China Desk

China News, Summarized 23 Oct 2026, 04:55 UTC 34 stories 22/23 sources

Today's brief

China is closing ranks: Xi around one core, industry around DeepSeek

Xi turns the Long March into a loyalty lesson

Xi wants one core. Marking the Long March's 90th anniversary on Thursday, Xi said the Red Army survived by rallying around Mao Zedong and that China now faces a "new Long March". It came four days before the fifth plenum, the Central Committee's closed-door meeting, set for 26 to 29 October.

The line aimed inward. Seoul's Kyunghyang Shinmun notes that Xi blamed "leftist dogmatism" for nearly sinking the Party before Mao took command in 1935. That passage was absent from the 80th-anniversary speech a decade ago. The paper reads it as aimed at the Party itself. My read: divided leadership nearly ended it, and one core saved it.

Huang Qifan is the context. On 13 October the Central Commission for Discipline Inspection, the Party's anti-graft body, opened a probe into Huang, 74, Chongqing's mayor from 2010 to 2016. He had outlasted the falls of his superiors Bo Xilai and Sun Zhengcai.

The count is the point. SCMP's tally, relayed by ua.news, is that 35 of the Central Committee's 205 full members have been investigated or lost their legislative mandates. On 17 October I argued that surviving one purge protects nobody from the next. A retired mayor fits that.

THE TELL: state broadcaster CGTN's English summary stresses "loyalty, truth, justice and unity" and the road to a modern socialist China. In the version I read, it never mentions leftism or a political core. SCMP and the Korean paper lead on both. No mainland outlet in today's feed covered the speech.

The AI treaty reaches the legislature. The National People's Congress Standing Committee meets 2 to 6 November and will consider ratifying the agreement creating the World AI Cooperation Organization. Twenty-nine countries signed it in Shanghai on 16 July, Russia, Brazil and Cuba among them. It takes effect once a third state deposits ratification. Beijing is building its own AI institution while Washington sanctions its AI investors.

Go deeper on this section: ClaudeChatGPT

DeepSeek's investors are industrial China

Nobody is on the record. DeepSeek has raised about 100 billion yuan (~$14 billion), Bloomberg reported, which prints $15 billion. The target rose twice, from 50 billion yuan (~$7 billion) to 80 billion (~$11 billion), then settled higher. Reuters says the round is still closing and figures could change. No report carries a comment from DeepSeek.

Suppliers and customers, not funds. CATL, the battery giant, and Tencent are among the biggest investors. Quartz, citing the Wall Street Journal, adds carmaker Geely and Lenovo-backed Legend Capital as newcomers, with a valuation near $74 billion going in. Reuters says many investors are state-backed or corporate. My read: this is a national-champion round, not a venture bet.

CATL is paying while bleeding. On 20 October I flagged that CATL's profit lagged its sales. TMTPost reports the shares then fell over 12% on 21 October, the biggest one-day drop since listing, to 227.9 yuan (~$32). Third-quarter gross margin slid to 20.69%, and nine-month operating cash flow fell 23.6%. It has also put over 10 billion yuan (~$1.4 billion) into AI-energy plays.

THE NUMBER: $1 billion. That is DeepSeek's annualised revenue by 24 September, per China Daily's account of The Information's report, more than double a few months earlier. Liang Wenfeng, the former hedge-fund manager who founded DeepSeek, told investors price rises hadn't driven customers away. On my arithmetic, the valuation is about 74 times that. Pricing power is what the money is buying.

Chinese coverage is a rewrite. Zhidx, the tech site, relays Bloomberg as China's largest AI financing and ends by saying products and an IPO must prove it. It omits the state-backed investors Reuters flagged. It does add that Moonshot AI is reportedly eyeing a Hong Kong listing in early 2027, while DeepSeek has hired CITIC Securities for a Shanghai listing.

Go deeper on this section: ClaudeChatGPT

Narrow models, hard kernels, robot faith

Two decision models in one day. On Thursday Ant Group's enterprise-technology unit released LingDT-Decision-1.0 at 4 billion and 27 billion parameters. China Telecom's AI lab open-sourced Xing4.0-29B-Intent. Both pick from a fixed list of options instead of writing an answer, targeting high-volume routing that general models handle expensively.

Self-reported, but specific. Ant says token use falls over 90%, and customer-service risk detection rises from 76% to 89% versus Qwen-3.7-flash. China Telecom scores 58 against 49 for the best of 18 general models on 6,000 public-dataset questions, at under 200 milliseconds. Ant adds ordered decisions expressed as a dependency graph.

