local firms. In March, the group’s industrial park published an eight-point incentive package granting tenants free tokens and development tools. In August, Bank of China, CITIC Bank, and Bank of Guangzhou introduced loan products in Haizhu district, across the city, that incorporate token consumption into credit assessment alongside conventional collateral.

Every step here has a predecessor. State enterprises building infrastructure assets. Local governments subsidizing consumption to generate activity metrics. Banks lending against those metrics. This is the operating system of land finance, where local governments sold land-use rights, subsidized development, and banks lent against rising property values. The token version follows a familiar sequence with a different asset underneath. Tokens cannot be stored, do not appreciate, and are not bounded by geography the way land is. The banks, at least, appear to have noticed.

Ten Cities and One Motive

Huangpu is not alone. At least 10 cities have launched or announced token-service platforms through state-owned enterprises since early 2026, across three models. In the first, state companies build computing clusters and sell tokens directly; Guangzhou’s Science City Group and Beijing’s JingSuan Company take this path. In the second, state companies aggregate models and compute from multiple providers and act as regional wholesalers; Jiaxing, Wenzhou, and Yangzhou work this way. Jiaxing’s platform operator described the role to Southern Weekend as that of a “mover,” not a producer. In the third, state entities organize token exchanges and distribute fiscal subsidies; Guangdong’s Token Exchange and Service Center and Inner Mongolia’s token-trading platform fit this model.

The three models differ in capital intensity. What they share is motive. National land-sale revenue fell from Rmb 8.71tn in 2021 to Rmb 4.15tn in 2025, a 52% decline over four years. More than 82% of local government financing vehicles had formally left official financing-platform lists by the end of 2025, though the underlying debt did not disappear with the change in classification. A Reuters investigation found that many government-backed data centers run at 20% to 30% utilization. More than 100 data center projects were cancelled in the 18 months before the report. Science and Technology Daily, a government-affiliated publication, profiled one computing center where fewer than half the racks were occupied and actual utilization ran at about 30%.

Token platforms are, in large part, a new retail layer for computing assets that need customers. They convert GPU-hours into a finer billing unit, add model routing and API packaging, and attach fiscal incentives. The research found no state-owned token platform that has disclosed auditable revenue, gross margin, or paid-customer retention. Where prices are published, they have not been tested in markets without subsidies.

Budget documents and credit-rating reports can begin to answer the question that platform announcements do not address.

Science City Group is among the few cases where public filings expose the balance sheet behind a local token platform. Separate bank disclosures in Haizhu district show how Guangzhou is beginning to test token data in lending. Whether these pieces form a system or merely coexist is what the numbers can start to test.

Hello China Tech publishes three analyses a week on China’s AI, chips, robotics, and EVs, built from primary sources rather than second-hand coverage. If tracing a state-owned token platform through its credit-rating report and district budget is the kind of reading you want more of, subscribe free and get each one as it publishes.

Read more

Kuaishou’s second-quarter results, published on August 19, disclosed one substantive Kling number: revenue above Rmb 850m (about $125m), up more than 200% year on year. Kling is the AI video generation unit that Kuaishou, the short-video platform that competes with ByteDance’s Douyin, carved out in July with about $2.8bn in signed investor commitments. The round can extend to $3bn, which would put the post-money valuation near $18bn. Kling has not filed publicly for a Hong Kong listing. But the financing agreement already puts an IPO on a contractual clock, with repurchase rights attached.

That is the official record, and it is thin. On August 25, National Business Daily, a Chinese financial newspaper, published figures from an unaudited internal data set it had obtained. The set covers the share of revenue from API clients rather than subscribers, its concentration in a few overseas accounts, a first-quarter loss, and an internal breakeven target. Kuaishou declined to answer the paper’s questions. None of these figures has appeared in an exchange filing.

Three records of Kling’s business now exist. The exchange filings show scale. Broker research shows part of the revenue mix. The leaked data set goes further, into customer concentration and losses. The numbers a future IPO investor would most want, who pays Kling, through which channel, and how easily they can stop, sit in the outer layers or nowhere at all. The public record supporting a valuation of roughly $18bn still contains that gap.

What Kuaishou does disclose follows a pattern. It publishes totals and counts. Quarterly revenue above Rmb 650m in Q1. An annualized revenue run rate approaching $500m in March. Business clients and developers exceeded 20,000 in mid-2025; by June 2026, Kuaishou said Kling had nearly 50,000 enterprise customers. More than 100m users. On the first-quarter earnings call, founder Cheng Yixiao described the business as driven by enterprise API and professional subscriptions together, and offered no ratio. Totals show scale. Counts show reach. Neither shows structure.

Even the historical financials carry an asterisk. The July financing announcement reports Kling’s 2025 revenue of about Rmb 1.1bn ($162m), its Rmb 1.9bn ($279m) net loss, and its negative net assets of Rmb 9m, all on an unaudited pro-forma basis. The basis assumes the restructuring had already been completed. The figures describe a company that did not yet exist in that legal form. A business priced at nearly $18bn in its latest private round rests, in public, on pro-forma numbers and a leak.

A prospectus will eventually have to answer a plain question: who actually pays Kling, and how easily can they leave?

Most of the answer is already visible, scattered where few bother to look: a broker’s report, an earnings briefing that never reached a filing, and the product pages of the storefronts that sell Kling’s videos.

Hello China Tech publishes three analyses a week on China’s AI, chips, robotics, and EVs, built from primary sources rather than second-hand coverage. If reading an $18bn valuation against the three layers of its public record, filings, broker notes, and a leak, is the kind of analysis you want more of, subscribe free and get each one as it publishes.

Read more

This is the fourth installment in a series tracking how China’s first publicly listed AI model companies navigate the economics of survival. The series began with “China’s AI Unicorns Are Running Out of Runway” in December 2025 and continued with “The First AI Company to Show Its Cards” in March 2026 and “Zhipu Bet Everything on One Equation” in April 2026.

MiniMax, the Shanghai-based AI company that listed in Hong Kong in January, reported first-half results on August 26. Revenue for the six months to June 30 reached $116.6m, up 283.1% from a year earlier. Six months of sales beat all of 2025 by nearly half. On the earnings call, management said its ARR measure passed $800m in August, against $150m in February, and that token consumption in July ran at 20 times January’s level.

The market was positioned the other way. Going into the report, short positions had climbed to 20% of MiniMax’s free float, the highest since listing, against 6% for Zhipu, the Tsinghua University-backed lab that listed one day earlier. MiniMax’s market value stood near $13bn in late August, about a quarter of Zhipu’s. The day after the results, the stock rose 3.8%.

In December I argued that both companies were caught in a structural trap: profitable per API call, insolvent as businesses. In March, after MiniMax’s first annual report showed revenue outrunning R&D, I set four conditions under which the escape could continue. This report grades them against a second set of numbers.

The grade is not the one I expected. One condition split, two held, and the third became harder to classify. Revenue again grew faster than R&D, and the adjusted loss more than doubled at the same time. What my March framework missed was a distinction: enterprise customers do not necessarily imply custom-delivery economics. And MiniMax used the call to state a survival equation of its own, a mirror image of the one Zhipu staked its listing on.

