China Desk

China News, Summarized 25 Sep 2026 34 stories archived day

Today's brief

Beijing wrote Taiwan into the readout and the White House left the page blank

Xi asked Trump to change one verb

The ask was four words long. In Thursday's meeting Xi Jinping said he hoped the US would "adhere to the correct position of opposing 'Taiwan independence'" — a move from Washington's long-standing not supporting to opposing, which analysts told SCMP would also imply further restraint on arms sales to the island.

Only one side put it in writing. The single public mention of Taiwan after the closed-door talks came from Beijing's foreign ministry statement; the White House did not immediately offer its own account. Taipei's officials responded cautiously, unsure what was actually said about them.

The rest of the day was 1945. Trump called China a World War II ally; both men kept returning to the joint fight against Japanese aggression, Xi describing bonds "forged in blood and fire". That script was written for Tokyo, three days after Japan's prime minister landed at JFK to a near-empty apron.

Then Trump produced "gung-ho" as a Chinese phrase. SCMP's gentle correction: it is neither everyday Chinese nor a classical idiom — and puzzled viewers at home would look it up on Baidu, "or increasingly Doubao", ByteDance's chatbot and now the country's most-used one. The default reference tool changed hands and nobody announced it.

On AI, Xi asked for the opposite of the trend. Per the ministry readout, "the right thing to do is to draw on each other's strengths, not guard against each other," plus risk dialogue, misuse prevention and human control, via SCMP. Anthropic's account sweep, meanwhile, keeps decoupling the two countries one invoice at a time.

Go deeper on this section: ClaudeChatGPT

Live fire on the 28th, a licence for an AI film

Zhuhai's maritime authority closed part of the South China Sea. Live-fire exercises on 28 and 30 September, 8am to 5pm, entry prohibited, per Jiemian. Days after the state dinner. The southern Yellow Sea closure I flagged Wednesday still runs daily to 23 October, ending three days before the Fifth Plenum.

And the state has decided AI cinema is an industry. Local governments are dangling compute vouchers, rent waivers and dedicated funds at AI filmmakers, Reuters reports. CCTV puts the cost of an AI short drama in the first half at a few hundred yuan a minute, down from 5,000 yuan (~$700). Bona Film Group has the first theatrical licence for a studio AI feature.

Same article, two different leads. Reuters frames the subsidies as part of the national AI push. The Chinese-language relay on Solidot leads on overcapacity — solar, EVs, robots, now this — and on DataEye's count: 221,900 new AI videos on Douyin in the first half, 1,055 of them past 100m views.

THE TELL: Mid-Autumn Festival box office passed 100m yuan (~$14m), and the top title was a re-release of Avengers: Endgame at 35.87m yuan (~$5.05m), per IT Home — 185 minutes with four new ones, seven years after the original took 4.27bn yuan.

Go deeper on this section: ClaudeChatGPT

Alibaba gave away the thing that locks customers in

Wu Yongming showed the chip, then the stack. Alibaba's chief executive presented T-Head's new Zhenwu V900 at the Apsara conference on 22 September as China's most powerful AI chip, 3x the previous M890. A day later the chip unit widened the open-sourcing of SAIL, its CUDA substitute: the PyTorch adaptation, a source-migration tool, a Triton port, DeepGEMM and FlashAttention ports.

The moat was never the compiler — it's the migration bill. Lu Shenghua, T-Head's senior director for software ecosystem, says the first thing customers ask is how much moving costs, and that the ones who want custom kernels won't hand over their algorithms for the vendor to tune. Open the code and both problems dissolve, per QbitAI. Zhenwu now claims 650+ customers; Xiaohongshu built its own migration-and-kernel-tuning agent on top.

Huawei put its superpod on general sale the same week. Atlas 950 SuperPoD: 1,024 NPUs under one memory address space, 2.1x chip-to-chip bandwidth, 1.5–1.7x training throughput, sub-10ms inference latency, and an orthogonal backplane that deletes 6,000 cables per cabinet and 50km of copper per system, per Zhidx. The air-cooled 850E does 32 to 768 cards with no liquid retrofit.

And a third-party operator got 6.87x out of PCIe cards. METASTONE's Meta-Infer took DeepSeek-V4.1-Flash from 1,932 to 13,274 input tokens/s on eight PCIe-only GPUs with no change to weights or structure — restoring a sparse-MLA prefill fast path, replacing slow FP8 kernels, rewriting collectives for PCIe bandwidth, via QbitAI. A B300 is still 4.9–5.6x that box. The gap between "runs" and "serviceable" is software, and that is where the arbitrage lives.

Go deeper on this section: ClaudeChatGPT

Chinese tech press reads Opus 5.5 as a DeepSeek paper

A TMTPost column filed both new frontier models as imports. Claude Opus 5.5 burns 119,000 output tokens per question at max effort, 84,000 of them thinking, against GPT-6 Astra's 27,000 — test-time compute, which the piece traces to Yao Shunyu, author of Tree of Thoughts and ReAct. GPT-6 Luna's $0.01 per million cached-input tokens, 90% off, it traces to DeepSeek's disk context caching from 2 August, same 90%.

The honest core of the argument is the incentive. MLA, disk caching, sparse attention and multi-token prediction came out of not having enough cards; whoever has the most has the least reason to invent them. The dishonest part is the word borrowing — Yao is a US-lab researcher, at OpenAI after Princeton. The flip worth noting is the self-image: Silicon Valley productises, China defines.

Users are voting with their own money. A V2EX poster who swore off Claude after an account ban abroad says Opus 5.5 found a flaw in his experiment's statistics that Codex's models kept talking around. Sentiment, not benchmark.