WHY IT MATTERS: generation is the wrong tool for routing. A refund, a renewal cancellation and an invoice request arrive in one sentence, and the answer set is closed. Ant concedes that spotting dependencies is not executing them: money and permissions still need business rules.

Qwen and Kimi share one attention recipe. At AICon Shenzhen a Qwen engineer laid out FlashQLA, a GPU kernel library for linear attention. Qwen runs three linear-attention layers to each softmax layer. Kimi K3 uses the same three-to-one split with its own variant.

The catch is occupancy. The recurrence allows one thread block per head, so a 32-head model leaves most of a GPU's 100-plus multiprocessors idle. The fix is sequence parallelism with a correction matrix, written as hand-built kernels because Triton can't schedule it.

Unitree's founder says he hasn't believed enough. Wang Xingxing spent 36 minutes at Unitree's Hangzhou conference urging the room to trust AI. Since August he has cut his estimate for robotics' ChatGPT moment from two-to-ten years to one-to-three. He showed GPT-6 driving a robot untrained. GeekPark's reading: if rivals supply the brains, the winner builds bodies cheapest. He wants 10 million units a year, not tens of thousands.

Belief carries a multiple. On 20 October Unitree itself warned of a trailing price-to-earnings ratio of 388. A chairman asking for more faith is also defending that number. That is my inference.

Go deeper on this section: ClaudeChatGPT

China's helium price crashed behind a wall

Helium fell 80% at home. Domestic high-purity helium peaked near 497.5 yuan (~$70) per cubic metre in April, after the Middle East conflict hit Qatar's output. It now sells for 61 to 105 yuan (~$9 to $15), while overseas prices still climb. Beijing banned helium exports on 10 July. Jiemian gives that equal billing with record Russian imports and new domestic plants.

Who it protects. China imported about 84% of its helium last year. My read: the wall insulates buyers who can't stockpile, since liquid helium boils away within roughly 45 days.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • The export draft is a ledger. On 20 October I read Beijing's draft as a registry more than a gate. ChinaTalk's three-day reading agrees. It counts 40 additions across AI, solar and biomedicine, 14 banned and 26 licensed. Interim registration applies even to exports already under way, which it reads as Beijing building a dataset before deciding what to license.
  • AI short drama got cheaper, not richer. On 20 October I said margin moves to rights and the shelf. GeekPark's interview with Huang Haonan, founder of studio Soy Sauce Culture, adds numbers. Cost per minute fell from 1,000 yuan (~$141) to 400 (~$56). Monthly output rose from 40–50 titles to 200–300, yet revenue stayed near 20 million yuan (~$2.8 million). Taste is a third owner: capacity scales, judgement doesn't.
  • Brussels is not softening. SCMP reports that Brussels and Berlin feel emboldened and are drafting fresh restrictions, while Beijing bristles at the idea it was strong-armed. On 15 October I said planners would keep talking while tightening. The tightening is holding; the talking is the fragile part.
  • Memory keeps climbing, on one poster's word. A V2EX user says a pair of 32GB sticks costs nearly 6,000 yuan (~$845), and a friend paid 7,100 (~$1,000) days earlier. That is one anecdote, not data, but it fits the shipment falls I have tracked since 14 October.

Go deeper on this section: ClaudeChatGPT

The river · 34

DeepSeek reportedly raises about $140 billion in its largest funding round

# Zhidx 1d

Bloomberg reports a roughly 1,000-billion-yuan ($150 billion) round for DeepSeek, with Tencent and CATL among the largest investors, to fund compute and model development.

Also: InfoQ China, IT Home

More
  • The report says the fundraising target rose from 500 billion yuan to 800 billion yuan and then to about 1,000 billion yuan (roughly $140B), setting a record for China’s AI sector.
  • The account links the financing to data-center infrastructure and continued development of open models for domestic and overseas markets.
  • It says the round could provide a base for an IPO and help narrow the funding gap with leading US labs, though neither outcome is assured.
  • The item notes DeepSeek’s recent releases of models and inference tools as part of a strategy that differs from closed, product-led lab models.
  • The claims are attributed to Bloomberg and people familiar with the financing; the article says whether capital translates into lasting performance or overseas revenue remains to be seen.
Open at Zhidx Ask Claude ChatGPT

Xi orders urgent rescue after a deadly Tibet earthquake

# SCMP 2h

A magnitude-7.0 earthquake in Tibet killed 16 and injured dozens, and Xi Jinping ordered search-and-rescue efforts in the high-altitude, cold-weather area.