The Numbers

A year ago MiniMax earned 69.7% of revenue from consumer products such as Talkie, its AI companion app, and Hailuo AI, its video tool. In this report, the open platform and enterprise services earned $73.9m, up 703.1%, or 63.4% of revenue. Consumer products grew 100.9% to $42.6m and fell to 36.6% of the mix. Two reporting periods were enough to reverse the revenue mix.

The loss line needs the same two-column reading as in March. The IFRS net loss narrowed 11% to $358m, mostly because last year carried a $253.9m fair-value charge on preferred shares that converted at listing. The adjusted net loss, closer to operating reality, more than doubled to $293m. R&D spending rose 138.8% to $296.9m, 2.5 times revenue.

Gross margin rose to 17.9% from 12.1% a year earlier. Sequentially it fell: full-year 2025 came in at 25.4%, which implies the second half of 2025 ran near 34%. The balance sheet offers signs of where some of the capital is going. Property and equipment jumped from $1.6m to $74.9m, and non-current prepayments rose from $0.9m to $100.8m, a shift consistent with the self-built cluster push management describes, though the filing does not itemize the assets. Cash and cash equivalents stood at $930.9m at June 30, within a broader cash-balance measure of $1.32bn that includes financial assets and deposits. A July placement and convertible bond raised HK$16bn in gross proceeds (about $2bn), and management put current reserves above $3bn.

The efficiency ratios in this report improve while the absolute losses widen. The $800m ARR, the 20x token growth and the 2m enterprise customers and developers explain why the same report supports opposite positions. Deciding between those positions requires reopening the four conditions I set in March, and asking what each of the three numbers actually measures.

If grading an AI company’s earnings against survival conditions set two reports ago 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.

Read more

On the evening of August 22, a humanoid robot ran 100 meters in 9.39 seconds at Beijing’s National Speed Skating Oval, the arena known as the Ice Ribbon. Usain Bolt’s human record is 9.58. It did not slow before the line. It hit a padded wall at full speed, went down, and was carried off by human handlers, the New York Times reported. Three days later the Tiangong robot, built by the Beijing Humanoid Robot Innovation Center, which goes by X-Humanoid in English, ran 8.86 seconds in the large-group second round.

One reading travelled fast in Western coverage, and on the stopwatch it holds up. The 2022 Guinness mark for a bipedal 100 meters stood at 24.73 seconds. Alan Fern, a robotics professor at Oregon State University, worked on it. Getting to 9.39 in four years is a huge achievement, he told the Times, and the gains came mostly from better hardware and control rather than a technical breakthrough. Michael Milford of Queensland University of Technology said most robot sports do not convert into the daily tasks people actually want done. The same report quoted Xinhua, the state news agency, conceding the point in its own commentary: factory robots do not need to win sprints.

I spent two evenings this week with the event schedule rather than the results table. The schedule reads as a grading sheet. The final column holds the rule on how much human help an event allows, and that column is where this year’s change sits.

If an autonomy weight coefficient printed beside each event on a robot competition schedule 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.

The Last Column

The table lists 51 events in two categories, competition and scenario. Each row carries a final column headed 自主权重系数, printed in the table’s own English as Autonomy Weight Coefficient.

Twenty of the 30 competition events carry a single line: fully autonomous mode only. These include every flat race, all three soccer formats, floor gymnastics, the dance events, tai chi and the optional boxing routine, tug of war, pitch-pot and table tennis. Two events, the 100m and 400m obstacle races, accept either mode. For 20 of the 21 scenario events the table sets a price rather than a ban: a fully autonomous run scores at weight 1, a remote-controlled run at 0.5.

Eight of the 30 competition events leave the column blank: the three jumps, both weightlifting classes and all three kickboxing weight classes. The comprehensive pentathlon, the final row of the table, is the ninth blank. One reading is that the table asks for autonomy where autonomy can be graded today and stays quiet where it cannot.

The two complaints I heard most about last year’s first Games were remote control and constant battery swaps. The rules moved on the first one. The municipal government’s own release says several competition events dropped manual remote control and now run autonomously end to end. The table uses two instruments. Twenty competition events shut remote control out. Twenty scenario events let it in at half credit. The weighting itself is not new, since Beijing’s robot half marathon in April already discounted remote-controlled entries in its scoring. What is new is that the autonomy question now runs across a 51-row program, nine rows of which still carry no coefficient at all.

The same release gives the scale: 16 countries, 666 teams, up 138%, 2,056 robots, four times last year’s count, 51 events over five days. Last year’s best 100m was 21.50 seconds and the standing high jump 95.641 centimeters. This year Tiangong Ultra cleared 2.8843 meters.

The largest upgrade at these Games is easier to find in the rules than in any single machine. In 2025 I argued that Chinese robot makers treat public imperfection as a commercial asset, shipping flawed machines and improving them in the open. The imperfection now comes with a weight coefficient attached.

One limit on all of this should be stated plainly. The column records what the organizers want, not what the machines delivered. It shows how autonomy is rewarded. It does not show how many autonomous entries finished, how often a human stepped in, or how the two modes compared on the clock. Those numbers are not in the published table.

Four Exams, Hardest Last

My own sorting, after a week of watching, is that the 51 events fall into four kinds of test. The medal count does not show the order.

A sprint is mostly a hardware and control result. The limits are how hard the joints can be driven, how fast the pack gives up its charge, and how long the machine sheds heat before the motors fade. Those limits move with the budget more than with the model. A bipedal machine beating a human record does show the control stack works. It says less about what the machine understands. The more telling detail is that machines sharing the Tiangong name hold both the sprint mark and the standing high jump. Winning two unrelated physical events points to how much performance now rides on a reusable hardware and control stack. It shows less about intelligence.

In April, Honor, the smartphone brand spun out of Huawei, won Beijing’s robot half marathon ahead of the specialist entrants. I wrote then that building good legs had become an engineering management problem. Four months later, the athlete’s oath at the opening ceremony was taken by an Honor robot named Lightning. The outsider from April was on the platform.

Fighting raises the difficulty in one specific way. A lane holds still. An opponent does not, and the disturbance arrives while the machine is already committed to a move. A Chinese business paper reporting from the floor described boxing robots wandering off in opposite directions. Most punches hit air. One machine raised a leg to kick and fell over first. A friend who was in the venue told me the boxing drew the loudest laughs of the day. My own read, after watching the rounds, is that most still did not dare to really hit. The table agrees in its way, since kickboxing is one of the events where the autonomy column is blank.

Table tennis is where the whole chain has to run at once. From the opponent’s contact to the machine’s swing there are a few tens of milliseconds. Inside that window the robot has to find the ball, project where it will be, and move the arm. A staged demo can cover a slow link in that chain. A rally in front of a camera is harder to stage. AgiBot is a humanoid startup in the listing queue. One of its vice presidents told the same paper that a table tennis system reached that level after a bit over two months of training.