Go deeper on this section: ClaudeChatGPT

Four thousand seven hundred girlfriends at a dollar forty each

Anthropic's GTG-15001 disclosure is an industrial process. One operation ran 20-plus dating apps on 4,700+ AI female accounts, sent 2.36m messages and worked at least 25,000 targets, with roughly three AI personas per human — the human appears only when a mark insists on video, per TMTPost.

The unit cost is the whole story. Under 10 yuan (~$1.40) per fabricated lover: a generated face for pennies, a cloned voice about 2 yuan, real-time face swap a few yuan, two weeks of chat a few yuan more. The FBI's first AI-tagged tally is $893m, of which $19m is romance fraud. Shanghai police arrested 77 people in January over one app cluster.

The legal twin is bigger. The same piece counts 4,300+ AI companion apps on the market. Loneliness is being monetised on both sides of the line; the criminals just skip the app-store cut.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • The Anthropic story went linguistic. Following Thursday's ban wave on AWS and Microsoft Greater China accounts: New Wisdom reports Chinese-language users hitting web rate limits when replies contain Chinese characters, here. Claim, not finding — and a different mechanism from the ownership rule, if it holds.
  • Liang Wenfeng's signature list is now 11 papers. Following Tuesday's DSec sandbox paper: Leiphone counts three years of the DeepSeek founder's name appearing on MLA, DeepSeekMoE, native sparse attention, mHC, DSpark and DSec — and never on the flagship model reports, here. Its reminder of the economics: Coder-V2 at $0.14/$0.28 per million against GPT-4 Turbo's $10/$30.
  • The dexterous-hand bottleneck got numbers. Unitree's Dex5-S hand, 21 September: 22 active degrees of freedom, human-palm size, from 39,900 yuan (~$5,620). The TMTPost map has the real constraint — roughly 500,000 hours of quality physical-interaction data worldwide against the 10m hours a general model needs, and a Stanford result where baselines scoring 80–96% in simulation land at 3–38% on real five-finger hands. One correction to that piece: Unitree's 61bn yuan was its offer-price valuation, not its first-day cap; it closed 19 August up 460% near 342bn yuan (~$48bn).
  • China now has its own Index. Mifeng has shipped 20,000 head-and-gripper capture rigs, claims 1m hours of data across 22 scene types, and paid one beta contributor over 5,000 yuan (~$700) in a month; Tencent Robotics X and Ant's Lingbo are buyers, per QbitAI. Figure's crowdsourcing app has 16m videos and $15m paid out — but sells to nobody, because it feeds Figure.
  • Where two of Huawei's model people went. Following Friday's robot-leadership exits: Li Yin, former CTO of Huawei Cloud's Pangu models, and Zhang Hanwang, former Huawei multimodal chief scientist and an NTU professor, raised seed and angel rounds at a $500m valuation to pretrain a "large physics model", via QbitAI. Their model ranks first on trajectory accuracy in World Arena 2.0's first track — a leaderboard, not a product.

Go deeper on this section: ClaudeChatGPT

The river · 34

Alibaba opens more of the software stack behind its AI chips

# QbitAI

Alibaba's chip unit is expanding the open-source T-Head SAIL stack around its Zhenwu accelerators, targeting the software portability and optimization barriers that keep customers tied to established GPU ecosystems.

More
  • SAIL connects frameworks such as PyTorch to Zhenwu hardware through migration tools, compiler and runtime components, acceleration libraries, communication libraries and debugging tools.
  • The newly cited projects include PyTorch adaptation, source-migration tooling, Triton support, and ports of DeepGEMM and FlashAttention; TensorFlow, JAX and additional communication components are still in progress.
  • Alibaba says Zhenwu serves more than 650 customers across over 20 industries, while adapted Qwen, DeepSeek and Kimi models on ModelScope had accumulated more than 348,000 downloads by September 2026.
  • The central business problem is migration cost: customers need to move existing models, operators and tuning knowledge without restarting their software stack or accepting a large performance penalty.
  • Open source also changes the division of labor. Customers such as Xiaopeng, Ant Group and Xiaohongshu can inspect and modify the stack, turning their workload-specific optimizations into ecosystem assets.
  • Xiaohongshu reportedly built an agent for model migration and operator optimization on top of SAIL, an early example of AI assisting the porting process itself.
  • This resembles the CUDA strategy in function, but the durable advantage will depend on fast support for new models, stable tooling and enough outside contributors to make Zhenwu more than a vendor-specific island.
Open at QbitAI Ask Claude ChatGPT

Huawei's Ascend 950 turns 1,024 NPUs into a commercial supernode

# Zhidx

Huawei launched a 1,024-NPU Ascend 950 supernode built around unified addressing and high-speed interconnects, targeting large-model training and inference without Nvidia hardware.

More
  • The Atlas 950 SuperPoD links 1,024 NPUs through more than 1,300 interconnect modules, 4,000 optical modules and a claimed 19 kilometers of cable tray, making system engineering as important as chip performance.
  • Huawei says its interconnect provides unified memory addressing across the node, with 2.1 times the bandwidth of the prior generation so remote memory access resembles local access.
  • Thermal and reliability design are central: each NPU module is reported at 950 watts, while cold plates, liquid-cooled optics and link-level retransmission address heat and optical-link failures.
  • The company claims 1.5-1.7 times higher training efficiency, two-to-three times better inference performance and sub-10 millisecond latency, but these are vendor-reported comparisons without workload definitions here.
  • The strategic significance is software and systems portability under export controls: a commercially deployable supernode can matter more than a standalone accelerator benchmark if Chinese customers can run changing models on it.
  • The unresolved risk is ecosystem depth, including kernel support, model portability, developer tooling and sustained cluster uptime outside Huawei's reference deployments.
Open at Zhidx Ask Claude ChatGPT

China and the US can manage rivalry without making peace

# SCMP

Chinese coverage frames repeated Xi-Trump recoveries as evidence that rivalry can be stabilized, even as tariffs, export controls and blacklists remain unresolved.