Also: IT Home

More
  • The report says the immediate priority is rescue and treatment, with authorities also monitoring aftershocks, weather and risks of secondary disasters.
  • The episode shows how disaster response in remote western China is framed through central leadership instructions as well as local emergency operations.
  • Fatality and injury counts may change as rescue work continues; the report is an early account rather than a final assessment.
Open at SCMP Ask Claude ChatGPT

iFlytek says its new models were trained on domestic chips

# AI Zhidx 15h

iFlytek unveiled a cloud-edge model family and speech model, reporting improvements in long-context training and distributed inference on Chinese compute.

Also: InfoQ China

More
  • The company says training a 2,000-billion-token, 256K-context workload on its Feixing II cluster fell from 4.5 months to 1.2 months after optimization.
  • It reports a four-layer KV-cache architecture that raised cache-hit rate from 63% to 95.7% and improved inference efficiency by 4.5x on its domestic hardware platform.
  • The edge 4B model routes work among local, cloud and hybrid execution; iFlytek says routing accuracy rose from 81.2% to 96.1% after joint training.
  • The technical story is about co-designing model training, distributed inference, caching and routing around heterogeneous domestic accelerators, not merely swapping chip brands.
  • These performance and benchmark claims come from iFlytek and need independent replication; the product release overlaps with the company’s broader speech-model announcement.
Open at Zhidx Ask Claude ChatGPT

China’s humanoid-robot boom depends on collecting human hand data

# Culture ChinaTalk 18h

A ChinaTalk contest-winning essay examines the labor and infrastructure behind first-person data collection for Chinese humanoid-robot training.

More
  • The essay cites reports of more than 150 humanoid-robot companies and 40 state-owned robot-data centers in China, placing data collection within national industrial strategy.
  • First-person video and hand-tracking records offer a lower-cost route to varied human-object interactions than collecting every example on expensive robot hardware.
  • The excerpt frames data as a human labor and power question: workers’ ordinary actions can become training assets, raising issues of consent, compensation and ownership.
  • It says robot firms are increasingly focused on hands because dexterity and coordinated manipulation remain difficult to reproduce.
  • The piece is a reported essay from a China-focused contest; the figures cited are attributed to other sources and are not independently audited here.
Open at ChinaTalk Ask Claude ChatGPT

China Telecom releases an open model for intent parsing and structured extraction

# AI QbitAI 21h

China Telecom’s Xing4.0-29B-Intent combines intent classification with slot extraction, returning structured JSON for customer-service and device-control workflows.

More
  • The model uses a schema-driven semantic-parsing framework to represent categories and parameters in a consistent output format for downstream systems.
  • Unlike a general model that may produce long answers, the company says the model returns a few dozen tokens with latency in the hundreds of milliseconds.
  • It supports multi-intent requests and parameter extraction, such as identifying a room, device and desired setting from one spoken instruction.
  • The lab says it is releasing model weights and 2.75 million training examples; the benchmarks are company-run comparisons and need independent reproduction.
  • The design is useful where outputs must be fast and machine-readable, but it still needs business rules and execution systems to safely carry out actions.
Open at QbitAI Ask Claude ChatGPT

Xi’s Long March speech stresses Party leadership; Li Qiang meets Putin

# Policy Sinocism 7h

Sinocism says Xi’s anniversary speech elevated the Party’s correct leadership as the central lesson of the Long March, while Premier Li Qiang met Vladimir Putin.

More
  • The newsletter contrasts the speech with the 2016 anniversary address: the new framing gives greater emphasis to Party leadership and the importance of a central leader.
  • The text quotes Xi describing earlier military losses as resulting from errors in the Party line, then presenting the 1935 Zunyi meeting as the beginning of Mao’s leadership.
  • The anniversary event drew senior leaders and about 3,000 attendees, signaling a major commemorative setting for historical messaging.
  • The newsletter also lists Li Qiang’s meeting with Putin, but the provided excerpt gives no substantive account of their discussion.
  • The speech is a political narrative about how the Party interprets its own history, not simply a commemorative address.
Open at Sinocism Ask Claude ChatGPT

Unitree’s Wang says faster foundation models could unlock physical AI

# AI GeekPark 16h

At Unitree’s 10th-anniversary event, Wang Xingxing argued that robots may soon inherit more general task skills, while hardware scale remains a key advantage.

