China Desk — 18 Aug 2026

Geely hands listed-company leadership to its next generation

# QbitAI

Geely replaced its listed-company chair and CEO while reporting record first-half revenue and profit, signaling a move from founder-led control toward institutional management.

Li Shufu remains chair of the wider holding group and honorary chairman, while An Conghui becomes chair and Gan Jiayue becomes CEO of the listed automaker.
The company reported 1.736 trillion yuan (roughly $245 billion) in first-half revenue and 96.8 billion yuan (roughly $13.6 billion) in core attributable profit, up 46 percent.
The timing matters: succession during improving performance suggests an organizational transition rather than an emergency response to failure.
The new leaders are long-serving internal operators associated with electric vehicles, brand integration, supply-chain management and operating efficiency.
Geely’s reported overseas sales rose 158 percent and new-energy exports 585 percent, making the leadership change relevant to a more international growth phase.
The strategic test is whether the new structure can preserve founder-driven speed while imposing enough process discipline for a multi-brand global company.
A strong first half does not prove the succession will work; currency effects, impairment and intense Chinese auto competition remain material risks.

Open at QbitAI Ask Claude ChatGPT

Chinese startup uses AI to optimize domestic-chip infrastructure

# QbitAI

Tsinghua-linked startup Qingmang says its self-optimizing infrastructure cuts complex operator tuning from about two weeks to one hour on Chinese chips.

The system closes a feedback loop across operators, compilers, memory scheduling, frameworks and inference runtimes rather than generating a single optimized kernel.
Its architecture separates autonomous optimization, which chooses the best approach for the current workload, from autonomous evolution, which stores experience for later models, chips and environments.
The company reports roughly 336-fold faster operator tuning, about 30-fold faster full-stack model adaptation and up to 30-fold acceleration for selected operators; these are company-reported results.
This addresses a central Chinese constraint: heterogeneous domestic accelerators require repeated porting and tuning, while access to leading foreign chips remains restricted.
Automating the software layer could raise utilization of available Chinese compute without requiring a new fabrication breakthrough.
The claims need independent benchmarks across models, chips and workloads; a few best-case operators would not establish a general infrastructure advantage.
If reproducible, the approach makes the compiler and runtime stack a strategic asset in China’s effort to build a self-reliant AI ecosystem.

Open at QbitAI Ask Claude ChatGPT

China's AI bubble may be industrial policy by another name

# ChinaTalk

Chinese financial language calls the AI funding rush a pass-the-parcel bubble, while this analysis argues that inflated valuations and mass company exits may be deliberate costs of state-led industrial policy.

The article says Chinese investors are applying the phrase to an early-stage market where valuations outrun revenue, money clusters in fashionable subsectors and many firms lack a path to profit.
It cites roughly 80,000 Chinese AI companies dissolving or deregistering over two years, but the excerpt does not define the database or distinguish failed startups from normal corporate cleanup.
The counterargument is historical: Beijing used concentrated state capital and overcapacity to build solar and electric-vehicle industries, so a bubble can coexist with the creation of a globally competitive supply base.
The analysis says national AI funds and hundreds of billions of yuan have accelerated capital formation; those figures and policy details need verification before being treated as a precise spending total.
The key question is allocation quality: can public money fund infrastructure, talent and durable companies, or will it mainly recycle valuations among investors before demand materializes?
For executives, the implication is that Chinese AI competition may remain intense even after a financing correction, because industrial policy can preserve capacity and strategic suppliers after private capital retreats.
The article's conclusion is an argument, not a forecast: the same mechanism that creates waste can also produce scale, lower costs and a few globally important firms.

Open at ChinaTalk Ask Claude ChatGPT

Xi leads China's farewell to former premier Zhu Rongji

# SCMP

Xi Jinping led a formal farewell for former premier Zhu Rongji, whose state funeral and official remembrance underline how the Party selectively preserves the legacy of market-oriented reform.

Zhu served as premier from 1998 to 2003 and is remembered for steering major economic reforms and China's entry into the World Trade Organization, according to the report.
The attendance of the Party elite and flags at half-mast make this a state-managed historical message, not simply a private funeral.
Zhu's reputation as a forceful reform administrator gives the ceremony resonance amid current debates about growth, private enterprise, and the role of markets.
The leadership can honor his competence while avoiding an endorsement of every policy associated with his era; official remembrance is often carefully bounded in this way.
The event matters because elite funerals reveal which past leaders can be publicly invoked and what parts of their legacy the current system wants to stabilize.
The supplied account does not indicate any new economic policy, so the signal is political and symbolic rather than a policy reversal.