Also: Jiemian

More
  • The analysis argues that both governments repeatedly step back from rupture after high-tension summits, suggesting crisis management rather than a durable settlement.
  • Its useful signal is the difference between tactical stabilization and strategic convergence: trade and technology disputes can pause without disappearing.
  • For technology executives, this implies that export controls and supply-chain restrictions may be renegotiated at the margin while remaining durable planning assumptions.
  • The piece treats leader-level diplomacy as a pressure-release mechanism, but the supplied excerpt does not establish any new agreement beyond that broader pattern.
Open at SCMP Ask Claude ChatGPT

DeepSeek's new sandbox shows where large-agent training is heading

# Leiphone

A DeepSeek paper describes DSec, a production sandbox system that can run 3 million isolated environments daily, making reproducible agent training and containment a core infrastructure problem.

More
  • The paper presents DSec as an internal platform for large-scale agent reinforcement learning and evaluation, rather than as a model architecture change.
  • The reported system can create roughly 3 million sandbox environments per day across a production unit spanning 160 CPU nodes.
  • Strict permission isolation is designed both to prevent agents from gaming an evaluation and to reproduce each environment's feedback precisely.
  • The broader signal is that agent research is becoming an infrastructure discipline: state management, environment reset, observability and containment matter alongside model quality.
  • The article places Liang Wenfeng's authorship on a sequence of foundational systems papers covering mixture-of-experts routing, sparse attention, speculative decoding and agent sandboxes.
  • If the reported scale is representative, the competitive advantage is not just training compute but the ability to generate, replay and score huge numbers of controlled interactions.
Open at Leiphone Ask Claude ChatGPT

Agibot delivers its 20,000th robot to a theme park

# Leiphone

Agibot delivered its 20,000th humanoid robot and deployed more than 300 at a Chinese theme park, offering one of the clearest tests yet of embodied AI as an operating business.

More
  • The first deployment places more than 300 robots across over 100 interaction points covering guiding, education, retail, hotels, sports, performances, and companion experiences.
  • The park is expected to scale the relationship to thousands of robots, turning a public venue with large, variable crowds into a long-duration deployment environment.
  • Agibot's 20,000-unit milestone follows an earlier industrial delivery, letting the company claim two different productivity markets: factory labor and visitor services.
  • The system combines motion, interaction, and task intelligence, with a platform for customizing voice, expressions, persona, actions, and appearance for the park's characters.
  • The reported architecture uses network slicing, local edge equipment, and security isolation to separate visitor traffic from robot operations, while still relying on cloud connectivity for some workloads.
  • The robots are designed for ten-hour endurance, compared with an industry average of about three hours cited by the company, and use 360-degree sensing and compliant control around people.
  • The decisive test is operational economics, not a stage demo: charging, maintenance, safety, scheduling, connectivity, and daily uptime will determine whether hundreds of robots can produce repeatable service value.
Open at Leiphone Ask Claude ChatGPT

Xi and Trump discuss AI, Taiwan and a fragile trade truce

# SCMP

Chinese coverage says Xi urged cooperation on AI risks and human control while also pressing Trump on Taiwan during a summit designed to prevent trade tensions from escalating.

More
  • Beijing’s account frames AI as an area where the US and China should draw on each other’s strengths, continue risk dialogue and prevent misuse.
  • The same meeting carried a harder security message on Taiwan, showing how cooperation language on AI can coexist with sharp demands on sovereignty.
  • The summit’s pageantry matters because both governments are using leader-level contact to stabilize competition without resolving the underlying technology and security disputes.
  • For technical executives, the implication is not relaxed rivalry: dialogue may reduce accident risk while export controls, supply-chain restrictions and model governance remain contested.
  • The supplied report does not document a signed AI agreement, so the practical outcome is continued communication rather than a new joint regime.
Open at SCMP Ask Claude ChatGPT

A Chinese quantum magnetometer tracks a subway train underground

# SCMP

Researchers at the University of Science and Technology of China demonstrated a pencil-sized, five-watt quantum magnetometer by tracking a subway train beneath a city street.

More
  • The device detects the magnetic disturbance of a moving train, a practical proxy for the harder task of locating metal objects such as submarines.
  • Its reported size and five-watt power draw matter more than the demonstration's theatrical setting: they suggest quantum sensing can move toward portable deployment instead of laboratory-only instrumentation.
  • The source identifies the team and a study published in a Chinese physics journal, but the excerpt does not provide range, noise floor, localization error or comparison with conventional magnetometers.
  • Potential uses include underground mapping, infrastructure monitoring and defense sensing, though the submarine application remains an intended use rather than a demonstrated capability.
  • The key engineering question is whether the sensor maintains sensitivity in urban electromagnetic noise and can be calibrated and networked at scale.
  • A subway test is therefore an encouraging field validation, not evidence that a handheld device can already detect submarines operationally.
Open at SCMP Ask Claude ChatGPT

Xi asks Trump to oppose Taiwan independence explicitly

# SCMP

Xi Jinping pressed Donald Trump to adopt a clearer public position against Taiwan independence, a request that could constrain US arms policy and strategic ambiguity.