The scenario events empty the seats. They also sit at the top of the difficulty ladder, and the two facts go together: work worth grading is rarely work worth watching. Twenty-one of them cover nine broad fields including home, hotel and logistics. The schedule breaks them into 12 operational tasks, eight dexterous hand challenges and one pentathlon, and that second group maps where the industry actually is. Picking beans with tweezers. Weighing powder. Opening a cap. Fastening a nail. Connecting a cable. Watch two of these events back to back and the slowness stops reading as confusion. It is sequencing: see, compute, move, check, repeat. Each step gets solved before it is taken. That may be the state closest to paid work, because a warehouse buys repeatable before it buys quick.

A year ago I wrote about a robot that spent 17 minutes cleaning a hotel room at the first Games, picking up nine pieces of trash. A human housekeeper finishes that in under five minutes. I used it as a punchline. I would not write it that way now. On the evidence so far, the binding constraint looks less like top speed and more like the hand: grip, force control, fine manipulation. Those 17 minutes were the real exam paper, and I graded the wrong question.

Who Writes the Exam

On August 25, with the Games still running, the Ministry of Industry and Information Technology opened public comment on a draft guide for the national humanoid robot industry standards system, 2026 edition. The document states that no formally released humanoid robot standard exists at home or abroad, with more than 100 standards under research or pre-research. It sets a 2028 target of at least 100 key standards, covering capability testing and evaluation methods, key technologies, platforms and systems, scenario applications and safety governance. It also aims to have them promoted and implemented at more than 200 companies.

The structure runs to six parts, from basic and common standards through to safety and ethics. Two categories in the first part map onto the Oval directly. One grades robot capability by level of intelligence and by task complexity. The other sets benchmarks and test methods for physical function, intelligence features, resource use and system collaboration. A third category sits closer still to what the Oval actually does: standards for reliability and repairability. They cover whether a machine completes a defined task under defined conditions for a defined period, and how it is brought back after a fault.

Put the two documents side by side and the division of labor becomes visible. The Games are an informal benchmark that runs once a year, with rules the organizers can rewrite between editions. The standards guide is the formal counterpart, still in draft, with a 2028 target and a company count attached. August 25 was also the day the sprint record fell again. The overlap in timing may be coincidence. The overlap in content is not.

The Games are hosted by the Beijing Municipal People’s Government and China Media Group, the state broadcaster. Two international bodies co-host: the World Robot Cooperation Organization and the Asia-Pacific RoboCup International Council. Operations are run by BAOC Group and Shougang Fund. A separate company, Shoucheng Holdings, describes itself as a deep participant in the Games through robot leasing, training and testing, and venue support. It is also an active investor in the sector, with holdings in Unitree, Galbot and several other Chinese humanoid startups. Kang Yu, general manager of its capital markets department, told reporters during the Games that the event works as a reliability test under the spotlight.

What follows is a question of degree. The institutions here are separate bodies. What converges around this one event is the functions: standard-setting, venue operations and investment attention, sitting closer together than a Western reader would expect at home. The effect worth watching is on price discovery rather than on any single result. Nothing in the record shows a medal moving a valuation. What the record does show is testing capacity, technical visibility and capital attention gathering at one venue. I have argued before that China’s state apparatus rarely builds the robots itself. It builds the machinery that lets robots be measured, ranked and priced. The last safeguard clause of the standards guide points at the next step. It asks domestic bodies to track key international standards, join international standardization work, and push mutual recognition between Chinese and foreign standards. That clause is where this reaches readers outside China. One industry report puts China at 97% of global humanoid shipments in the first half of 2026. That share of output gives China a reason to want influence over how humanoids are tested. Whether it gets that influence depends on adoption outside China and on participation in international standard-setting, and the draft settles neither.

What the Exam Cannot Measure

Kang Yu also warned against the obvious misuse of the event her firm helps support: medals rank single metrics, and single metrics do not add up to a company. Her sharper point concerned quantity. The audience sees one robot set a record. Some companies brought more than ten machines to the venue. Going from one to ten raises the consistency question, and consistency is the problem mass production has to solve. She makes the same case about joints, which carry a large share of a robot’s value and a large share of its reliability faults.

At the World Robot Conference, a trade show running in the same city that week, the mood ran the other way. Reporting from the show caught two founders trading valuation tactics over lunch. Go and find out what peers are quoting, one told the other: if a rival asks Rmb 1bn, we ask Rmb 800m. The sector’s valuation math I set out last week has not changed. One number in it has. Unitree founder Wang Xingxing pushed his timeline for an embodied-AI ChatGPT moment out to two or three years at best, five or ten at worst.

Two shipment counts circulate for the first half of 2026. The 40,000 units behind that 97% share come from an industry report. The head of Beijing’s Shijingshan humanoid robot data training center gives 23,000 for the domestic market, counted on a different basis. His estimate carries the harder qualification: of those 23,000 units, under 5% were stably doing real work on a production line. The reporting that carries his estimate proposes a stricter commercial test than shipment counts. An order for 100 machines followed by a larger repeat order the next year would say more than any first deployment. Nothing in these materials records one.

AgiBot’s argument that more real-scenario data will feed back into the models is reasonable on its own terms. It runs into the problem I set out earlier: robot training data has no common denominator across different bodies and tasks.

Midnight in the Prep Hall

Kang Yu’s best material, in the same interview, was not on the track. Near midnight, she said, she passed the soccer preparation hall and found the lights still on. Middle school students were still tuning their robots. One was doing summer homework while code ran. One had fallen asleep on the grass, then sat straight up when a teammate said the machine seemed off. Those people, she said, are the industry’s future.

A fast time travels well online. Taking a hit, holding an object and finishing a shift draw far less attention online. The table now tests versions of all three, though it still does not grade them equally.

So here is the reading order for next August. Open the new event table first, and the new records second. If the blank cells in the autonomy column start filling in, and if the weight for remote control drops below 0.5, the exam got harder. That would mean the organizers believed the field could pass it. If that column comes back unchanged, then this year’s rules were the ceiling rather than a step.

On August 19, LandSpace (蓝箭航天) launched its Zhuque-3 rocket from northwestern China carrying a satellite. About 6 minutes later, the first-stage booster landed on 4 legs at a site 390 kilometers downrange in Gansu province. The second stage later placed the satellite into orbit. LandSpace, one of China’s leading private rocket companies, became the second Chinese entity to recover an orbital-class first stage. The Long March 10B, built by China Aerospace Science and Technology Corporation (CASC), had done it roughly 6 weeks earlier, catching its booster in a net at sea. The two approaches reflect different optimization targets: nets save the weight of landing legs, while legs eliminate the need for a ship, which could speed turnaround.

The landing succeeded. The engineering story that matters, though, sits further back in the supply chain. LandSpace builds Zhuque-3’s body from ordinary stainless steel and hires engineers from China’s auto industry, where 34.5 million vehicles were produced in 2025. The company concentrates aerospace-specific technology on a few decisive points: the methane engines, thin-wall welding, and return flight control. LandSpace aims to route the rest through China’s general industrial base.

This division of labour could lower launch costs. But the potential rests on questions the August landing did not answer: refurbishment time, maintenance scope, and whether the reuse cycle can beat the manufacturing cycle. The reuse model began producing inspection data on August 19. It has not yet produced results.