Open at SCMP Ask Claude ChatGPT

China's top general invokes Jiang's legacy to stress loyalty

# SCMP

China's top general uses Jiang Zemin's response to past crises to argue for Party unity and military reform, showing how the leadership is using historical memory to reinforce loyalty.

The article by Zhang Shengmin reportedly links the Soviet collapse and the first Gulf War to the need for political cohesion and military modernization.
The timing is deliberate: commemorations of Jiang's centenary give the current leadership a controlled way to invoke a predecessor while defining which parts of his legacy remain useful.
The emphasis on loyalty signals that military reform is being framed not only as technology and organization, but as political reliability under external pressure.
This fits a wider Chinese governance pattern in which historical episodes are converted into lessons about Party discipline and institutional obedience.
The story is important because Zhang is China's most senior general, so the argument is closer to an institutional signal than to a civilian academic retrospective.
Its practical meaning will depend on whether the rhetoric precedes personnel changes, training priorities, or new requirements for political work inside the armed forces.

Open at SCMP Ask Claude ChatGPT

US advisers urge Washington to treat data as an economic asset

# SCMP

A US-China advisory commission says the United States risks losing AI and robotics leadership unless it develops a national data strategy, drawing an explicit lesson from China's data commercialization push.

The report describes China's approach as treating large data collections as economic infrastructure rather than merely as privately held byproducts of online services.
The comparison is strategically important because data policy affects model training, robotics, industrial optimization and the ability to set technical standards, not only privacy compliance.
The US recommendation is not simply to collect more data; it implies national rules for access, stewardship, sharing and commercial incentives.
The account offers no detailed US policy design and does not establish that China's data assets are broadly accessible or high quality, so the comparison should not be mistaken for proof of Chinese advantage.
The story matters to technology executives because data governance can become industrial policy: permissions and interoperability may determine which companies can build durable AI systems.

Open at SCMP Ask Claude ChatGPT

China’s senior mode leaves millions of older users stranded

# Jiemian

A report finds that China’s senior-friendly interfaces often enlarge text without removing the complexity, advertising and navigation traps that keep older users dependent.

China had 161 million internet users aged 60 or above by June 2025, while more than 140 million domestic smartphones and smart TVs had reportedly undergone accessibility upgrades.
The mismatch is stark: accessibility coverage expanded, but smart-TV daily activation fell from 70 percent in 2016 to below 30 percent in 2022.
Complaints about senior modes focus on enlarged fonts layered over pop-up ads, confusing return paths and interfaces that still assume a digitally fluent user.
Industry standards have evolved from a simple mode requirement toward audio-visual, interaction, safety and health dimensions, but the report says product implementations remain largely cosmetic.
The problem is not just usability. A nominal accessibility mode can shift the burden to relatives, suppress device adoption and turn a regulatory compliance label into a misleading proxy for inclusion.
For technology companies, the lesson is architectural: older users need simpler state transitions, fewer commercial interruptions and recoverable flows, not merely a scaled-up presentation layer.
The story also shows how Chinese consumer technology is judged publicly: a celebrity complaint can rapidly turn a narrow product defect into a nationwide debate over whether firms are monetizing vulnerable users.

Open at Jiemian Ask Claude ChatGPT

China merges hospital digital ratings into a new evaluation regime

# Jiemian

China is replacing several hospital IT assessments with one digital-hospital framework and remote review, reopening a market squeezed by payment reform and anti-corruption controls.

The new framework combines electronic medical records, smart services, interoperability and smart management assessments into one evaluation standard.
Hospitals rated at level five or above will reportedly move from on-site inspection to remote system demonstrations, making audit trails and repeatable evidence more important than staged installations.
The policy arrives as diagnosis-related-group payments and medical-sector anti-corruption efforts have tightened public-hospital cash flow, delaying nonessential IT spending.
For vendors, the opportunity is less a return to indiscriminate software procurement than a shift toward measurable data exchange, workflow integration and evidence-backed operational outcomes.
The reform may revive demand, but it also threatens suppliers that survived by helping hospitals prepare for separate compliance checklists rather than improving the underlying systems.