Also: 36Kr

More
  • Wang cited demonstrations in which a robot connected to a general model changed paint colors and placed Go stones without task-specific training; conditions are not independently documented.
  • He says sensor and actuator alignment remains the major obstacle between language-level planning and reliable manipulation in changing physical environments.
  • The report says Unitree’s G1 has shipped about 10,000 to 20,000 units, while the company is developing a cycle of code generation, simulation, hardware tests and self-evaluation.
  • Wang’s strategy depends on increasingly capable outside models while Unitree concentrates on robot bodies, manufacturing scale and a self-improving control stack.
  • The prediction that household robots could arrive soon is a forecast, not a demonstrated capability; repeatable task success in unfamiliar homes remains unproven.
Open at GeekPark Ask Claude ChatGPT

AI comics are cheaper to produce, but harder to monetize

# Culture GeekPark 20h

Chinese AI-comic studios say output has surged while revenue has barely moved, as falling production costs bring more creators into the market.

More
  • One producer says an 80-minute series fell from roughly 100,000 yuan (about $14,000) to 40,000 yuan (about $5,600) in production cost, while output rose sharply.
  • The same account says monthly production increased severalfold without a comparable rise in revenue, as lower prices and token costs absorb labor savings.
  • Platforms are shifting toward individual creators and small studios, weakening the position of large production houses that once negotiated with platforms as intermediaries.
  • The article describes a new scarcity: with AI reducing production barriers, judgment about story, pacing and intellectual property matters more than execution capacity.
  • Industry figures and earnings examples are interview-based; the report does not establish that individual creators will reliably capture the value platforms once paid to studios.
Open at GeekPark Ask Claude ChatGPT

A Chinese critique calls out sensational coverage of young Japanese skipping meals

# Culture China Digital Times 1h

A translated Chinese commentary argues that domestic outlets exaggerated a Japanese survey by presenting meal-skipping as proof of economic decline.

More
  • The critique says the source survey found time pressure and lack of hunger were more common reasons for skipping meals than saving money, contrary to headlines that emphasized hardship.
  • This is valuable as a rare internal critique of Chinese foreign-news framing: the writer argues that traffic-driven stories use another country’s difficulties to make China look better.
  • The commentary does not deny food-price pressure in Japan; it objects to selective emphasis and the use of an anecdote as a national trend.
Open at China Digital Times Ask Claude ChatGPT

Recursive opens Flint, a locally runnable coding model with 40 billion parameters

# AI Zhidx 1d

Recursive released a MoE coding model that activates about 2.5 billion parameters per token and runs on consumer GPUs and Apple silicon.

More
  • The model has about 39.8 billion total parameters and reportedly generates more than 150 tokens per second on Nvidia RTX Spark and 125 on an M4 Pro MacBook.
  • Recursive reports scores of 54.9% on Terminal-Bench 2.1, 89.7% on SWE-Bench Pro v2 and 74.2% on SWE-Bench Verified; these are company-provided benchmark results.
  • The model was trained from scratch, including its tokenizer, pretraining, supervised tuning and reinforcement learning, with the company saying the full pipeline took four months.
  • The active-parameter count helps explain how a large MoE model can fit local inference constraints, though memory and quality still depend on quantization and workload.
  • The release advances a researcher’s self-improving-AI agenda, but the report says complex tasks may need clearer decomposition and user prompts.
Open at Zhidx Ask Claude ChatGPT

EU considers easing sanctions on a Chinese textile maker in a reciprocal deal

# Policy SCMP 8h

EU diplomats are weighing sanctions relief for China’s Yusha Group after Beijing reportedly pledged to restrict its sales to Russia-linked buyers and remove a German firm from its own controls list.

More
  • The proposed exchange links EU sanctions, Chinese export controls and assurances about sales to Russian military-linked customers, creating an unusual piece of reciprocal leverage.
  • The reported arrangement would require trust in Beijing’s commitment and a way to monitor future shipments; the article does not describe enforcement terms.
  • The case suggests that sanctions lists may become bargaining instruments in wider trade and security negotiations, not only unilateral penalties.
  • The account is based on sources familiar with discussions and describes a proposal under consideration, not a final EU decision.
Open at SCMP Ask Claude ChatGPT

A China-developed cancer drug reaches US patients after medical trips to Shanghai

# Analysis SCMP 6h

AstraZeneca says it will begin US sales of a cancer drug first developed in China after some Americans traveled to Shanghai to obtain it.