A Rocket Priced Off the Steel Market

Traditional rocket bodies use aerospace-grade aluminum alloys. The most advanced, aluminum-lithium alloys, offer high strength-to-weight ratios. They also concentrate manufacturing knowledge inside state aerospace institutions. A rocket-tank designer at CASC noted in 2021 that Al-Li alloy prices ranged from Rmb 300,000 to Rmb 1m per ton, compared with roughly Rmb 100,000 for standard aluminum-copper. The equipment and expertise to fabricate large-diameter Al-Li tanks resided almost entirely within China’s state aerospace system.

Stainless steel operates in a different economy. Standard 304 stainless trades on a transparent spot market at about Rmb 12,600 to 14,550 per ton in late August 2026. LandSpace has not disclosed its alloy grade or procurement cost. Dong Kai, Zhuque-3’s deputy chief designer, described the company’s material only as “a mass-market product, tens of thousands of yuan per ton.” The gap between stainless and aerospace aluminum is more than a price ratio. It is a supply chain shift. China produced 40.9 million tons of stainless steel in 2025, about 64% of global output. Tank fabricators across the Yangtze Delta handle large stainless vessels as routine work. Few commercial suppliers outside the state aerospace system work with thin-wall aerospace aluminum at comparable scale.

Elon Musk followed the same logic when he switched Starship from carbon fiber to 301 stainless in 2018. Carbon fiber cost $135 per kilogram before waste. With 35% scrap, the effective price approached $200 per kilogram. Stainless cost $3 a kilogram, he told Popular Mechanics. But his decision also reflected thermal properties: steel’s high melting point reduces heat shielding, which simplifies reuse.

For China, the substitution carries a further implication. By choosing a material with a deep, competitive domestic supply base, LandSpace connects the rocket’s largest structural component to the general economy rather than the aerospace procurement system. The material is cheap. The question is what it takes to make commodity steel precise enough to fly.

The steel has a transparent market price. The process that makes it flyable does not.

How commodity steel converts material cost into process cost 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.

Read more

On August 20, Alibaba reported a 75% drop in net income. US-listed shares fell 4.6% in pre-market trading, then closed the day at $130.53, up 1.26%. The next day they closed at $119.34. Then on Sunday, August 23, Alibaba announced a HK$80bn ($10.2bn) share placement in Hong Kong to fund AI investment, the largest primary follow-on offering ever by a Hong Kong-listed company. The shares priced at HK$112.70, an 8.4% discount to Friday’s close. In early Monday trading in Hong Kong, the stock fell as much as 10%. Day one shrugged. Day two sold. Day three put a price on the AI bill.

The 75% headline is a poor guide to Alibaba’s quarter. The decline combined operating pressure with weaker investment income: operating profit fell 57%, while lower disposal gains and smaller mark-to-market changes widened the gap in GAAP earnings. Adjusted EBITA fell 30%. The adjusted EBITA decline reflects Alibaba’s deliberate acceleration of AI investment, as management explained on the call.

Three weeks ago, I argued that Alibaba’s AI empire had no checkout: equity gains, cloud orders and hardware costs sat in three separate ledgers, and no public filing reconciled them. This earnings report installs half of one. Starting this quarter, Alibaba reorganized its financial disclosures into four segments: Alibaba E-commerce Group, AI Cloud and Compute Services, AI Labs and Applications, and All Others. The restructuring pulled AI model spending out of the opaque “All Others” residual and gave it a visible line. For the first time, investors can see what the infrastructure side earns and what the application side burns.

The half that remains missing is the external portfolio. Alibaba’s AI investments, including stakes in Moonshot and MiniMax, two Chinese frontier AI model companies, and CXMT, China’s leading DRAM manufacturer, still generate paper gains that flow into non-operating income. Cloud orders from portfolio companies remain undisclosed. Hardware-cost relationships go unmapped. The internal AI P&L gained a register. The external one did not.

The Ledger Alibaba Chose to Open

The new segment structure splits Alibaba’s AI business across two reporting lines.

AI Cloud and Compute Services, which combines the Cloud Intelligence Group with chip design unit T-Head, posted revenue of Rmb 48.4bn, up 45% year over year. Growth hit a 22-quarter high. AI-related product revenue reached Rmb 12.4bn, the 12th consecutive quarter of triple-digit growth, now accounting for 35% of external cloud revenue. Adjusted EBITA rose 133% to Rmb 5.6bn. The margin climbed from roughly 7% a year ago to 11.6%.

AI Labs and Applications houses model development, the Qwen consumer app and QwenWork enterprise agents. It reported revenue of Rmb 3.3bn, up 16%, and an adjusted EBITA loss of Rmb 13.9bn. The recast year-earlier loss was Rmb 3.2bn, previously buried inside “All Others” alongside Amap (Alibaba’s mapping and navigation app), Alibaba Health, gaming and other units. The current deficit is more than 4 times as large.

The disclosure points toward one of the five markers I set in July. I asked whether Token Hub, the organizational group CEO Eddie Wu created to unify model services, would become a reporting segment with its own margin. The result is close but incomplete. AI Labs and Applications partially fills that role. But MaaS revenue, the model-as-a-service business Wu has positioned as the cloud’s highest-margin product, still sits inside the cloud line. The register is clearer. It is not yet clean.

The surrounding numbers explain why the segment restructuring appears to have been timed to this quarter. Revenue grew 9% to Rmb 269bn. The E-commerce Group generated Rmb 39.7bn in adjusted EBITA, stable year over year, providing the cash that funds the rest. Quick commerce revenue, which includes Taobao Instant Commerce and Freshippo, grew 45% to Rmb 53.3bn. Alibaba said Taobao Instant Commerce improved unit economics sequentially. Core e-commerce advertising, measured by customer management revenue, fell 7% on a reported basis and grew just 1% on a like-for-like basis after accounting reclassifications. The cash engine still runs. Its output is flat. Like-for-like CMR growth fell from 8% the prior quarter to 1%, a deceleration steep enough to test how long e-commerce profits can keep funding the AI bill.

On the same call, Wu outlined the investment math behind Alibaba’s AI spending: payback periods, asset lifetimes and the conditions under which growth can become cash generative. The model, and the gaps it leaves, sit behind the paywall.

Hello China Tech publishes three analyses a week on China’s AI, chips, robotics, and EVs, built from primary sources rather than second-hand coverage. If the gap between cloud AI profitability and model-layer losses inside a single company is the kind of reading you want more of, subscribe free and get each one as it publishes.

Read more

One thing separates this toolbox from generic industry analysis. These frameworks were not imported from strategy textbooks and applied to China. Each one was extracted from a specific mechanism inside China’s tech system: how the state architects markets, how coordination compresses time, how listings engineer scarcity, how procurement substitutes for benchmarks. To understand the framework is to understand the mechanism. That is the point of this index. A few of the frameworks turned out to travel beyond China, and where that is true, the entry says so plainly.

Each entry gives the idea in one line, names the piece of China’s system it captures, says when to apply it, and links the articles that built it. Where a framework carries a live, falsifiable test, the entry names it.

Two notes on scope. First, the six field-level principles are not repeated here. They live in The China Tech Field Guide, which is free and covers the top of the stack. This index works one level down, at the tools I reach for while writing. Second, the index is alive. New frameworks are added as they earn their place, roughly every two months, and entries are updated when the evidence moves.