Open at Jiemian Ask Claude ChatGPT

Chinese space tug passes engine and systems tests before 2027 debut

# IT Home

Chinese launcher company CAS Space says its 20kN upper stage completed a 566-second integrated firing and is ready for a first flight in the first quarter of 2027.

The upper stage is designed to move payloads between orbits after the launch vehicle has finished its ascent, a capability useful for multi-satellite deployment and higher orbits.
The engine completed multiple starts and variable-condition tests, with total accumulated firing time reported at 5,211 seconds and a measured specific impulse of 315 seconds.
A four-tank parallel-feed design kept propellant imbalance below 2.10%, helping preserve center of mass and attitude-control performance during flight.
The system is intended to be modular across launch vehicles and to support more than 20 restarts, which would make it useful for multi-orbit missions rather than a single fixed profile.
The company says system-level verification is complete, but a first flight remains the meaningful test of packaging, guidance and restart reliability in space.
The report is based on company and state-media disclosures, so independent flight data is not yet available.

Open at IT Home Ask Claude ChatGPT

Alibaba open-sources MyContext for enterprise agent memory

# AI QbitAI

Alibaba's Qwen Work team has open-sourced MyContext, a layer that turns chats, documents, meetings, and business rules into permission-aware, traceable context for agents.

The system is designed to handle changing timelines, conflicting facts, and continuously updated enterprise records rather than stuffing a static document into a prompt.
Each conclusion reportedly retains an evidence trail back to source conversations or files, while access follows user and organizational permissions.
This attacks a real production bottleneck: agents can call tools yet still fail because they do not know the local decisions, people, constraints, and history surrounding a task.
The architecture resembles other enterprise semantic layers, but Alibaba's choice to release an implementation could help it recruit developers around the Qwen Work ecosystem.
The supplied report says the project passed 1,000 GitHub stars in its first week; that is an early attention signal, not proof of adoption or security maturity.
The key enterprise test is whether context extraction stays accurate and affordable as records change, rather than becoming another opaque memory store.

Open at QbitAI Ask Claude ChatGPT

Alibaba’s Qwen Work now connects to WeCom

# AI QbitAI

Alibaba’s Qwen Work agent can read and write WeCom documents and tables, schedule meetings and assign tasks, completing its coverage of China’s three major enterprise-messaging platforms.

Also: Leiphone

The connector joins DingTalk and Feishu integrations, letting users invoke workplace actions through natural-language requests rather than separate commands or tools.
The capability is strategically important because it places an agent above the collaboration layer: the user states an outcome while the model coordinates data, permissions and workflow actions.
The same pattern raises enterprise governance questions around delegated authority, audit logs and accidental changes across documents, calendars and messaging.
Qwen Work also supports Slack, suggesting Alibaba is testing whether a China-built agent can sit across domestic and international collaboration ecosystems.

Qwen Work expands across China’s enterprise platforms Leiphone

Open at QbitAI Ask Claude ChatGPT

Chinese team releases a cross-body world simulator for robots

# AI QbitAI

Current Robotics introduced CurrentWorld-0, a simulator that predicts action, multi-camera views and force-touch outcomes across different robot bodies instead of merely generating plausible video.

The system keeps each robot's own action subspace while modeling shared physical consequences, so pushing a cup should produce the same object behavior whether the actor is a fixed arm, mobile platform or humanoid.
It generates multiple viewpoints consistently and tracks object state through occlusion, addressing a common failure where one camera shows success while another reveals that the object never moved.
Force and touch prediction are included for tasks such as gripping chips or peeling a cucumber, where visual alignment alone cannot establish contact or load.
The design targets evaluation and training: a failed action must remain failed in simulation rather than being visually corrected into a successful trajectory.
This is a technically meaningful distinction from ordinary video models, but the report supplies demos rather than quantitative rollout error, contact accuracy or transfer results.
If validated, a controllable cross-body simulator could reduce the cost of resetting physical scenes and allow policies to be compared under repeatable intermediate states.

Open at QbitAI Ask Claude ChatGPT

Hong Kong's heat deaths expose a gap in climate adaptation

# Culture Initium

Hong Kong recorded a 36.9-degree day and a 30.2-degree night as workers and older residents faced suspected heat deaths, prompting questions about whether the city treats heat as a public-health emergency.