More
  • The report describes a gap between US regulatory approval and domestic availability that drove patients to seek the medicine in China.
  • The episode reverses the usual direction of medical tourism and is a notable case of Chinese-origin drug development becoming relevant to US patients.
  • The article does not identify the drug or give enough trial and access details to assess how broad the effect is.
Open at SCMP Ask Claude ChatGPT

Ant’s LingDT targets structured decisions rather than open-ended answers

# AI Leiphone 18h

Ant Digital says its 4B and 27B models support multi-label choices and ordered actions, aiming to reduce token use and latency in business workflows.

Also: InfoQ China

More
  • The model is designed for bounded tasks such as customer intent, risk classification and tool selection, where a valid choice set is known in advance.
  • Its claimed directed-acyclic-graph ordering feature represents dependencies among actions, such as checking an order before issuing a refund.
  • Ant reports over 90% lower token use in some deployments and ranks its 27B model first on a 23-benchmark comparison, but the article does not provide independent replication.
  • The approach fits alongside general LLMs: use a specialist for high-volume structured decisions and retain broader models for open-ended reasoning.
Open at Leiphone Ask Claude ChatGPT

StartLux opens Prism, a decision-model inference engine built for shared context

# AI InfoQ China 0m

StartLux says its engine speeds decision-model inference by up to 5.3x on one H200 by reusing shared context and computing only candidate outputs.

More
  • Unlike general text-generation servers, Prism targets models that score a finite set of choices; it avoids calculating logits for an entire vocabulary when only a few candidate tokens matter.
  • A tree-attention approach reuses the common document context across several questions, while tensor and data parallelism target latency and throughput respectively.
  • StartLux reports 92.94% accuracy versus 92.97% at BF16 baseline across 12,351 decisions, but the speedups are company tests and depend on context overlap.
Open at InfoQ China Ask Claude ChatGPT

Chinese state media presses KDDI to apologize over a Mongolian wildfire

# Policy SCMP 1h

SCMP says Chinese state media called for Japanese carrier KDDI to apologize to China over a fire in Mongolia, as Beijing-Tokyo tensions run high.

More
  • The fire reportedly damaged about 80% of a Mongolian nature reserve near the China border, and KDDI has linked it to activity by its majority-owned subsidiary.
  • The framing turns a cross-border environmental incident into a China-Japan political issue, distinct from the company’s apology to Mongolia and cooperation with its investigation.
  • The account is a useful example of how state-media coverage can recast a neighboring-country incident through a bilateral diplomatic lens.
Open at SCMP Ask Claude ChatGPT

Chinese internet users read France’s protests through domestic and geopolitical lenses

# Culture SCMP 55m

Chinese online reaction to France’s protests mixes geopolitical cynicism, pride in China and concern over damage to French landmarks.

More
  • The protests began with complaints about school conditions and widened into grievances over living costs and public services, according to the report.
  • The reported online response shows how foreign unrest can become a mirror for domestic comparisons, rather than simply a discussion of French politics.
  • The article describes a mixture of reactions, not a representative survey of Chinese public opinion.
Open at SCMP Ask Claude ChatGPT

Tesla reports higher vehicle margins but sharply lower operating profit

# Analysis QbitAI 21h

Tesla’s third-quarter operating profit fell 56% as research and selling costs rose, even while vehicle margins improved and paid FSD users grew.

More
  • The report says Shanghai factory wholesale sales reached 95,366 vehicles in September, including a 58% year-over-year increase in exports.
  • Tesla said paid FSD users reached 1.74 million globally, with monthly subscriptions exceeding half of the total; the article does not give a China-specific FSD deployment date.
  • The company’s growing R&D spend includes AI and other projects, while its dependence on Model 3 and Model Y remains pronounced.
  • The China signal is mainly manufacturing and export performance; the story is otherwise a US-company earnings report.
Open at QbitAI Ask Claude ChatGPT

BroadLink says specialized models can bring AI appliance costs near $1 per year

# Analysis GeekPark 15h

BroadLink says a collection of small expert models can bring appliance AI costs to about 1 yuan per device per year, with larger models used mainly for review.

More
  • The company argues that household devices need deterministic, low-latency and low-cost decisions, rather than open-ended generation on every interaction.
  • Its proposed architecture uses small task-specific models in the appliance stack and a larger model as a supervisory reviewer that can analyze errors and support iteration.
  • The interview says existing connected devices can receive some upgrades over the air, avoiding a new hardware certification cycle; actual activation and usage remain weak in many appliance categories.
  • The 1-yuan annual cost is a company claim, and the article does not disclose volumes, accuracy or which inference costs are included.
Open at GeekPark Ask Claude ChatGPT

Survey finds young Chinese adults are saving and buying insurance for security

# Culture 36Kr 16h

A 908-person survey reports average annual savings of 25,400 yuan (about $3,600) among people born in the 2000s, with nearly two-thirds holding commercial insurance.