Here is the full toolbox, twenty frameworks in five groups:

Capital and valuations: The Scarcity Trap · The Perpetual Raise · Denominator Skepticism · The Consensus Machine · Sell-First vs Build-First · Geopolitical Arbitrage and the Two Walls

Compute and chips: The Credential Economy · Token Export · The Token Factory Test · The Sanctions Paradox

Tokens and AI economics: The Token Hangover and Cost per Completed Task · The Seat Outlives the Agent · Harness Engineering · The Siphon Effect

Platforms and interfaces: The Two Tencents · The Infrastructure Trap · The Do-Engine

Hardware businesses: The Two-Week Wall · The Chip Bag Test · Collect From Both Ends

One entry follows in full, so you can see what the format delivers. I chose the one whose prediction has already been settled in public.

This index is a founding-member reference at Hello China Tech, independent analysis of China’s AI, semiconductor, robotics, and EV sectors, built on Chinese-language primary sources. The best place to start is free: selected analyses each month, including The China Tech Field Guide.

Subscribe free →

The Scarcity Trap

From: Zhipu and the Scarcity Trap (February 2026) · Reused in: Zhipu Bet Everything on One Equation, Moonshot Sells the Right to Say No, CXMT’s 466% Debut on a 6.73% Float

The idea in one line: When a sliver of free float lets concentrated buying set the price for every share outstanding, the valuation is a price of scarcity, not of the business. Lock-up expiries put that scarcity on a timer.

What it reveals about China’s system: Float engineering. Chinese listing structures routinely bring single-digit floats to market, with cornerstone allocations and lock-up schedules printed in the filings. The rally and its reversal are both written into the share register before trading starts.

Use it when: a Chinese listing posts an extreme post-IPO rally on a single-digit float, and the unlock dates are already printed.

Where it stands: Round one is settled. The February piece flagged Zhipu’s 8.5% float; both Zhipu and MiniMax fell in July’s unlock week, then placed more than HK$40bn combined. The mechanics moved to CXMT: unlocks in January and July 2027, the second one roughly five times the opening float.

Read more

Baidu closed at $90.87 on August 18, down 12.73% on the day it reported second-quarter results. The company’s market capitalization settled around $31bn.

Kunlunxin, Baidu’s AI chip subsidiary, is seeking a dual listing in Shanghai and Hong Kong at a reported target valuation of $50bn, according to The Information. Baidu holds about 58% of Kunlunxin through Baidu (China) Co., Ltd.

At the reported target, Baidu’s stake in Kunlunxin would equal roughly 93% of Baidu’s market capitalization. The comparison implies a large holding-company discount.

An Arithmetic That Does Not Close

In January 2026, when Kunlunxin first filed confidentially for a Hong Kong IPO, I estimated the implied valuation gap at $36.7bn. At the time, Baidu traded at $51.4bn. I estimated Kunlunxin at about $18bn using one method. A separate comparable-company calculation suggested Baidu’s stake could be worth as much as $22bn, or about 45% of the parent’s market capitalization at the time.

Seven months later, the components have moved in opposite directions. Baidu’s market capitalization dropped to roughly $31bn. The reported Kunlunxin IPO target was $50bn. At 58% ownership, Baidu’s stake would be worth about $29bn on paper. That is roughly 93% of the parent’s entire market value before accounting for AI cloud, Apollo Go, Rmb 283.1bn in cash and investments, or any other asset.

Baidu still trades at a large discount to a simple sum of its disclosed asset values. The pattern suggests a structural divergence between where value is created and where it is priced.

The Collapse That Wasn’t

The headline number in the Q2 earnings release was a 68% drop in net income, to Rmb 2.3bn. Baidu’s shares fell 12.73% that day. Bloomberg led its coverage with the decline. But the number obscures the operating picture.

Baidu General Business, the reporting segment that excludes streaming subsidiary iQIYI, posted operating income of Rmb 3.1bn, down about 6% year over year. Operating margin held near 12%. Legacy business revenue fell 23% and online marketing declined 19%, but AI revenue growth and lower operating expenses partly offset these declines. The net income drop comes from below the line. Fair value gains on long-term investments shrank from Rmb 3.3bn a year ago to Rmb 1.1bn. Foreign exchange losses reached Rmb 1.2bn. Lower investment gains and a larger foreign-exchange loss account for most of the net income decline.

The real action is in the balance sheet and the cash flow statement. Capital expenditure for Q2 reached Rmb 11.4bn ($1.7bn), roughly triple the Rmb 3.8bn a year earlier. Combined with Q1, Baidu spent about Rmb 17.3bn on capex in the first half. Operating cash flow remained positive at Rmb 3.4bn. But free cash flow remained deeply negative because of the capex surge: negative Rmb 8.0bn for Q2, and negative Rmb 11.2bn for the first half.

Borrowing increased in the same period. Short-term loans rose from Rmb 7.6bn to Rmb 26.3bn. Non-current long-term loans rose from Rmb 3.4bn to Rmb 20.8bn, while the current portion fell from Rmb 14.8bn to Rmb 2.0bn. Lease right-of-use assets doubled from Rmb 9.3bn to Rmb 18.6bn. Total liabilities rose from Rmb 159.4bn to Rmb 174.8bn in six months. Baidu increased its borrowing while holding Rmb 283.1bn ($41.7bn) in total cash and investments.

The question is where the return on that spending shows up.

Baidu spent Rmb 17.3bn on capex in the first half of 2026 while total liabilities rose by Rmb 15bn. GPU Cloud revenue grew 283%. Kunlunxin is reportedly seeking a $50bn valuation in its planned IPO. The missing data is how much of Baidu’s cloud infrastructure uses Kunlunxin chips, and how much of Kunlunxin’s revenue comes from Baidu.

If the idea that one subsidiary could account for 93% of its parent’s market value 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.

Read more

Unitree Robotics, the Hangzhou company that has become the public face of China’s humanoid robot boom, listed on Shanghai’s STAR Market, the mainland’s Nasdaq-style technology board, on August 19. The stock closed its first day at Rmb 845 a share, up 460.34% from the Rmb 150.80 offer price. The close valued the company at Rmb 341.77bn, about $50bn. Against the Rmb 288m net profit reported in its March IPO filing, that equals roughly 1,190 times earnings.

The next morning, on August 20, Wang Xingxing spoke at the World Robot Conference, an annual industry event in Beijing. It was his first public speech as the founder of a listed company. He used part of it to reset his own deadline.

Wang has a benchmark for the moment humanoid robots become an industry rather than a stage act. Take a robot across a range of unfamiliar environments, homes included, and have it complete about 80% of tasks in about 80% of those settings through voice or language instructions. He calls this the sector’s ChatGPT moment, and by March 2026 he had given it that measurable form. What keeps changing is the date. At this valuation, investors have left themselves little room for the slow end of his range.

This piece builds on Hello China Tech’s March reading of Unitree’s IPO filing, which flagged the 80/80 threshold as the line between a hardware vendor and a platform company. If tracking that kind of claim across filings, speeches, and listings is useful to you, subscribe free to get every new analysis as it publishes.

Subscribe now

Two Augusts, Two Timelines

Wang has now dated what he calls the ChatGPT moment twice, at successive World Robot Conferences, one year apart.