The supplied analysis says the city endured nine consecutive days above 33 degrees and that the 30.2-degree minimum was its hottest night in 143 years.
It reports several suspected heat-related deaths, including older hikers and an outdoor renovation worker, illustrating the danger of persistent nighttime heat rather than isolated daytime peaks.
Hong Kong's dense, humid urban form magnifies exposure: cooling access, outdoor labor, housing conditions and social isolation matter as much as the official temperature.
The city's disaster-management culture is historically stronger on typhoons and rainstorms, while heat produces less visible disruption and weaker political urgency.
The labor dimension is central because outdoor workers may have limited ability to stop work, find shade or absorb lost income during warnings.
Useful adaptation would require targeted outreach, workplace enforcement, cooling centers and neighborhood-level monitoring, not only weather alerts.
The story matters as a governance test for an aging, high-density city where climate risk is increasingly a chronic service and labor problem.

Open at Initium Ask Claude ChatGPT

DeepSeek raises cached-context prices as storage becomes the bottleneck

# Analysis Leiphone

DeepSeek reportedly raised its cached-token price as much as 11 times, exposing how long-context AI shifts cost from matrix computation toward KV-state storage, movement and scheduling.

The reported off-peak cache price rose from 0.025 yuan to 0.3 yuan per million tokens, roughly four cents, while peak pricing rose fivefold according to the article.
Cache hits reuse previously computed key-value state, avoiding repeated prefill work for codebases and long documents; the savings disappear when high concurrency constantly evicts and reloads that state.
The analysis describes compressed attention and hot-cold storage across GPU memory, CPU memory and NVMe, making data placement and movement part of inference architecture.
At extreme load, GPUs can spend more time moving cache blocks than doing useful inference, turning a compute cluster into an I/O system with poor effective throughput.
The pricing change is therefore a capacity-management signal, not simply a monetization decision: cheap caching encouraged customers to retain enormous contexts without paying for the storage burden.
For agent builders, context design, reuse patterns and cache locality may matter as much as model choice, especially when long-running coding agents generate repeated state.
The technical claims come from the supplied analysis and should be validated against DeepSeek’s published pricing and serving implementation.

Open at Leiphone Ask Claude ChatGPT

China’s state media urges standard Chinese terms for AI

# Policy SCMP

Chinese state media is urging people to replace English AI terms such as agent and LLM with standardized Chinese equivalents, linking vocabulary to national discourse power.

The argument is not merely linguistic: the commentary presents control over terminology as part of China’s effort to build an autonomous technology narrative.
Standardized domestic terms can make technical concepts easier to disseminate through education, policy and mass media, while also reducing dependence on the vocabulary of US-led platforms.
The move reflects a recurring state preference for translating imported technical categories into language that fits official ideological and administrative frameworks.
It may create friction with engineers who work across global documentation, open-source repositories and APIs, where English names remain the interoperability layer.
The policy does not show that Chinese researchers will stop using English internally; it shows that public-facing discourse is being treated as an arena of technological sovereignty.
For foreign companies operating in China, terminology choices can affect localization, policy reception and whether a product is framed as part of China’s own innovation system.
The important question is whether the campaign remains a media editorial or becomes a requirement in standards, procurement, education and product labeling.

Open at SCMP Ask Claude ChatGPT

A Chinese edge SoC pairs an NPU with a CUDA-compatible GPGPU

# AI Leiphone

Chengheng Micro's CH3715 combines a 48-TOPS NPU with a 1-TFLOPS FP32 GPGPU, CPU, DSP, FFT and image processing to replace multi-chip edge-vision systems.

The architectural choice reflects heterogeneous workloads: the NPU handles neural inference, the GPGPU supplies programmable high-precision parallel compute, DSP and FFT handle signal processing, and the CPU schedules the system.
The four-core GPGPU is reported as CUDA-compatible, which could reduce migration costs for customers with existing parallel algorithms while the NPU targets INT8 and FP16 inference.
The SoC also integrates dual ISPs, a rendering GPU, high-speed interfaces and real-time control cores, aiming to reduce board area, data movement and cross-chip latency in robots and industrial vision.
The company claims more than 50% higher system throughput, 30% lower power and 40% smaller equipment, but those figures are vendor claims and lack workload details.
The design lesson is that edge AI is not a pure TOPS contest: perception, geometry, signal processing and control often need different numerical precision and software models.
Chengheng says its next chip will remove less-used blocks, improve memory bandwidth and target roughly half the cost, suggesting that customer deployments will determine whether the broad first-generation integration was justified.