More
  • According to the reported Chinese foreign ministry account, Xi asked Washington to adhere to what Beijing calls the correct position of opposing Taiwan independence.
  • This is more consequential than summit ceremony because it targets the wording and policy space that underpin US deterrence and arms sales to Taiwan.
  • A public US commitment would represent a meaningful shift from longstanding ambiguity; the supplied account does not say Trump accepted it.
  • The request shows Beijing using leader-level diplomacy to move the US from general one-China language toward an explicit political restriction.
  • For technology and defense planners, Taiwan remains the point where trade, chip supply chains and military escalation risk converge.
Open at SCMP Ask Claude ChatGPT

China pushes AI video toward an industrial-scale production model

# Solidot

Chinese local governments are courting AI video studios as a new industrial sector: costs have fallen sharply, output is exploding, and regulators have begun approving AI-made films for theaters.

More
  • The source says AI short-drama production costs fell from about 5,000 yuan per minute to a few hundred yuan, roughly from $700 to under $100, during the first half of 2026.
  • More than 221,900 new AI videos reportedly appeared on Douyin in the first half of the year, but only 1,055 exceeded 100 million views, a sharp sign of potential overcapacity.
  • Local governments and industrial parks are calling studios to bring AI film businesses to their regions, echoing China's earlier strategy of building capacity in solar, electric vehicles, and robotics.
  • China's film regulator granted a theatrical release license to a 90-minute AI science-fiction film, described as the first large-studio AI film approved for cinemas.
  • The industrial opportunity is therefore paired with a familiar Chinese risk: cheap production and local subsidies can create enormous volume before durable audience demand, distribution economics, or creative quality are proven.
  • The important shift is from AI as a creator tool to AI video as a regional industrial policy, with local governments competing to own studios, pipelines, and production capacity.
  • If overcapacity develops, the surviving advantage will likely be rights, distribution, distinctive creative direction, and workflow integration rather than the ability to generate footage cheaply.
Open at Solidot Ask Claude ChatGPT

Alibaba's CT model brings esophageal cancer screening closer to clinics

# TMTPost

Alibaba's DAMO EAGLE model uses routine chest CT scans to flag esophageal cancer and precancerous lesions, with validation reported across 80,000 cases in three countries.

More
  • The model is designed for triage, not autonomous diagnosis: it identifies patients who should receive confirmatory endoscopy without adding an invasive screening step to everyone.
  • The report gives 90% sensitivity for esophageal cancer, 52.5% for precancerous lesions and 99.2% specificity, though these are reported study figures rather than a guarantee of clinical performance.
  • The key China-specific shift is from laboratory accuracy to clinical translation, with hospitals and provincial validation centers being assembled as regulatory test beds.
  • Remaining gates include prospective multi-center validation, medical-device registration, hospital-system integration, data authorization and a sustainable insurance or reimbursement code.
  • Alibaba's broader portfolio combines specialty cancer models with an open general imaging model, suggesting a platform strategy in which deployment generates more real-world data and workflow integration.
  • The report explicitly warns that a top-journal result is not equivalent to legal clinical authorization or a scalable commercial service.
Open at TMTPost Ask Claude ChatGPT

Unitree's dexterous hand shows why hardware is ahead of robot brains

# TMTPost

A Chinese analysis of Unitree's $5,620 Dex5-S argues that 22 active degrees of freedom do not solve the harder problems of tactile data, control and failure recovery.

More
  • The hand is positioned as a research and industrial procurement tool for universities, robot makers and labs rather than evidence that household robots are imminent.
  • The useful abstraction is a multiplicative capability stack: degrees of freedom, tactile resolution, control precision, training data and task distribution must all work together.
  • Flexible objects, occlusion, slip and recovery remain difficult because simulation usually assumes rigid bodies, clean sensing and perfect communication; the article cites large simulation-to-real performance drops.
  • The analysis estimates that high-quality real physical-interaction data is far scarcer than the millions of hours general-purpose embodied models may need, because force and tactile traces cannot simply be scraped from the web.
  • The strategic implication is that the next moat may be data-collection environments and deployment feedback loops, not another headline hardware specification.
  • Unitree's product is therefore a marker of an industry pivot: actuators are becoming purchasable while the general-purpose policy layer remains unproven.
Open at TMTPost Ask Claude ChatGPT

Former Huawei model leaders raise $70M for a physical-world foundation model

# AI QbitAI

Physical-AI startup Xirang Kaiwu says it raised several hundred million yuan, roughly $30-60 million, at a $500 million valuation to train a reusable large physics model for robots.

More
  • The founders previously held senior model roles at Huawei, including cloud-model leadership and multimodal research, and are betting that robotics needs a shared physical foundation model rather than a separate model for every body and task.
  • The proposed large physics model would learn from video, robot trajectories and real interactions, then transfer physical knowledge across robot bodies and environments.
  • The company argues that visual-language-action systems still require heavy retraining for each embodiment and deployment setting, limiting the economics of household and public-service robots.
  • Xirang reports first-place trajectory accuracy and second-place physical-adherence results on a World Arena benchmark, but the supplied account provides no independent validation or benchmark details.
  • The team also claims performance has improved as data and training scale increase, which is an early, company-reported signal for its hypothesis that physical intelligence has a scaling path.
  • The strategic bet is unusually capital intensive: large training runs, data collection and real-world deployment all compete for scarce funding, while the payoff depends on transfer working across embodiments.
  • If the approach works, the value shifts from bespoke robot integration toward a general model layer; if not, the company may simply add another expensive task-specific training loop.
Open at QbitAI Ask Claude ChatGPT

China's crowdsourced robotics-data market targets the next bottleneck

# AI QbitAI

Mifeng has launched a hardware-and-app platform for crowdsourcing human motion data, claiming 20,000 devices and 1 million hours of real-world recordings for robot training.