More
  • Respondents most often worried about illness or job loss interrupting household income, while many preferred savings they could access quickly rather than locking funds away.
  • The survey says 64.9% had bought commercial insurance online and estimates average cumulative coverage at about 1.34 million yuan (roughly $190,000).
  • The findings suggest younger adults are treating insurance and liquid savings as defenses against uncertain income and medical costs, rather than saving only for major purchases.
  • The survey was commissioned by a youth-focused publication and is not presented as a probability sample of all Chinese young adults.
Open at 36Kr Ask Claude ChatGPT

Former US official discusses realistic goals for AI-risk talks with China

# Policy The Wire China 20h

Seth Center, who led early US-China AI discussions in the Biden administration, argues that bilateral dialogue should focus on achievable risk-reduction goals.

More
  • The interview is about what the two governments can realistically discuss on AI risk, not a report of a new agreement or formal negotiation.
  • Its value is the policy question of whether US-China channels can address shared hazards despite strategic competition; the supplied summary provides no specific recommendations.
Open at The Wire China Ask Claude ChatGPT

Service robots are moving from demos into China’s commercial venues

# Culture TMTPost 25m

Nearly 30 robotics firms are reportedly deploying in tourism, retail and hospitality, where bounded tasks can provide real-world operating data and revenue tests.

More
  • The article’s examples include robots serving coffee, greeting visitors and delivering food; operators say these sites are more forgiving than factories or homes while still exposing robots to unpredictable people and conditions.
  • One company says its drink system has about one failure per 1,000 cups, while another says a wheeled service robot can work through an eight-hour day; these are company claims, not common benchmarks.
  • The choice of wheeled platforms over bipedal robots for some shifts highlights a practical trade-off: uptime and task reliability can matter more than humanoid form.
  • The report also describes a shift in public response from treating robots as novelties to expecting them at some venues, though its evidence is anecdotal.
  • The harder commercial test is whether deployments reduce total labor and service costs once supervision, maintenance and venue operations are included.
Open at TMTPost Ask Claude ChatGPT

Quad leaders will meet as Manila signals possible talks among US allies

# Policy SCMP 13h

The Philippines’ envoy signaled possible “Squad” talks ahead of an ASEAN gathering as the Quad confirmed a leaders’ summit focused on the region.

More
  • The informal Squad group is described as part of a US-led effort to counter China’s growing regional influence; the article does not report a new China-related agreement.
  • The timing alongside ASEAN and a future Trump-Xi summit makes the meeting a signal of regional alignment, but the item is about plans rather than decisions.
Open at SCMP Ask Claude ChatGPT

An AI agent completes a reported home sale in China without a commission

# Analysis 36Kr 3h

A property agent called Yiju Xiaoxin reportedly matched buyers and sellers and supported negotiation for a 1.95-million-yuan home sale, while people handled viewings and closing.

More
  • The model separates information work such as matching, valuation and negotiation from in-person fulfillment such as inspections, financing and transfer services.
  • The seller reportedly waived a 40,000-yuan commission, passing the amount to the buyer; this is a single transaction, not proof the model is economically sustainable.
  • The test of the approach is not only whether an agent can facilitate a deal, but who bears responsibility when its advice or matching causes a dispute.
  • Replication depends on local service networks, homeowner willingness to list and consumer trust in direct buyer-seller negotiation.
Open at 36Kr Ask Claude ChatGPT

Jiuhe Advanced raises a pre-Series A round for metal 3D printing

# Analysis 36Kr 2h

The aerospace materials and additive-manufacturing supplier raised several tens of millions of yuan (roughly $4-7M) and says it has shipped parts to leading commercial-space customers.

More
  • The company’s thesis is that metal additive manufacturing is shifting from printer and process performance toward joint optimization of alloy composition, powder production, printing and heat treatment.
  • It says its aluminum alloy reaches yield strength above 500 MPa without rare-earth elements, while costing several hundred yuan per kilogram (roughly tens of dollars per kilogram).
  • For a rocket-pump impeller, the company says its unsupported-printing approach cuts support material by 96%; it also reports fine internal channels as narrow as 1 mm.
  • The technical challenge is managing different thermal histories across alloy melting, powder atomization, printing and post-processing, rather than optimizing only the print stage.
  • The claims are company-reported, but the article describes a meaningful manufacturing strategy: materials design and process control are being treated as one system.
Open at 36Kr Ask Claude ChatGPT

Meituan plans to offer its AI store manager free to restaurants nationwide

# AI Leiphone 2h

Meituan says its AI operations assistant can diagnose listings, menus and promotions, and will be offered free to restaurants across China.

