China Desk — 08 Oct 2026, 15:31 UTC

Manus reopens a Beijing team after raising more than $500 million

# QbitAI 6h

Manus’s parent raised over $500 million, roughly 3.55 billion yuan, and reopened 17 Beijing roles for a China-focused agent product after previously shifting its center abroad.

Also: Zhidx

The hiring spans agent product, harness engineering, evaluation, model research, security, virtualization, DevOps, growth, and support, suggesting a full local product organization rather than a small satellite office.
Manus previously cut much of its Chinese team after moving headquarters to Singapore, so the return raises questions about employee trust and whether China is now a durable market or a tactical hiring pool.
The company says it is building domestic products and working with Chinese model vendors and ecosystem partners, while continuing its global business.
The reported financing was led by Boyu and IDG, with Tencent, Sequoia China, and ZhenFund continuing to invest; the deal ties capital structure to a renewed domestic strategy.
The important signal is organizational: an agent company that once optimized for overseas expansion now sees enough Chinese market, talent, and ecosystem value to rebuild locally.

Manus raises more than $500 million after returning to independent operation Zhidx

Open at QbitAI Ask Claude ChatGPT

Geely claims 2.25 MW charging with AI thermal control

# QbitAI 6h

Geely launched a charging system claiming 2.25 megawatts peak power, sub-65-degree Celsius battery temperature, and a 20 percent cycle-life improvement.

The system uses a PowerMind energy model, five-part liquid cooling, and a fully immersed DC bus converter to coordinate the vehicle, charger, battery, storage, and grid.
Geely says its AI temperature controller predicts battery temperature 30 seconds ahead and dynamically changes power to keep average temperature below 55 degrees Celsius.
The company reports 10-to-97 percent charging in 8 minutes 40 seconds under a test configuration, with the peak measured at 1.093 megawatts rather than the system’s 2.25-megawatt maximum.
The claimed 20 percent cycle-life gain depends on a battery-health algorithm and pulse-repair technique; independent long-duration testing is not supplied.
The deployment challenge is infrastructure: Geely says its network has more than 2,500 stations and 12,000 guns across 232 Chinese cities, but megawatt charging requires substantial grid and thermal capacity.

Open at QbitAI Ask Claude ChatGPT

Huawei says China may share US AI safety concerns later

# Pekingnology 9h

Huawei chairman Eric Xu says China must accelerate AI now but may encounter the same frontier risks as US labs later, while predicting hardware shortages could last to 2029.

Xu’s argument links safety perception to capability: Chinese firms may not yet feel risks that US frontier labs face because they have less compute, but he expects the concerns to converge as China advances.
That is a revealing Chinese industry position—risk management is endorsed, but slowing development before reaching the frontier is rejected as strategically premature.
Huawei expects global shortages in chips, memory, optical networking, and other AI infrastructure to persist for years, with China potentially taking longer to catch up.
The technical discussion reportedly covers Huawei’s interconnect and computing architecture, showing that supply constraints are being treated as a systems problem rather than only a chip-fabrication problem.
The remarks align corporate strategy with national competition: accelerate domestic capability, manage risks as they become concrete, and build an ecosystem less dependent on Nvidia.
Because the account is a media interview, it should not be read as a formal Chinese government position.

Open at Pekingnology Ask Claude ChatGPT

Alibaba plans much larger Qwen models and AI-assisted chip design

# Leiphone 8h

Alibaba says future Qwen models could reach 5 trillion to 10 trillion parameters while using sparse architectures, and reports that Qwen autonomously reduced a chip design’s area by 42 percent.

Alibaba’s roadmap moves from a reported 2.4-trillion-parameter Qwen 3.8 toward models several times larger, while seeking efficiency through sparse attention, linear attention, and external memory.
The company says Qwen 3.8 Flash activates only 6 billion parameters per inference, cuts training cost to one-ninth, reduces input pricing to one-quarter, and improves long-context prefill throughput 8.6 times.
In a reported NoC experiment, Qwen 3.8 Max used EDA tools for more than 60 hours and reduced physical area 42 percent and estimated power 59.5 percent without round-by-round human intervention.
The strategic shift is from AI as an application to AI as an engineering participant in model training, chip design, and inference-stack adaptation.
These are company-reported results; the missing details are design baseline, constraints, verification coverage, and whether the process generalizes beyond a carefully selected module.