More
  • The platform uses head-mounted panoramic cameras, wrist cameras and lightweight tactile grippers to collect egocentric, hand and environment data outside laboratory settings.
  • Mifeng says its devices cover 22 scene categories and capture color, depth, inertial, tactile and audio modalities, with production capacity in the tens of thousands per month.
  • Its data engine reconstructs hand and body motion under occlusion, rapid movement and multi-person interaction; the company claims sub-centimeter reconstruction in difficult scenes, which remains a vendor claim.
  • A grading system assigns data to different uses, including imitation learning, general pretraining, error recognition and hard-case evaluation, rather than discarding every imperfect recording.
  • The proposed loop runs from collection through reconstruction, quality grading, training, deployment and failure-data return, making data operations part of the robot product rather than a one-time research cost.
  • Unembodied human data can teach general object, action and spatial relationships, but the article acknowledges that each robot still needs real-body post-training.
  • The business test is whether low-friction contributors can produce consistent, legally usable data at a cost below bespoke robotics-lab collection.
Open at QbitAI Ask Claude ChatGPT

The Xi-Trump summit produced stability more than a settlement

# Policy SCMP

Chinese coverage presents the Xi-Trump summit as a carefully staged peer-to-peer reset with trade and Taiwan on the agenda, but the concrete commitments remain limited and the structural rivalry intact.

Also: The Wire China, Initium

More
  • Xi made his first state visit to the United States in more than a decade and met Donald Trump repeatedly during a 43-hour trip, including formal talks, a ceremonial welcome and a state dinner.
  • The agenda covered tariffs, trade and Taiwan, while the reporting emphasizes the symbolic value of Washington treating Beijing as an equal major power.
  • The summit's practical value is continuity: both sides left with a reason to keep meeting, which could reduce the risk of an immediate return to last year's trade-war escalation.
  • The reporting does not identify a comprehensive settlement of the underlying disputes; the contrast between grand ceremony and few concrete deliverables is the central analytical point.
  • For companies, relative stability may improve planning around tariffs, export controls and supply chains, but it should not be mistaken for normalization.
  • The next test is whether follow-up talks produce enforceable measures rather than another cycle in which pre-summit threats are exchanged for post-summit calm.
Open at SCMP Ask Claude ChatGPT

Alibaba launches a full software and hardware stack for Qwen Office

# AI Zhidx

Alibaba's Qwen Office added enterprise context management, agent security controls, a $169 recording device and an $127 desktop robot while signing new corporate users.

More
  • Enterprise Context compresses and structures documents, chats, knowledge bases and operational updates so agents can retrieve current business state instead of relying on a giant undifferentiated context window.
  • The security center provides least-privilege access, sandboxing, approvals, interception and audit trails, which positions agent governance as a product layer rather than an afterthought.
  • The QwenNote A2 captures meetings and offline interactions as external context, while the Eva desktop robot uses a phone as its display and supports gaze, touch and voice activation.
  • The strategy links enterprise software, agent orchestration and hardware capture: more context improves task execution, while more tasks create pressure for permissioning and traceability.
  • Alibaba says companies are connecting procurement, sales, contracts and other workflows, but the supplied report gives no independent productivity or deployment reliability data.
  • The architectural bet is that proprietary enterprise context, not a larger general model, will become the durable control point for office agents.
Open at Zhidx Ask Claude ChatGPT

Changsha presents a central-China model for industrial AI

# Analysis Zhidx

Changsha launched a 100 billion yuan (about $14.1B) future-industries fund and showcased AI systems moving from screens into factories, hospitals, mines and logistics.

More
  • The fund's first 30 billion yuan (about $4.2B) targets AI, life sciences, quantum technology and advanced materials, pairing government capital with market-oriented venture management.
  • The city is building around existing industrial strengths rather than trying to imitate Beijing or Shenzhen: engineering machinery, manufacturing, mining and broadcast-content ecosystems provide deployment sites.
  • Reported 2025 figures include 850 billion yuan (about $120B) in AI-industry scale, more than 400 key firms and 5,200 PFLOPS of intelligent computing capacity, though these are official or local figures.
  • The exhibition's chain runs from data collection and labeling through foundation models, agents, domestic compute, medical AI, humanoid robots and industrial systems, illustrating an ecosystem strategy rather than a single champion.
  • The policy model is to align capital, compute, talent and industrial demand in one regional cluster; its weakness is that showcase deployments may not translate into repeatable commercial unit economics.
  • For US technical executives, Changsha is a useful example of how Chinese inland cities use public funds and incumbent manufacturing to create AI test beds outside the coastal hubs.
Open at Zhidx Ask Claude ChatGPT

Kimi turns browser actions into reusable agent skills

# AI QbitAI

Moonshot AI's new browser extension lets Kimi operate a user's Chrome or Edge session and record successful web workflows as reusable skills, pushing agents from reading pages toward repeatable action.

More
  • The extension can open pages, read content and click controls from a sidebar, while retaining the older bridge that lets a local agent operate the user's real browser.
  • Recording a successful sequence as a skill addresses a practical failure mode: agents otherwise need repeated prompt tuning for the same dynamic website.
  • The product is aimed at recurring tasks such as checking news and exporting data, not just one-off browsing assistance.
  • The design pairs with Kimi Code Desktop, which handles project files and an internal browser while the extension reaches external user accounts and websites.
  • The browser remains an unstable execution environment: page redesigns, dynamic loading and permission changes can still break a recorded workflow.
  • The strategic implication is that agent vendors are accumulating procedural knowledge at the interface layer, where the hard problem is reliable state and authorization rather than language generation.
Open at QbitAI Ask Claude ChatGPT

Xi's rare laugh becomes a viral moment during Trump's White House tour

# Culture SCMP

Chinese coverage highlights Xi Jinping's unusual laughter at Trump's autopen joke, while the US-facing story is more interested in the summit's optics and policy stakes.