More
  • The product uses Meituan and merchant data to generate operational recommendations and help follow up on them, extending AI from advice toward platform-mediated execution.
  • Meituan cites a small restaurant whose weekly orders rose more than 30% after changes and a chain whose new-store output rose 25%; these are company examples, not controlled results.
  • The rollout may lower the barrier for small businesses to use analytics, while deepening merchants’ reliance on Meituan’s marketplace data and operating tools.
Open at Leiphone Ask Claude ChatGPT

Tencent opens its search infrastructure to AI developers

# Analysis GeekPark 20h

Tencent turned its Yuanbao search stack into an API, positioning trusted, current information as infrastructure for agents rather than a consumer search destination.

More
  • The company says the system serves more than 50 internal products, indexes information at the hundreds-of-billions scale and draws on 157 authoritative sources.
  • The shift is from returning webpages for people to returning structured, timely information that models and agents can use as inputs.
  • Search quality depends on crawling, parsing, source ranking and filtering low-quality generated content, all costly infrastructure for smaller developers to build themselves.
  • This is a Tencent product narrative; usage figures and quality claims are company-reported, and the supplied excerpt does not detail independent comparisons.
Open at GeekPark Ask Claude ChatGPT

Chinese drone-show supplier says German customer ordered 7,000 aircraft

# Analysis 36Kr 4h

Shenzhen-based Damoda says a German customer expanded purchases to 7,000 drones, while the company emphasizes automated launch, charging and recovery systems.

More
  • The engineering challenge is fleet coordination at scale: synchronized positioning, routing, communications and collision avoidance, combined with rapid ground deployment.
  • Damoda says its vehicle-based system can set up in about 60 seconds, reducing the labor needed to stage large shows.
  • The company says equipment sales make up about 70% of revenue and its systems reach more than 50 countries; these figures come from the company.
  • The export model sells equipment, software, training and local operating support rather than relying solely on the Chinese supplier to staff every performance.
Open at 36Kr Ask Claude ChatGPT

China’s first AI feature film opens to weak presales

# Culture TMTPost 39m

The AI-generated film Sanxingdui: Future Tales entered cinemas with low presales, testing whether production-cost promises can translate into paying audiences.

More
  • The report says advance ticket sales were about 144,000 yuan (roughly $20,000), with a box-office forecast of about 2.17 million yuan (roughly $300,000), against a reported production cost in the tens of millions of yuan (roughly $1-2M).
  • The film’s production involved extensive human iteration: the team reportedly generated more than 1.2 million images and made over 100 small revisions after approval.
  • The contrast is between investor enthusiasm for AI production and audience willingness to pay; cheaper generation does not remove the need for continuity, performance, editing and distribution.
  • Short-form and platform-distributed AI video may have clearer monetization than a theatrical feature, where defects are magnified and viewers commit time and ticket money.
  • The first film is a process precedent for certification and release, not proof that AI can yet deliver a commercially successful feature.
Open at TMTPost Ask Claude ChatGPT

Chinese astronomers identify a stable three-star system with FAST

# Signal SCMP 3h

Researchers using China’s FAST radio telescope report a stable system containing two dead stars and a living companion about 7,000 light-years away.

More
  • The observation offers a real-world counterpoint to the chaotic three-sun system in Liu Cixin’s science fiction, while the actual system is described as stable.
  • The short report provides no journal, methods or detailed orbital parameters, limiting what can be assessed about the finding.
Open at SCMP Ask Claude ChatGPT

China’s average model-service price and throughput both roughly doubled

# Signal IT Home 1h

A Chinese Academy-affiliated report says token use rose sharply, average domestic throughput doubled, and normalized model prices nearly doubled in 2026.

More
  • The report puts average domestic output throughput at 78.23 tokens per second in August, versus 38.25 in 2025; its overseas comparison averaged 112.82 tokens per second.
  • Domestic normalized average pricing rose from 6.8 yuan (about $1) in late 2025 to 13.46 yuan (about $1.90) in 2026; the report’s normalization basis matters when comparing providers.
  • Average first-token latency also increased to 1.26 seconds, which the report attributes partly to larger models, longer contexts and heavier demand rather than simple service deterioration.
  • Nearly half of monitored models reportedly support one-million-token contexts, while 95.7% support at least 200,000 tokens.
  • The report finds substantial performance variation for the same open model across providers, making deployment and serving optimization a competitive layer beyond model weights.
Open at IT Home Ask Claude ChatGPT

Huang says Nvidia is open to buying memory from China’s CXMT

# Analysis IT Home 10m

Nvidia CEO Jensen Huang said the company is open to working with CXMT and called China exceptionally strong in manufacturing.