Open at Leiphone Ask Claude ChatGPT

Beijing and Brussels head into trade talks with tensions unresolved

# Sinocism 19h

Chinese policy coverage expects Europe’s trade commissioner to press Beijing on the trade surplus while Beijing resists export curbs, framing talks as a test of whether both sides can avoid escalation.

The European side is reportedly considering limits on Chinese plug-in hybrid exports, while Chinese officials reject a quota-style solution and propose monitoring trade flows instead.
The Chinese framing treats consultation as a way to delay or redirect punitive measures rather than concede structural constraints on exports.
The dispute reflects a broader pattern: Beijing is willing to discuss verification and sector-specific mechanisms but resists arrangements that look like externally imposed limits on industrial capacity.
The talks matter for technology executives because automotive exports, batteries, chemicals, and other industrial sectors are increasingly tied to market access and retaliatory risk.
The supplied newsletter is analysis, not an official negotiating record, so the reported positions should be treated as informed interpretation until confirmed by both sides.

Open at Sinocism Ask Claude ChatGPT

Huawei is designing storage specifically for inference KV cache

# Leiphone 11h

Huawei and Nvidia are introducing dedicated context-memory products as KV cache becomes large enough to manage separately, but SSD requirements remain unsettled.

Externalizing KV cache trades storage cost for recomputation: raising cache hit rate from 90 to 95 percent halves the portion that must be recomputed, but the last few percentage points can require disproportionate capacity.
The replacement is not one terabyte of DRAM for one terabyte of SSD; the report describes practical ratios of one to five or one to ten and cost gaps of several dozen times.
Workload lifetime determines the design: minute-scale cache churn demands high write endurance, while week-scale retention shifts the problem toward capacity and retrieval performance.
The industry is discussing enterprise SSDs at roughly three to nine drive writes per day, but the target may move as workloads stabilize.
Huawei’s OceanStor M900 and Nvidia’s CMX indicate that cache tiering is moving from an internal server optimization toward a shared storage product category.
A common specification may still take six to twelve months to emerge, followed by roughly another year of product development and validation.

Open at Leiphone Ask Claude ChatGPT

China says a foreign warship lost power while shadowing its destroyer

# SCMP 2h

Chinese coverage presents a rare South China Sea encounter in which a foreign warship allegedly lost propulsion and radar while closely following a Chinese Type 055 destroyer.

China’s state broadcaster says the foreign vessel conducted aggressive close shadowing during routine training by the destroyer Zunyi; the account does not identify the ship or provide independent confirmation.
The incident resembles earlier reports involving US Navy destroyers, allowing Chinese media to reverse the usual narrative of Chinese harassment and portray US forces as vulnerable and provocative.
The propulsion, power, and radar claims matter operationally because they frame close encounters as both escalation risks and demonstrations of relative systems resilience.
The absence of the other navy’s account is material: this is a Chinese state-media narrative, not a settled incident record.
If corroborated, the episode would add pressure to already dangerous rules-of-behavior questions in the South China Sea; if not, its main importance is information signaling.

Open at SCMP Ask Claude ChatGPT

Brake-pedal failure sends Chinese luxury carmaker JAC into a trading halt

# TMTPost 51m

A third-party test showed three Huawei-JAC V800 vehicles breaking at the same brake-pedal bracket; JAC shares hit the limit down as a safety dispute became a trust crisis.

Also: Jiemian, IT Home

The vehicles failed during repeated 100-to-zero kilometer-per-hour braking tests, and the report says the cars were produced across August and September rather than coming from one obvious unit.
JAC and Huawei’s premium joint venture is carrying the automaker’s high-end transformation story, so a basic mechanical failure can reprice expectations far beyond the cost of one component.
The company says the test was nonstandard and that no customer failures have been reported, but it still plans a design improvement and a free upgrade option for delivered vehicles.
The market reaction was amplified by a 40-plus-percent run-up in JAC shares over the prior six trading sessions, making the roughly 6.2 billion yuan, about $870 million, one-day market-value loss partly a reversal of speculative expectations.
The dispute illustrates a recurring Chinese consumer and market dynamic: domestic premium brands are judged not only against formal standards but against highly visible online tests that can become a proxy for engineering credibility.
The key unresolved question is whether the repeated fracture reflects a narrow extreme-load weakness or a broader supplier, validation, or design-control problem.