More
  • The reported moment occurred during a White House tour and stood out because Xi is usually presented as restrained in public settings.
  • Chinese-facing coverage treats the laugh as a humanizing, shareable image inside a carefully choreographed state visit.
  • The contrast matters because leader-to-leader symbolism is part of the diplomatic message in China, where public demeanor can be read as evidence of rapport or control.
  • The scene says little about negotiations themselves; it is an optics signal that may help both governments present engagement without claiming a substantive breakthrough.
Open at SCMP Ask Claude ChatGPT

Taiwan reacts cautiously to Trump's warm welcome for Xi

# Policy SCMP

Taiwanese officials and observers remain uncertain about what Trump and Xi discussed, while Beijing's public statement foregrounds opposition to Taiwan independence.

More
  • The report says Beijing's statement was the only immediate public account to mention Taiwan after the leaders' private talks.
  • Taipei's caution reflects a familiar information problem: the island must infer strategic risk from language released by Beijing when Washington provides few details.
  • The contrast between a warm summit image and an explicit Chinese demand on Taiwan makes the meeting politically uncomfortable for Taipei even without a disclosed concession.
  • The source does not establish any change in US policy, arms sales or military posture, so uncertainty itself is the main signal.
Open at SCMP Ask Claude ChatGPT

US defense weaknesses may be handing Beijing intelligence gifts

# Analysis ChinaTalk

A ChinaTalk discussion argues that US forces are overextended in the Middle East, short on missiles and ships, and exposing capabilities that Beijing can study.

More
  • The podcast discussion says a large share of US naval forces and missile stockpiles is tied up in the Middle East, constraining Pacific readiness; these are the participants’ assessments, not independently verified figures here.
  • The “gifts for Beijing” framing points to captured drones, failed missiles, and visible operational patterns as intelligence opportunities for China.
  • The discussion also questions whether an AI branch modeled on special operations forces fits technology that should be pervasive across the military rather than isolated in one organization.
  • The strategic issue is force availability: a China contingency depends not only on platforms and software but on whether munitions, logistics, and command attention are already consumed elsewhere.
  • Because the item is a podcast discussion, its value is analytical framing and expert disagreement rather than a new factual disclosure.
Open at ChinaTalk Ask Claude ChatGPT

Beijing may decide which Latin American diplomat becomes UN chief

# Policy SCMP

After three inconclusive Security Council ballots, China is reportedly weighing candidates from Costa Rica and Guyana for the next UN secretary-general, highlighting Beijing's growing leverage in the selection.

More
  • The report says Beijing is considering Rebeca Grynspan of Costa Rica and Carolyn Rodrigues-Birkett of Guyana after three rounds of informal voting failed to produce a clear favorite.
  • China's position matters because the Security Council recommends a candidate before the General Assembly formally appoints the secretary-general.
  • The account is based on conversations with UN delegations rather than a public Chinese announcement, so the reported preference should be treated as provisional.
  • The choice would let Beijing shape the profile of the UN's next diplomatic center while preserving room to bargain with other permanent members.
  • The larger signal is that China is using multilateral institutions as an arena for influence, not merely reacting to US-led initiatives.
Open at SCMP Ask Claude ChatGPT

Unitree founder says millimeter-scale accuracy is robotics' bottleneck

# AI Jiemian

Unitree founder Wang Xingxing says robots now handle longer tasks and voice-generated motion, but reliable work still depends on closing errors of only a few millimeters.

More
  • Wang frames the bottleneck as execution accuracy rather than a lack of impressive demonstrations: a robot must repeatedly place, grasp or manipulate objects within a narrow tolerance.
  • Unitree says its systems can generate actions from spoken instructions and execute long sequences with some disturbance resistance, but current inference introduces a small delay.
  • The comments reflect a familiar embodied-AI split between high-level intent generation and low-level control, calibration and contact dynamics.
  • Wang expects broader arbitrary-action generation by the end of this year or next year, but the source offers no benchmark or production deployment data to support that timetable.
  • The three-meter passenger mech shown at the event is presented as an outdoor platform, suggesting Unitree is also testing whether large robots can find a practical niche beyond humanoid demos.
Open at Jiemian Ask Claude ChatGPT

ByteDance's Doubao is reorganizing around monetization

# AI TMTPost

A Chinese analysis says Doubao is shifting staff and systems toward transactions, office software and enterprise services as its huge consumer user base remains expensive to monetize.

More
  • The immediate dispute involved an 11-person reorganization, not the rumored halving of the conversation team; the source says three people left and other functions moved to commerce and work products.
  • The deeper move follows ByteDance's decision to fold Feishu product teams into Doubao and align go-to-market functions with Volcano Engine, making enterprise capability part of the AI unit's operating base.
  • The article cites Doubao monthly active users approaching 380 million while daily computing costs are estimated in the tens of millions of yuan, roughly $1.4 million to $5.6 million, against daily revenue below 1 million yuan, roughly $140,000.
  • Feishu is positioned as a monetization bridge: its enterprise accounts, documents, permissions and paid workflow can give Doubao access to customers that are less resistant to paying than consumer users.
  • This is a strategic shift from free conversational scale toward transaction commissions and enterprise agents, but the revenue and cost figures are attributed estimates rather than audited disclosures.
  • The reorganization therefore matters less as a staffing event than as evidence that Chinese AI platforms are confronting the gap between user growth and sustainable unit economics.
Open at TMTPost Ask Claude ChatGPT

Chinese components appear in Russia's faster attack drones

# Analysis Initium

A China-related supply-chain angle emerges as Russia deploys jet-powered attack drones with reported Chinese engines and communications hardware, reducing Ukraine's interception rate.