Open at Leiphone Ask Claude ChatGPT

Chinese naval software expert disappears from academy records

# Policy SCMP

Vice Admiral Zhao Xiaozhe, a senior Chinese naval software expert involved in combat-information systems, has disappeared from the Chinese Academy of Engineering's online membership list.

Zhao reportedly helped set research priorities for the armed forces and worked on systems that process battlefield information and support command decisions.
His removal from the academy site is not itself proof of dismissal, investigation or political trouble, but it adds to a pattern of prominent defense experts becoming less publicly visible.
The episode illustrates the tightening boundary around military technology personnel as software, data fusion and decision systems become central to combat capability.
For outside observers, public biographies and institutional websites are increasingly unreliable indicators of who currently holds influence inside China's defense research system.
The strategic signal is therefore about opacity as much as personnel: important military-technology networks may be harder to track precisely when their work is becoming more consequential.

Open at SCMP Ask Claude ChatGPT

DJI redesigns the sensor stack behind 8K 360-degree cameras

# Analysis GeekPark

DJI’s Osmo 360 II reflects a longer Chinese imaging strategy built around custom square sensors, drone-derived stitching and high-speed stabilization rather than incremental camera packaging.

The company developed a 1/1.1-inch square CMOS so the circular fisheye image uses more of the sensor area, enabling native 8K output from two 4K views without upscaling.
DJI says the custom sensor’s effective imaging area was 64 percent larger than the previous leading 1/1.3-inch design and that it achieved 8K at 50 frames per second.
The technical foundation came from drone mapping, multi-camera stitching, spatial reconstruction and gimbal stabilization, where motion and parallax expose errors more severely than ordinary handheld shooting.
A 360-degree camera changes the capture workflow: users record everything first and choose the view later, which suits motorcycles, sports and hands-free work where framing cannot be monitored.
The product history shows DJI using adjacent businesses as an engineering laboratory, then rebuilding commodity components around the actual geometry of a new form factor.
The strategic question is whether this approach creates a durable imaging platform or only a premium niche whose advantages competitors can copy.

Open at GeekPark Ask Claude ChatGPT

RL infrastructure turns model training into a live production line

# AI Leiphone

A Chinese AI infrastructure provider describes reinforcement learning as a coupled generator, environment and trainer system where throughput and model freshness must be balanced continuously.

In the proposed loop, the generator produces rollouts, an environment supplies rewards and a trainer updates weights before the process repeats.
The architectural bottleneck is matching trainer throughput with effective generator throughput: a slow generator starves training, while a fast one creates stale rollouts.
The system also shares weights, rollout data, KV cache, environment state and task queues, making scheduling and state management as important as raw accelerator count.
Reported experiments show continued RL can improve smaller reasoning models, but the commercial claims come from the provider and its research partners.
The broader signal is that post-training infrastructure is becoming a product category: model labs need a persistent data-generation and evaluation loop, not just a one-time pretraining cluster.

Open at Leiphone Ask Claude ChatGPT

PLA drill uses drones to deliver cash and gold in a Taiwan scenario

# Policy SCMP

A Chinese state-media report described a People's Liberation Army exercise in which a drone delivered cash and gold, apparently simulating wartime financial support for a Taiwan operation; the report was later deleted after online debate.

The exercise reportedly used a multirotor drone to drop camouflage cases at designated locations under the Eastern Theater Command.
The unusual payload suggests planners were testing how to move financial resources into contested or disrupted areas, not merely practicing logistics for ammunition or supplies.
Chinese online discussion became part of the event: the report was deleted after heated reaction, showing the sensitivity of publicizing operational details that can look theatrical or politically provocative.
Deletion does not prove the drill was fabricated, but it does limit what can be established from the supplied account and may indicate concern about the message it sent.
The broader military signal is that Taiwan planning is being imagined as a whole-of-society problem involving money, local administration, and sustainment as well as combat.
Its importance would rise if similar exercises include communications, banking continuity, civilian logistics, or occupation-governance components.