More
  • The comment is not a purchasing commitment, but it signals that a leading US chip company is willing to consider Chinese memory suppliers as part of a global supply chain.
  • CXMT’s relevance is strategic: memory is a critical input for AI infrastructure, and any commercial relationship would be shaped by product qualification and US-China restrictions.
  • Huang’s broader remarks frame Chinese manufacturing capability as something competitors must take seriously, rather than as a marginal low-cost alternative.
Open at IT Home Ask Claude ChatGPT

Chinese security startup tests token-based pricing for AI penetration testing

# Analysis 36Kr 2h

ChainReactors raised a seed round of several million yuan and is selling an AI-driven security-testing platform through open-source community channels and usage-based pricing.

More
  • Its agent system is designed to search attack paths without fixed roles or workflows, while its commercial service aims to cover a full penetration-test lifecycle.
  • The company says customers bring their own model tokens and pay the platform based on usage, an attempt to replace custom project delivery with a more scalable product model.
  • A survey of 100 trial users reportedly found about 70% open to token-based billing, but pricing, model-consumption costs and performance across industries remain unsettled.
  • The strategy trades traditional relationship-led sales for open-source adoption and benchmark visibility; the startup says current customers largely came through its community.
  • Automated testing is dual-use: authorization boundaries and safeguards will matter as the tool lowers the skill threshold for offensive security.
Open at 36Kr Ask Claude ChatGPT

MiniMax is selling token subscriptions through Tmall

# Analysis TMTPost 21m

MiniMax and other Chinese AI vendors are turning model-call quotas into retail subscriptions, making tokens easier for consumers to compare and buy.

More
  • The model is analogous to prepaid mobile data: vendors bundle usage into monthly or annual tiers, while Tmall’s AI hub gives buyers a shared storefront across providers.
  • The shift is strategic as well as commercial: AI firms gain consumer discovery and brand exposure beyond their established developer and enterprise channels.
  • The article says subscriptions commonly use rolling time windows and weekly quotas, a way to manage peak compute demand while packaging capacity for retail.
  • A standardized storefront could make token usage feel like a consumer utility, but demand, unit economics and retention are not demonstrated by storefront presence alone.
Open at TMTPost Ask Claude ChatGPT

Epic says Chinese developers are central to its next Unreal Engine push

# Analysis TMTPost 46m

At an event in Shanghai, Epic’s Tim Sweeney said China accounts for nearly 20% of Unreal Engine’s global monthly users and that Epic is sharing training material with Chinese AI labs.

More
  • Epic says it is providing leading AI labs, including five Chinese companies, with Unreal Engine source code, documentation, content and reinforcement-learning examples to improve agent use of 3D tools.
  • The technical aim is not just code generation: models need to manipulate real-time 3D editors and produce coherent scenes, a more structured tool-use problem than generating images.
  • Sweeney cited internal tests in which AI accelerated a store-client rebuild and enabled a developer to make a 3D scene in a day, but these are company anecdotes.
  • The arrangement gives Chinese labs access to substantial engine-specific training data while Epic may benefit from agents that lower the cost of creating content for its ecosystem.
  • The claims about exponential capability gains and near-term breakthroughs remain forecasts; the valuable signal is Epic treating model training data as an ecosystem investment.
Open at TMTPost Ask Claude ChatGPT
Source ledger
ChinAI 403 0/0 HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk 200 1/20
Pekingnology 200 0/20
Ginger River Review 200 0/20
Sinocism 200 1/20
QbitAI 200 3/10
36Kr 200 10/30
Solidot 200 0/20
Leiphone 200 15/20
GeekPark 200 7/30
IT Home 200 12/60 +37 over cap
SCMP 200 10/50 +7 over cap
Initium 200 4/15
Jiemian 200 8/30 +20 over cap
China Digital Times 200 1/15
The Wire China 200 1/50
Techmeme 200 2/15
Hacker News 200 0/20
V2EX 200 8/50 +35 over cap
Zhidx 200 10/20
InfoQ China 200 8/20 +10 over cap
New Wisdom 200 8/15 +3 over cap
TMTPost 200 8/20 +11 over cap