At the 2025 conference, he put it this way, per Sina Finance’s unreviewed transcript: “At the fast end, the next one or two years, or two to three, is quite possible. At the slowest, I estimate three to five years, with high probability.” He added an invitation: keep watching this.

On August 20, per a Tencent Tech transcript republished by 36Kr, the estimate became: fast, two to three years; slow, somewhere between five and ten.

The arithmetic needs care, because two clocks run here. Measured as remaining time, the earliest estimate moved from one year to two, and the far end moved from five years to ten. Measured on a calendar, the shift is larger, because a year had already passed. The earliest date moved from 2026 to 2028. The slow window moved from 2028–2030 to 2031–2036. The full range doubled in width, from four years to eight.

Both versions carried an optimistic note. Last year it was “quite possible.” This year, Wang said the pace of progress might make the moment arrive “a tiny bit faster” than his own estimate. His optimism stayed in place. The dates behind it moved.

This reads as an update, not a retraction. Founders adjust timelines, and a transcript of a conference speech is an imperfect record. But each adjustment has pointed the same way: outward.

The Threshold Is the Multiple

A founder’s soft dates would not matter much if the stock were priced on the business that exists. It is not.

The business that exists is a profitable hardware company. Listing-day disclosures show revenue of Rmb 1.152bn for the first half of 2026, up 48.54% from a year earlier. Net profit reached Rmb 274m, against a Rmb 32m loss in the same period of 2025. The hardware business is financially real. It is also decelerating: year-on-year revenue growth fell from 335% for full 2025 to 68.49% in the first quarter and 48.54% for the half. The prospectus, as I noted in March, describes embodied AI models that have not been scaled into commercial products and were tested only in controlled pilots. It calls the whole field still at the “research and testing stage.”

The 80/80 threshold marks the border between that company and the one the multiple requires. Below it, Unitree remains what its filing shows: a seller of research platforms to a customer base with natural limits. Above it, robots justify deployment through versatility, and the company changes category. A price near 1,190 times reported earnings is far easier to defend under the fast case than under a five-to-ten-year wait.

The Prospectus, Read Aloud

What Wang said on stage was franker than the share price suggested. Many of the polished robot routines audiences have watched for years, he said, relied on actions trained in advance. The new demos, generated in real time from voice commands, carry a delay of several seconds. The system recognizes the speech, generates and checks an action in the cloud, then sends the command back to the robot. A model trained thoroughly on one fixed scene approaches 100% success; change the object or the environment slightly and the success rate collapses. Each new task still requires retraining. Unitree’s own world-model effort began in early 2020, was shelved by the end of that year for poor results, and was revived in 2025.

None of this is new in the legal sense. The risk section of the prospectus carried the substance five months ago, in filing language. What changed is how directly it was said. Risk factors are written to be skimmed. A founder telling a conference hall that the smooth demos were trained in advance is meant to be heard. In a sector that has often rewarded being seen over being ready, plain speech from its most visible founder is worth recording.

His proposed way across the border is a self-evolution loop. A frontier language model drives an agent that reads new papers, writes robot control code, and tests it in simulation. The code then runs on physical machines, and AI plus human judgment scores the results. He attached no budget, no timeline, and no benchmark to it. It is a research direction, announced the day after the listing, in an area Wang described as the company’s largest current commitment of money and people.

A 7.44% Float Sets the Price

The listing measured demand more clearly than it measured conviction about Wang’s timeline. Retail subscriptions ran to more than 8,000 times the shares available. Regulators had approved the listing in 104 days, the fastest review in the STAR Market’s history. The stock opened at Rmb 1,100, up 629.44%, before settling at Rmb 845. Chinese financial media quoted a closing multiple near 585 times, computed on a different earnings basis from the figure above, against a sector reference average of 38.56.

The structure helps explain the valuation. Only 30.09m shares, 7.44% of the total, traded freely. CXMT, China’s leading memory-chip maker, rose 466% in July on a 6.73% float. Thin floats let small pools of money set large prices, and the same report ties Unitree’s swings to its scarce tradable supply. The stock fell more than 17% on its second day, the day of the speech. The float can amplify moves in either direction; the available record does not settle a single cause.

A price built this way measures scarcity more than belief in any timeline. Chinese media describe Unitree as the first humanoid robot stock in mainland China’s domestic A-share market. If the share price carries little information about the threshold, then the week’s material statement on it came from the stage, not the market.

The Clock to Watch

The test Wang set has two virtues rare in this sector: he defined it himself, and anyone can check it. A robot, unfamiliar environments, voice instructions, and an 80% task-success rate across about 80% of those settings. At the 2025 conference he told his audience to keep watching. One year of watching produced one data point: the full range widened, and every revision pointed outward.

The useful evidence from here will not be another polished demo. It will be a reported success rate in unfamiliar settings, less retraining for each new task, or revenue from deployments running Unitree’s own models. If Wang updates the range again at the 2027 conference, investors get another public data point. A narrower range would mean buyers at roughly 1,190 times earnings reached the platform company early. A wider one would leave them holding a hardware maker whose first-half revenue grew 48.54%, valued as though general-purpose autonomy were much closer.

Two reports, published a day apart in August 2026, described the same project in different terms. LatePost, one of China’s most authoritative tech publications, reported on August 6 that ByteDance, TikTok’s parent company, was discussing a model exceeding 5 trillion parameters, still at an early stage and not necessarily destined for release. The Financial Times reported the following day that 3 people familiar with the effort described a model of up to 10 trillion parameters in early-stage training. One of them said pre-training had begun; the final scale remained undecided.

The reports diverge on the numbers. Both point to the same organizational commitment: ByteDance is concentrating talent, data, and compute behind a single long-duration technical bet.

In late July, ByteDance founder Zhang Yiming made a rare appearance at a Seed team all-hands alongside Seed chief Wu Yonghui. He told the team to accept falling behind competitors for a period and to push toward the frontier of model capability rather than benchmark positions. He also instructed them to stop relying on distillation of rival models. CEO Liang Rubo reinforced the message at a company-wide all-hands on August 6, calling AI the long-term strategic priority and noting that more than 90% of new clients for Feishu, ByteDance’s collaboration suite known internationally as Lark, had purchased AI products.

ByteDance built its empire on a specific operating system: launch parallel products, compare measurable outcomes, and concentrate resources once one route pulls ahead. The 5-to-10-trillion-parameter project retains one half of that formula, the belief that overwhelming scale can force a breakthrough, while discarding the other, internal competition among parallel teams. The question is whether concentration without the safety net of parallel bets can survive a project this large.

The Operating System That Built ByteDance

ByteDance’s founding myth is a scaling story. According to LatePost, the company launched 13 apps in its early years. Neihan Duanzi, a humor platform, and Jinri Toutiao, a news aggregator powered by algorithmic recommendation, emerged as its first large-audience successes.

The 2016 short-video push refined the template. Zhang Yiming told employees in a 2019 anniversary speech that ByteDance had decided not just to enter short video but to build two products, simultaneously pursuing domestic growth, international expansion, and acquisitions. Three products launched within months: Xigua Video, Huoshan, and Douyin, TikTok’s Chinese twin. Early user data favored Huoshan. Longer-term retention favored Douyin. Resources shifted accordingly.