Open at SCMP Ask Claude ChatGPT

Six AI agents generate, playtest and repair video games

# AI QbitAI

A Chinese AI startup uses specialized agents in a generate-run-check-repair loop, testing whether a game is actually playable instead of merely compiling.

The system converts a natural-language idea into rules, character behavior, level goals, win conditions and numerical parameters before producing code and assets.
Its key evaluation loop runs the game, observes failures such as impossible jumps or broken hit detection, then repairs the affected components rather than declaring success when code compiles.
This treats playability as a systems property emerging from interactions among physics, maps, feedback and difficulty, which ordinary unit tests cannot fully capture.
The company plans to move beyond code and engine rendering toward a world model that predicts the next game state directly from player actions and history.
The report offers a useful product boundary: AI can already accelerate prototypes, but dependable games still require runtime evaluation and human judgment about whether an unexpected behavior is a bug or a viable mechanic.

Open at QbitAI Ask Claude ChatGPT

OPPO says agent capability now outruns industry coordination

# Policy Leiphone

OPPO argues that models are already strong enough for useful consumer agents, but fragmented services, permissions and accountability prevent them from completing real tasks reliably.

The company and Alipay are building a trusted agent-to-agent protocol that combines device-side computation, contextual authorization and payment-grade command execution.
The proposed architecture separates a system agent that understands the user from specialist agents that perform domain tasks, rather than expecting one model to own every service.
OPPO says the partnership already connects nearly 200 functions and more than 18 public-service scenarios, including payments, benefits, permits and utilities.
The deeper bottleneck is institutional: services must expose stable interfaces, identify who authorized an action and assign responsibility when an agent makes a mistake.
This is a distinctly Chinese deployment strategy because device makers, payment platforms and public-service workflows are being connected through a small number of dominant ecosystems.
The reported increase of several million daily active users for Alipay is company-provided and does not establish that agent actions are profitable or safe at scale.

Open at Leiphone Ask Claude ChatGPT

Honor uses a robot phone to search for a new identity

# AI Leiphone

Honor’s $1,400 Robot Phone compresses a four-degree-of-freedom robotic gimbal and agent software into a consumer device, using embodied hardware to move beyond smartphones.

Also: Jiemian, 36Kr, IT Home

The phone can track users, follow music and reorient its camera, while Honor claims more than 400,000 reservations; both figures come from the company’s launch messaging.
The hardware reportedly inherits miniature motors, reducers and motion-control algorithms from an internal robotics program that shifted away from humanoids toward nearer-term products.
Honor’s broader architecture divides intelligence among personal devices, cloud models and physical terminals, with an agent operating system coordinating actions.
The product may be more strategic demonstration than mass-market category: internal sales expectations reportedly range from 20,000 to 200,000 units, tiny beside mainstream smartphone volumes.
The move shows how Chinese consumer-electronics firms are using embodied AI as a new identity and investment narrative while they search for growth beyond phones.
Its commercial test is whether the moving camera solves a recurring user problem strongly enough to justify price, thickness, reliability and battery costs.

Open at Leiphone Ask Claude ChatGPT

Xi tells Ecuador’s president Latin America should choose its partners

# Policy SCMP

Xi used an Ecuador meeting to defend Latin American autonomy and promote Chinese cooperation in trade, infrastructure, energy and digital industries.

The message combines a diplomatic principle against outside interference with a practical offer spanning mining, finance, green development, new energy and AI.
The framing is consistent with Beijing’s effort to present China as a development partner rather than a regional hegemon or ideological patron.
The report gives no Ecuadorian commitment beyond the meeting and no evidence that the rhetoric changes existing US or Chinese influence on the ground.
The event is part of an active China strategy in Latin America, where commercial access and political autonomy are deliberately linked.

Open at SCMP Ask Claude ChatGPT

China says county consumption is accelerating with urbanization

# Policy Jiemian

China’s Commerce Ministry attributes faster county-level consumption to migrants settling in county seats and rising rural incomes, while targeting goods and services for smaller cities.