More
  • The report says Ukraine's interception rate against older propeller drones approached 90%, but fell to about 65% against faster jet-powered variants in a cited September attack.
  • Reported upgrades include thermal imaging, AI-assisted target locking, anti-jamming antennas and Chinese-made turbojet engines or mesh-network modems, though the exact provenance of each component is not fully established.
  • The strategic lesson is industrial rather than merely tactical: Russia is iterating designs directly through mass production and battlefield feedback, while Ukraine must develop interceptors that can handle higher speeds.
  • This is adjacent China coverage because the Chinese dimension is component supply into a foreign weapons program, not a Chinese military operation.
  • The figures and battlefield claims come through media and Ukrainian sources; the report does not establish Beijing's intent or state involvement in the supply chain.
Open at Initium Ask Claude ChatGPT

Unitree leads China's four-legged robot market, but the rankings disagree

# Signal TMTPost

Two market reports agree Unitree leads global quadruped-robot shipments, yet disagree sharply on volumes and the runner-up because one counts units and the other hardware revenue.

More
  • Counterpoint estimates about 35,000 global quadrupeds shipped in the first half of 2026, while IDC estimates more than 49,000; the gap of 12,000 to 14,000 units exceeds half of Unitree's reported 2025 sales.
  • Counterpoint ranks Unitree at 37 percent by shipments and AGIBOT's quadruped unit at 16.3 percent, while IDC ranks by hardware revenue and puts Deep Robotics second at 9.2 percent.
  • Deep Robotics illustrates the distortion: its industrial quadrupeds reportedly average 287,500 yuan, roughly $40,500, versus Unitree's 30,300 yuan average, roughly $4,300.
  • The reports also define market size differently: Counterpoint includes sensors, software and services and estimates nearly $700 million for 2025, while IDC counts hardware and estimates more than $490 million.
  • Unitree's lead is therefore strongest in accessible unit volume, not necessarily in industrial revenue; Deep Robotics and AGIBOT may be stronger in higher-value applications.
  • The broader lesson is that robot-market rankings are meaningless without a declared unit, revenue, product and application boundary.
Open at TMTPost Ask Claude ChatGPT

Chinese CTA strategies suffer heavy losses in extreme markets

# Signal TMTPost

China's futures trading competition saw aggregate losses of 6.48 billion yuan (about $913M) by July 31, with only 22.25% of accounts profitable.

More
  • The report says trend-following and crisis-hedging strategies were hit by abrupt reversals, limit moves and synchronized sector rotations rather than a simple lack of volatility.
  • That matters for Chinese markets because the usual story that managed futures provide crisis alpha is being challenged by feedback loops and delivery constraints.
  • The supplied analysis argues that mean reversion can fail when forced buying, low inventories or storage limits push prices farther from fundamentals.
  • Its proposed response is regime detection through volatility thresholds and asymmetric hedges, but those are the author's analytical recommendations, not evidence of a regulator-approved framework.
Open at TMTPost Ask Claude ChatGPT

Oppo bets on on-device memory as phone AI's next moat

# AI GeekPark

Oppo says model differences are narrowing and is focusing instead on on-device models, persistent personal memory and an AI operating layer that can act on user intent.

More
  • The architecture separates cloud foundation models from the phone-specific layer: Oppo wants to own the device model, memory system and orchestration while consuming an open cloud-model ecosystem.
  • The difficult component is not storing data but maintaining a useful memory body that can update, correct, forget and rank events, such as distinguishing a current flight from stale travel records.
  • On-device inference reduces privacy exposure and cloud cost but moves the constraint to battery, thermals, bill of materials and hardware-software co-design.
  • Oppo reports that a system memory feature improved background retention by 55.6% on comparable devices, though the metric is vendor-provided and not a measure of assistant intelligence.
  • The strategy reflects a broader phone-maker response to model commoditization: own the user's long-term context and control plane rather than compete directly with frontier-model labs.
Open at GeekPark Ask Claude ChatGPT

Cheaper, lower-sugar mooncakes are winning China's holiday market

# Culture Jiemian

Chinese mooncake demand is shifting from gift boxes to self-consumption: average prices fell while lower-sugar, simpler products gained share, revealing a broader consumer reset.

More
  • Retail monitoring cited by the report puts the average price per 100 grams at 6.48 yuan in 2024, 5.84 yuan in 2025 and 4.89 yuan in 2026, roughly $0.91, $0.82 and $0.69.
  • Self-use and family sharing accounted for 50.4 percent and 55.9 percent of purchases in the cited survey, while traditional gifting was 42.5 percent.
  • Simple gift boxes priced around 68 to 200 yuan, roughly $10 to $28, reportedly captured more than 60 percent of sales, while individually wrapped and freshly baked products also gained ground.
  • The shift suggests gift-giving norms and discretionary spending are weakening, while consumers still preserve the festival ritual by buying smaller and more functional products.
  • Lower sugar and low-GI claims show brands responding to health concerns and self-consumption, but the article does not establish whether these claims translate into measured health benefits.
Open at Jiemian Ask Claude ChatGPT

China's AI micro-drama boom hits a brutal economics wall

# Culture TMTPost

More than 380,000 AI micro-dramas launched in China in the first eight months, while active production vendors fell 42 percent in one quarter as platform payouts weakened.