Open at TMTPost Ask Claude ChatGPT

Western open models are chasing China’s production advantage

# InfoQ China 4h

Reflection and Mistral launched open-weight models against Chinese references, while Chinese models accounted for about 41 percent of Hugging Face downloads and 56 percent of one gateway’s production tokens.

Reflection’s Beam is a 501-billion-parameter mixture-of-experts model with 23 billion active parameters, trained on 23.8 trillion tokens and more than 1 million GPU-hours of reinforcement learning.
The reported benchmarks put Beam near the previous Chinese frontier but behind newer GLM, Kimi, and DeepSeek systems, so its strongest claim is inference efficiency rather than absolute capability.
The production signal matters more than the leaderboard: the report says open-weight models rose from under 10 percent to 56 percent of Vercel AI Gateway tokens between December and August.
Chinese models are therefore functioning as the reference architecture for global open-model builders, not merely as low-cost alternatives for Chinese users.
Reflection’s heavy reliance on GB300 compute also complicates the DeepSeek comparison: it is an open-weight challenger built with very expensive frontier infrastructure.

Open at InfoQ China Ask Claude ChatGPT

Chinese underwater-robotics startup raises five rounds in nine months

# 36Kr 12h

Weilan Deep Sea completed five financings in nine months to train underwater embodied models and automate offshore inspection, where human support can exceed $100,000 per day.

The company’s platform combines acoustic, optical, and magnetic sensing with a modular robot and an underwater energy-and-communications base station, addressing the information and power constraints of deep water.
It reports sub-one-meter positioning error, more than 95 percent positioning accuracy, 10-kilogram-plus arm payload, 48-hour endurance, and 12-knot maximum speed; these are company claims without independent field validation.
The commercial model starts with inspection services for existing subsea infrastructure, then moves toward standardized services and eventually robot sales, prioritizing customer outcomes over model demonstrations.
The company says a fourth-generation robot has already performed fully autonomous inspection of subsea energy facilities, but the scale, intervention rate, and failure handling are not disclosed.
Underwater robotics is a strong early embodied-AI domain because machines can replace not just a worker but saturation-diving vessels, support crews, and hazardous exposure.
The proposed path from human-in-the-loop autonomy to full autonomy mirrors staged industrial automation, with deployment data and corrective feedback driving later generalization.

Open at 36Kr Ask Claude ChatGPT

Chinese lung-cancer drug misses its global Phase III endpoint

# TMTPost 2h

A global Phase III trial of furmonertinib missed its independently reviewed progression-free-survival endpoint, wiping more than 10 billion yuan, roughly $1.4 billion, from Chinese drugmaker Ailex’s market value.

The trial enrolled 398 patients and tested two furmonertinib doses against platinum chemotherapy for first-line EGFR exon 20 insertion lung cancer.
The 240-milligram arm reached 11 months median progression-free survival versus 9.5 months for chemotherapy, but the independently reviewed hazard ratio was 0.75 with a 0.0654 p-value and a confidence interval crossing 1.
Investigator-assessed results looked better, illustrating why regulator-facing trials prioritize prespecified independent review over a favorable local analysis.
The unusually strong control arm compressed the apparent treatment advantage, but that explanation does not change the failure to meet the primary endpoint needed for a straightforward overseas filing.
The result weakens the Chinese biotech sector’s licensing-out narrative: headline milestone values and global rights can be worth far less when a pivotal trial fails.
Ailex proposed a 100 million to 200 million yuan, roughly $14 million to $28 million, buyback, but the market treated the problem as a reset of future expectations rather than a liquidity crisis.

Open at TMTPost Ask Claude ChatGPT

Porsche cuts 9,000 jobs as Chinese demand reshapes its strategy

# Analysis QbitAI 8h

Porsche plans to cut up to 30 percent of jobs and extend combustion models after China deliveries fell 56 percent from their 2021 peak, exposing the limits of its electric pivot.

Porsche’s operating margin fell from 18 percent in 2023 to 1.1 percent in 2025, while China deliveries dropped from 95,671 vehicles in 2021 to 41,938 in 2025.
The company is responding with fewer models, lower overhead, delayed electric projects, longer combustion-car lifecycles, and a roughly 20 percent increase in average prices for top-end vehicles.
The China problem is not merely volume: domestic premium EV brands compete on software, assisted driving, cabin technology, and power while traditional brands still rely heavily on badge and driving dynamics.
Porsche plans to remove up to 40 percent of management roles and invest 2.1 billion euros, roughly $2.3 billion, in German production and R&D through 2035.
The episode is a strategic warning for global manufacturers: a Chinese market downturn can force a global product reset when local competitors move faster than a multinational’s powertrain cycle.