Rural retail sales rose 2.4 percent year on year in the first seven months, 1.3 percentage points faster than urban sales, according to the ministry briefing.
County and township sales represented 39.1 percent of total retail sales, up 0.3 percentage points from a year earlier.
The ministry links demand to county urbanization: county-seat and county-level-city populations rose by more than 30 percent in the latest census, while more than 15 million returning entrepreneurs came back from cities.
Rural disposable income rose 42.8 percent cumulatively during the current five-year planning period, and rural households had a higher marginal propensity to consume than urban households.
The policy response is aimed at a structural bottleneck rather than only a stimulus burst: 49.4 percent of surveyed rural residents cited limited product variety as their main offline-shopping problem.
The push toward electric vehicles, smart green goods and charging infrastructure shows Beijing trying to turn rural consumption into an outlet for industrial capacity while integrating smaller markets into national supply chains.

Rural retail sales outpace urban growth in China Jiemian

Open at Jiemian Ask Claude ChatGPT

Fake mobile agents leave Xi'an residents with surprise loans

# Culture Jiemian

Chinese media describe alleged fraudsters posing as mobile-company agents who use victims' unlocked phones and face scans to open installment loans for supposedly free handsets.

One 71-year-old Xi'an resident said a caller offered a free phone, then used his unlocked handset and photograph to arrange a 2,155-yuan loan, roughly $300, with monthly payments and interest.
The case illustrates how device subsidies, carrier-style sales scripts, identity verification and consumer credit can combine into a high-friction fraud channel.
The warning is operationally important for companies handling identity workflows: a trusted-brand pretext can turn routine phone access into authorization for financial contracts.
The article presents the residents' accounts and a warning, not a final regulatory finding or adjudicated fraud judgment.
The broader signal is that face recognition and handset control remain socially legible as proof of consent even when older users do not understand the transaction they are authorizing.

Open at Jiemian Ask Claude ChatGPT

Inner Mongolia builds a state-backed token economy

# Policy Jiemian

Inner Mongolia's government has issued a broad plan for token supply, trading, export and safety governance, tying AI development to green computing, data assets and local industrial policy.

Six regional agencies propose token factories, public computing access, a standardized trading platform and cross-border service experiments through the free-trade zone.
The plan links token production to domestically sourced computing, renewable electricity, model deployment, data-set construction and industry-specific applications such as energy, agriculture and manufacturing.
It also supports small one-person companies, specialized suppliers, talent programs and financing paths that could turn data sets and token services into recognized economic assets.
Security provisions include network-protection requirements, registration of generative-AI services and cross-border risk controls, showing that commercialization is being built inside a compliance perimeter.
The initiative is notable less for immediate demand than for how a provincial government is trying to make an AI supply chain legible to planners, lenders and local firms.
Whether the program creates durable businesses depends on actual buyers and usable data, not the number of supported platforms or demonstration projects.

Open at Jiemian Ask Claude ChatGPT

NetEase builds an AI layer for a young social community

# AI QbitAI

NetEase Media introduced Bee AI as a shared foundation for search, personal agents and interactive content inside a youth-oriented community, aiming to turn AI users into creators.

Also: Leiphone

The product strategy is not to launch another general model but to embed information retrieval, task assistance, personalization and generative interaction into an existing social product.
Its personal agent can answer questions, assist group chats and organize life information; the company says interactions rose nearly 90% over three months and AI-to-user interaction contributed 40% of activity, reaching 70% at the high end.
The interactive-content system lets users generate small games and branching experiences whose state changes in response to clicks, turning consumption into participation and remixing.
NetEase emphasizes user authorization and the ability to edit or delete personal information, acknowledging that continuity and personalization create governance obligations.
The report is company-led and offers no retention baseline, revenue, safety audit or independent measurement of the claimed engagement gains.

NetEase's Bee AI turns a social app into an agent platform Leiphone

Open at QbitAI Ask Claude ChatGPT

Alibaba releases HappyShrimp for end-to-end AI music creation

# AI QbitAI

Alibaba's HappyShrimp turns a natural-language mood, memory or story into lyrics, composition, arrangement and vocals through one end-to-end music-generation model.

The model reportedly plans the song as a single long-form structure instead of separately generating lyrics, melody, arrangement and voice and stitching them together.
Its product pitch centers on interpreting everyday cultural and emotional language, not only tags such as genre, tempo or instrumentation.
The launch includes a partnership with Taihe Music Group, making rights, artist participation and commercial distribution as important as model quality.
The report provides no audio benchmark, copyright framework, artist compensation model or evidence that long-term musical structure is consistently preserved.
The model is available in China and overseas, so the launch is also a test of whether a Chinese creative model can build a global consumer workflow rather than remain a domestic novelty.