The system had four components. Ship fast. Measure in days, not quarters. When data converges, consolidate. When it does not, cut. Dual-month OKR cycles forced teams to show progress within 60 days. Zhang Yiming described the approach in that same speech: “Looking back, many of our methods weren’t good at the start, but we were very dedicated, very focused. Sheer effort can work miracles.”

Where those conditions held, the method scaled. Where they did not, it produced expensive failures. ByteDance invested in or acquired more than 20 gaming companies through its gaming arm, with estimated spending of roughly Rmb 30bn. Management later acknowledged the strategy had been “big but not focused.” In March 2026, ByteDance sold Moonton, the Southeast Asian studio behind Mobile Legends, to Saudi Arabia’s Savvy Games Group at a valuation exceeding $6bn, according to Reuters. The sale ended an unsuccessful gaming strategy, though Moonton itself had appreciated from its roughly $4bn acquisition price. Its tutoring division, Dali Education, grew to over 10,000 employees by 2020, spanning K-12, early childhood, and adult learning. China’s tutoring crackdown in July 2021 forced closures, though the pattern of rewarding short-cycle metrics over long-cycle quality was visible before regulation intervened. Wukong Q&A, a Quora-style platform with an estimated Rmb 2bn in subsidies according to PE Daily (投中网), a Chinese venture-capital media outlet, shut down in January 2021.

The pattern is consistent. The operating system works when products ship quickly, feedback is quantifiable, and failed teams can be redeployed cheaply. Consumer apps met all three. Education, gaming, and community products required longer development cycles, harder-to-measure quality, and domain expertise that resists reallocation.

The question has always been which side of that boundary AI falls on. ByteDance’s consumer AI assistant, Doubao, sits comfortably inside it: 382 million monthly active users as of June 2026, according to QuestMobile, a Chinese mobile analytics firm, leading the country’s AI apps by a wide margin. Training a frontier model sits firmly outside. One run can consume months, cost several hundred million dollars, and offer limited intermediate signals to steer by. The feedback loop is as long as anything ByteDance has attempted.

The Seedance comparison grows less reassuring once model state, cluster architecture, and training data enter the calculation. Those are the three areas where ByteDance’s video-model advantage offers the least protection.

If the organizational history behind ByteDance’s AI research, from internal horse racing to concentrated bets, 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.

Read more

In a story circulating through China’s robotics industry, an embodied AI company paid Rmb 100m, partly in equity, for a few minutes of exposure at a televised gala. The board was not asked. Investors later gathered in the founder’s office; at least one pounded the table. The episode comes from a contributed deep-dive published August 13 by TMTPost, a Chinese tech business outlet, which named neither the company, the show, nor anyone in the room.

The same report describes a dinner at which founders compared cash runways. One said his company had 50 months of cash. Another claimed 100. Neither number can be traced to a financial statement.

The stories may be true. The verifiable record is much thinner: robot makers did appear on CCTV’s 2026 Spring Festival Gala, China’s most-watched broadcast, and when 36Kr, a Chinese tech news site, relayed chatter that companies had bid Rmb 60m to Rmb 100m for gala slots, one company denied it and another went silent. Even the sector’s most public spending exists only as deniable rumor.

The money behind the rumors is real. By one count from QbitAI, a Chinese AI-focused outlet, embodied AI companies raised about Rmb 43.8bn (roughly $6.4bn) in 2026 through June 12; the figure is a media tally, not a regulatory dataset. Single rounds are enormous: TARS Robotics, a Chinese embodied AI startup, closed a $455m Pre-A round in April that the company billed as a sector record.

China’s embodied AI sector is described by two competing records. One is a folklore ledger: gala invoices, runway boasts, hoarded billions, sourced to people who cannot be named, making claims no filing can check. The other is the record produced under mandatory disclosure, and for a sector that has absorbed tens of billions of renminbi it is thin: three disclosure events. An IPO prospectus. An acquisition announcement. A liquidation file. Each shows a company losing the right to choose its numbers, and where those numbers touch the folklore, the folklore fares badly. In a market where consensus is manufactured faster than evidence, unverifiable spending stories help set prices whether or not anyone designed them to.

A ledger built from hearsay

The folklore’s consistent theme is cash as identity. The TMTPost report describes a house style of extreme thrift that insiders reportedly nickname the “outlast-them” playbook, with companies parking funding in wealth-management products and running teams off the interest. It describes two sets of accounts, one sent to an investor by mistake, killing a Rmb 20m to 30m round, and institutions posting finance and anti-fraud staff inside portfolio companies; not one of these stories names a company.

The anonymous accounts rarely follow the thrift through to its operating consequences. In the report’s grimmest anecdote, a company froze research hiring, cut component-validation batches and postponed data collection to stretch its runway by nearly a year, then watched rivals poach its engineers while a product meant for small-batch delivery failed to reach production because of component-compatibility problems. A hardware executive quoted in the report has a name for the condition: a well-funded stall.

Private-company stories describe discipline. Mandatory disclosure measures it: spending mix, losses, liquidity, who actually runs the company. The three cases that follow are not a sample of the sector; they are the moments it became measurable.

If the folklore ledger reading 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.

Read more

FlashPoint is our premium quick-strike column on market-moving China tech events. Today: Hugging Face’s summer report finds the platform’s 25 most-downloaded and 25 most-liked open models share exactly 1 repository. In January, I analyzed China’s robot makers buying visibility before viability. In July, Kimi K3’s launch repriced Moonshot’s IPO option. Now Hugging Face has put numbers on that trade.

In January, China’s robot makers were buying airtime on the Spring Festival Gala, the country’s most watched broadcast, while one leading maker said its flagship ran at 30 to 50% of human efficiency. The AI-lab version of a Gala slot is the trillion-parameter open-weight release. Hugging Face, the main repository for open models, set the 25 most-downloaded models of the first 7 months of 2026 against the 25 most-liked; they share 1 entry. No 2026 model makes the download list; 13 of its 25 entries date from 2022. all-MiniLM-L6-v2, a small embedding workhorse, drew 1.55bn downloads and 5,156 likes: about 300,000 per like. Kimi K3, Beijing-based Moonshot AI’s 2.8tn-parameter flagship and largest open-weight commitment to date, collected about 60 per like. What a frontier release buys is belief, and belief is the asset China’s model labs are selling this year.

The two counters record different acts. A like says a release matters; it arrives within weeks and clusters on frontier models. A download says a model is wired into a scheduled pipeline, accruing to small, stable workhorses over years. The report finds Chinese frontier labs are the only accounts whose download volume concentrates above 70bn parameters: effectively all of MiniMax’s 2026 pulls, 88% of Moonshot’s. The count excludes API calls and private deployments, a boundary that matters for Moonshot. The open question is what applause without pipelines purchases, and how quickly the investors paying for it can change their minds.

This piece extends the visibility-over-viability lens I first built covering China’s humanoid robot IPO rush. If that is a useful way to read China’s AI capital cycle, subscribe free and get every new analysis as it publishes.