More
  • The cited broadcaster data imply extraordinary supply growth: more than 380,000 AI micro-dramas appeared in eight months, but active contractors fell from 1,216 to 698 in the first quarter.
  • The industry's unit economics resemble an oversupplied content market: cheaper generation raises output faster than audience revenue, pushing teams into loss or withdrawal.
  • The analysis compares the moment with Japanese animation's post-bubble crisis, when studios expanded beyond episode sales through shared investment and merchandise-linked production committees.
  • China's AI micro-drama ecosystem already has export distribution and platform infrastructure, but the article argues it still lacks a durable mechanism for turning attention into diversified IP revenue.
  • The comparison is an argument, not a forecast; the key test is whether a few durable characters and franchises emerge from the current flood of low-cost output.
Open at TMTPost Ask Claude ChatGPT

The prompt-engineer boom shows how AI reshapes job boundaries

# Culture TMTPost

Chinese analysis argues that prompt engineering was a real but narrow skill that quickly became a product feature, offering a useful labor-market lesson for AI organizations.

More
  • The article compares prompt engineers with earlier clerks, typists and human computers whose specialized value declined as tools became usable by everyone.
  • It cites a study of 20,662 LinkedIn postings that found only 72 prompt-engineer roles, suggesting public fascination far exceeded the occupation's actual scale.
  • Model vendors began absorbing prompt construction through prompt generators, richer context handling and reusable skill packages, reducing the scarcity of manual instruction design.
  • The durable role is shifting from writing clever prompts to designing evaluations, data, tools, permissions and workflows around models.
  • For Chinese employers and workers, the lesson is that AI may create visible job titles briefly while redistributing the underlying tasks across ordinary engineering and operations roles.
Open at TMTPost Ask Claude ChatGPT

AI romance scams scale through fake personas and automated chat

# Culture TMTPost

A reported scam network used more than 4,700 AI female accounts across 20 dating apps to send 2.36 million messages, while Chinese police have exposed similar operations.

More
  • The reported workflow combines generated faces, automated conversation, human video handoffs and small recurring charges, making fraud more scalable than traditional one-to-one impersonation.
  • The Chinese connection is concrete: police cases in Shanghai involved AI-generated personas and large financial losses, while domestic virtual-companion apps show how the same emotional infrastructure can also be commercialized legally.
  • The economics are alarming because image, voice, live face replacement and model calls can be assembled at very low marginal cost, letting operators relaunch after an API ban.
  • This is also a social signal about loneliness and monetized intimacy: the product is not only deception but continuous behavioral analysis of a target's emotional responses.
  • The figures are attributed to Anthropic, police disclosures and industry commentary; the report does not provide an independently audited loss total for the network.
Open at TMTPost Ask Claude ChatGPT

A Hangzhou vocational school's drone program doubles in size

# Culture Jiemian

A Hangzhou technical school planned 84 drone students but enrolled 170 after its first cohort found jobs before graduation, offering a concrete signal about China's emerging skills economy.

More
  • The program grew from three classes to five, while its first 17 students reportedly secured jobs in surveying, emergency response, education and hardware development.
  • The school opened the program in 2021, early among Chinese vocational institutions, and its students combine flight operations with programming, maintenance and modeling.
  • This is labor-market evidence rather than a startup forecast: employers are hiring technicians who can operate and maintain drones, not only engineers who design them.
  • The appeal also reflects a Chinese education pathway outside conventional universities, where visible technical skills and competition results can attract students who might otherwise choose general high school.
  • The source does not show national demand or wage data, so the enrollment surge should be read as a local signal of opportunity, not proof of a nationwide shortage.
Open at Jiemian Ask Claude ChatGPT

Chinese analysis says Anthropic and OpenAI are borrowing DeepSeek ideas

# Analysis TMTPost

A Chinese tech analysis portrays Claude Opus and GPT-6 as adopting ideas associated with Chinese researchers: test-time reasoning and cheaper context reuse.

More
  • The article contrasts more inference-time computation with more efficient reuse of long prefixes, framing model progress as a shift from pretraining scale toward serving architecture and reasoning budgets.
  • It links the first idea to test-time search and action research and the second to DeepSeek's reported disk-based context caching, where repeated prefixes are stored outside expensive accelerator memory.
  • The important China angle is attribution and reception: Chinese coverage treats DeepSeek's efficiency work as an influence on the global model race rather than as a local curiosity.
  • The claims are interpretive and the supplied text uses provocative language about copying; they do not establish that either foreign company directly copied DeepSeek's implementation.
  • For infrastructure planners, the underlying issue is concrete: long-context agents make KV-cache placement, cache hit rate and adaptive inference effort major components of cost.
Open at TMTPost Ask Claude ChatGPT

Wondershare pitches an end-to-end AI video stack for global content

# Culture Zhidx

Chinese software maker Wondershare launched an AI video suite spanning film, advertising, commerce, music and editing, alongside a plan to export more than 10,000 AI productions.

More
  • The product strategy is vertical integration across script, character assets, storyboards, generation, editing and collaboration rather than selling a single image or video model.
  • Wondershare is using Hunan's broadcast and media ecosystem as an application base while targeting international distribution, a pattern common in China's AI-content industrialization push.
  • The company says it is committing billion-yuan-level resources with partners, but the item does not specify the exact funding split, production quality or audience demand.
  • The strategic risk is the same as the wider Chinese AI-video market: cheap generation can create overwhelming supply before durable distribution and rights models exist.
Open at Zhidx Ask Claude ChatGPT