Open at QbitAI Ask Claude ChatGPT

Taiwan uses FPV drones against armored vehicles in a live-fire drill

# Policy SCMP

Chinese coverage highlights Taiwan's first live-fire FPV-drone exercise against armored vehicles, framing it as an asymmetric response to the People's Liberation Army while Taiwan focuses on battlefield adaptation.

The exercise applies drone tactics developed in the Ukraine war and expands Taiwan's strike options beyond rockets and heavy artillery, according to the report.
The technical shift is toward inexpensive, rapidly deployable first-person-view systems that can provide precision attack capacity without relying only on large platforms.
For Beijing, the exercise is a visible reminder that a Taiwan contingency would involve distributed commercial-style systems and improvised tactics, not just conventional force balances.
The report does not establish how many drones, what payloads, or what level of operational readiness Taiwan has achieved.
The signal becomes more consequential if the training is linked to domestic production, resilient communications, and mass mobilization rather than remaining a limited demonstration.

Open at SCMP Ask Claude ChatGPT

Chinese surgeons report rare double-thigh replantation success

# Signal SCMP

Doctors in eastern China reportedly helped a man walk again after both legs were severed at the thighs, a rare reconstructive-surgery result after nearly 20 months of recovery.

The patient is reportedly walking indoors with a frame, handling most daily tasks independently and regaining sensation down to the ankles.
The case is technically significant because replanting both upper legs requires reconnecting bone, vessels, nerves and muscle while managing prolonged rehabilitation.
The article calls it a global first, but the supplied account does not identify an independent medical publication or explain long-term functional outcomes.

Open at SCMP Ask Claude ChatGPT

Wang Yi heads to Seoul amid doubts about US security commitments

# Policy SCMP

China’s foreign minister is preparing to visit Seoul after Trump called for fewer US exercises with South Korea, giving Beijing an opening to court an uneasy regional partner.

The timing lets Beijing present itself as a predictable diplomatic actor while US alliance commitments appear less certain to South Korea and Japan.
The visit is not evidence of a strategic realignment: Seoul remains embedded in the US security system and the supplied report describes opportunity, not an agreement.
For China, the immediate objective is influence around the Korean Peninsula and a chance to exploit allied anxiety without escalating militarily.

Open at SCMP Ask Claude ChatGPT

China report says embodied-AI training sites remain a costly bet

# AI Jiemian

A China Academy of Information and Communications Technology report says embodied-AI training facilities have formed an early industry chain but still struggle to cover heavy fixed costs.

Training sites combine scene construction, data capture, robot hardware, compute, networking, storage and management systems, making them capital-intensive infrastructure rather than ordinary data centers.
The report describes an inverted market structure: upstream technology supply is active while downstream applications remain narrow.
Selling data is the fastest route to revenue, but the report says data sales alone are unlikely to repay the facilities’ fixed investment.
The proposed answer is a platform model combining data products, model and robot validation, joint research and training as a service.
This is a useful Chinese industry signal because it exposes the economic bottleneck behind embodied-AI demos: repeated, standardized physical data collection is expensive before a scalable customer base exists.

Open at Jiemian Ask Claude ChatGPT

Baidu's AI revenue stays above half of the business

# Signal Leiphone

Baidu reported 31.3 billion yuan (roughly $4.4 billion) in quarterly revenue, with AI accounting for half of general-business revenue for a second straight quarter as two US funds increased holdings.

The company reported 25.2 billion yuan (roughly $3.5 billion) in general-business revenue, with AI contributing 50% according to the source.
The result gives investors a clearer measure of AI's weight inside Baidu, but it does not separate the economics of model services, cloud, advertising or autonomous driving.
The fund purchases are market sentiment, not proof that Baidu's search-to-AI transition is succeeding; they sit alongside the unresolved capital-allocation question raised in the separate analysis.

Open at Leiphone Ask Claude ChatGPT

Inner Mongolia is turning inference into an export commodity, and that makes the AI bubble a policy choice

Six departments make tokens an export good

Half-mast for Zhu, loyalty essays from the generals

Trump trims the drills, Wang Yi books Seoul

DeepSeek is now billing for memory, not maths

Honor bet 9,999 yuan on a camera that moves

Threads we are pulling