Read more

On August 13, the morning after Tencent reported second-quarter results, the stock closed at HK$441.00, down 4.46%. The Hang Seng Tech Index rose 0.33% that day, leaving Tencent 4.79 percentage points behind its own benchmark on a day the benchmark went up. The US-traded ADRs had priced the release a day earlier, falling 5.44% to $56.28 on August 12, then a further 1.79% to $55.27 on August 13.

This has now happened twice. On March 19, after Tencent posted its strongest annual results in years, the stock fell 6.81% and lagged the same index by 4.62 points. The absolute decline was larger in March. Measured against the index, the August reaction was at least as severe.

The two sell-offs share more than a shape. In March, President Martin Lau asked investors to treat Tencent’s AI spending as a strategic investment rather than margin compression, and to judge it separately from the core business. In August he made the request explicit. “You can look at Tencent businesses and break it into two businesses,” Lau told analysts on the earnings call. One is the existing franchise. The other is a new AI-native business with its own model, its own applications, and its own compute.

A simpler story is available for the sell-off. IFRS net profit attributable to shareholders came in at Rmb 56.0bn against a Reuters consensus of about Rmb 61.8bn, and free cash flow turned negative. A miss plus a cash shock accounts for one bad day without any theory about disclosure. It does not account for what the research did next. Analysts used Tencent’s excluded-profit figure to confirm the core business was healthy, then went on charging the AI spending against the value of the group. The price move is a day of evidence. The published notes are the durable part.

Two Profit Numbers, One Cash Flow

Tencent now publishes two operating profit figures. Non-IFRS operating profit for the quarter was Rmb 75.6bn, up 9%. Excluding new AI products, it was Rmb 86.1bn, up 19%. The difference is the effect of what Tencent calls new AI products, which its footnote says consist “primarily” of Hunyuan, the in-house foundation model family its results materials brand as Hy, plus the Yuanbao chatbot, the CodeBuddy coding tool, the WorkBuddy office agent, and Xiaowei, the WeChat-native assistant. The perimeter remains management’s to redraw. Operating margin diverged the same way: 36.9% blended, down 0.6 points year on year, against 42.0% excluded, up 2.8.

The cash statements moved in the opposite direction. Capital expenditure reached Rmb 52.8bn, up 176% year on year and 65% from the first quarter. Free cash flow was negative Rmb 13.8bn, which Tencent attributes to capex payments and large compute prepayments in a seasonally weak quarter for game receipts; excluding the prepayments it puts the figure at positive Rmb 37.6bn. Net cash fell 60% to Rmb 58.2bn from Rmb 146.9bn, absorbing Rmb 59.3bn of capex payments, Rmb 41.6bn of 2025 dividends and about Rmb 14.7bn of buybacks. The same quarter a year earlier produced Rmb 43bn of free cash flow on Rmb 19.1bn of capex, and left Rmb 74.6bn of net cash.

Lau framed the spending as finite. The AI-native capex is “more of a sort of a lump sum that we are going to be investing this year and next year,” he said, and paying for it “should not be just measured against our operating cash flow” but against the balance sheet, the investment portfolio and a prudent level of debt capacity. That is a request to move an expense out of the earnings frame and into the funding frame.

Bernstein’s Robin Zhu asked what these purchases do to depreciation over the coming quarters. Chief Strategy Officer James Mitchell answered with the rental option, a use Tencent has explicitly ranked last, and Lau answered with the funding source. Property, plant and equipment rose from Rmb 149.9bn at the end of December to Rmb 198.9bn at June 30, a net figure that already absorbs depreciation and disposals and is not broken out by use. The direction is not in doubt: a capex step this size lifts depreciation for years. Tencent has not quantified the timing, or how much lands inside the excluded perimeter

Two profit numbers, one cash flow statement, and a request to weight the higher one. What the sell-side did next was more interesting than agreement or refusal.

This piece extends The Two Tencents, the March analysis of a company asking to be valued as a cash machine and an AI venture at the same time. If that lens is useful, subscribe free and get every new analysis as it publishes.

Read more

The Denial

On July 30, The Wall Street Journal reported that Tesla advisers had discussed a spinoff, sale or closure of the China business, and that some executives had been told to prepare for a separation ahead of a possible merger with SpaceX. The planning reflected a standing instruction from Musk, according to the Journal’s sources: run the company with a “laser” down the middle, so a US-China rupture would take out one side and leave the other trading.

Musk answered on X hours later. “This has never even come up in a discussion ever. Absurdly fake news.” Tesla’s China unit told Chinese media the report was untrue. No 8-K followed.

The denial is broad in an unusual way. It rejects the conversation itself, a step further than denying a decision. The analysis that followed took the same frame. Reuters Breakingviews argued a buyer would demand a discount and that a carved-out unit would stay dependent on licenses from its former parent. A Reuters legal analysis walked through why shared software, IP and data governance would slow any deal. All of it argues about a transaction that may never happen.

Operational separation does not require a transaction. Tesla’s China business has regionalized along four lines: a hard data boundary set by Chinese rules, software stacks diverging under rules on both sides, supply chains split by end market under tariff and continuity pressure, and a local funding base from Chinese banks. A sale would transfer the equity easily enough. The software licenses, data governance and valuation gap are the hard parts. Soft isolation may therefore be an equilibrium rather than a waypoint to a transaction.

Two Walls, One Company

Beijing built the first one. Since 2021, data from China-market vehicles has stayed on servers in Shanghai. Foreign carmakers cannot independently hold the surveying qualifications needed to collect road-mapping data, so Tesla licenses maps and lane-level navigation from Baidu, the country’s dominant search company. Material over-the-air updates that change driver-assistance behavior must be filed with or approved by MIIT, China’s industry ministry. A June 2025 draft guidance classified data used to train autonomous driving as important data, needing a security review before it leaves the country.

Each requirement reads as ordinary compliance. Stacked, they cost Tesla something. When it shipped a feature-limited China release in February 2025, it could neither move Chinese road data to its US clusters nor train on an equivalent local set, and Baidu engineers were in its Beijing office by March fitting the system to Chinese lane-level navigation. China joined Tesla’s supervised FSD markets in May 2026, still as a Level 2 system. Chinese industry media report that local training capacity is being built, though Tesla has not disclosed its scale or legal entity.

Washington built the second. The Commerce Department’s connected vehicles rule took effect on March 17, 2025. Software restrictions bind from model year 2027, connectivity hardware from model year 2030. March 17, 2026 is the legacy line: older code keeps its exclusion only while entities under Chinese jurisdiction refrain from maintaining, enhancing or modifying it. Example 38 reaches software developed through a China-based subsidiary that retains servicing obligations or contractual rights.

Tesla helped shape that perimeter. In an October 2024 comment letter it asked Commerce to keep firmware outside the definition and to exclude software that merely supports other systems. The final rule did both, narrowing covered software to what directly enables connectivity or automated driving.

The rule reaches Tesla through legal entity rather than corporate nationality. Chinese supplier filings then show who is financing the regionalization below the automaker.

This piece runs the Identity Squeeze framework in reverse, applying it to an American company being pulled apart by two regulators rather than a Chinese one assembling an American identity. If that lens is useful, subscribe free and get every new analysis as it publishes.

Read more