Open at QbitAI Ask Claude ChatGPT

Chinese startup unveils robots for round-the-clock convenience stores

# AI QbitAI 2h

Zhengxing Innovation combines a humanoid robot, a narrow aisle manipulator, and a management platform for zero-retrofit retail, targeting commercial service in 2027.

The architecture separates embodiment from orchestration: H1 handles human-facing work, C1 handles narrow-aisle transport and restocking, and M1 assigns tasks, monitors state, and supports human takeover.
The company claims its lightweight action model reaches 98.6 percent average success on LIBERO while running 24 times faster than mainstream alternatives, but the benchmark is not equivalent to open-store reliability.
C1 uses multi-sensor perception, depth cameras, and a narrow omnidirectional base to work around existing shelves; the report claims obstacle braking within 10 centimeters.
A reinforcement-learning post-training system uses task data and human corrections, with a reported 95 percent success rate for a specified restocking task.
The strategic bet is that retail front-of-house environments are a harder and more commercially meaningful test than structured warehouses, but the promised 99 percent task autonomy remains a vendor claim.
The company plans both hardware sales and robotics-as-a-service, making utilization, maintenance, and exception handling more important than demo performance.

Open at QbitAI Ask Claude ChatGPT

Shanghai HiSilicon argues that every IoT device should become lightly intelligent

# AI Zhidx 4h

Shanghai HiSilicon is pushing a low-power edge-AI strategy built around chips, cloud coordination, and a developer ecosystem, arguing that deployment—not raw compute—is the bottleneck.

The company’s concept of light intelligence combines connectivity, sensing, interaction, and modest local inference so ordinary appliances can react without round trips to the cloud.
The article correctly identifies embedded constraints that generic coding agents cannot solve alone: board support packages, pinouts, memory, thermal limits, power budgets, and hardware-in-the-loop debugging.
HiSpark is positioned as a full-stack bridge from silicon and models to SDKs, tools, demos, and deployment support, reflecting a shift from chip vendor to platform vendor.
The strategic target is a fragmented Chinese IoT market dominated by smaller firms that cannot independently optimize models, runtimes, and hardware.
The pitch remains ecosystem marketing; the report does not provide chip specifications, developer counts, or independent deployment results.

Open at Zhidx Ask Claude ChatGPT

Former US officials warn that transactional Taiwan policy raises risk

# Policy SCMP 31m

Chinese coverage stresses Washington’s dependence on Beijing and weakened military readiness, framing Taiwan policy as a strategic-risk problem rather than a bilateral dispute.

Former US officials told a Global Taiwan Institute event that Washington is tying Taiwan policy more closely to its relationship with Beijing, according to the report.
The Chinese framing foregrounds US readiness and policy inconsistency, while the available Western-facing account emphasizes warnings from former American officials rather than Beijing’s preferred interpretation.
The report also links the concern to the war in Iran, arguing that diversion of US military resources could affect contingency planning in the Taiwan Strait.
For Chinese policymakers, this kind of argument supports a narrative that US commitments in Asia are conditional and that regional actors should reassess the reliability of American support.
The account is analysis rather than a government decision, so its value lies in mapping a Chinese media frame, not proving an imminent change in military posture.

Open at SCMP Ask Claude ChatGPT

Expanded Japan-US exercises complicate PLA planning around Taiwan

# Policy SCMP 3h

Chinese coverage highlights a larger Keen Sword exercise involving missiles, air defenses, communications, and about 45,000 Japanese personnel near Taiwan.

The exercise is scheduled for October 19 to 29 and includes about 45,000 Japanese personnel, 6,100 vehicles, 40 ships, and 260 aircraft, up from the 2024 edition.
The relevant change is not merely troop count: rehearsing movement of anti-ship missiles, air defenses, and communications networks onto islands tests the logistics and command relationships that a Taiwan contingency would stress.
For the PLA, dispersed allied infrastructure creates more targets and more uncertainty about how quickly Japan and the United States could support one another.
The exercise also signals that Japan’s role is moving beyond rear-area support toward active regional denial and communications integration.
These are planning implications, not proof that the exercise rehearses a Taiwan war directly; the political effect may be as important as the military one.

Open at SCMP Ask Claude ChatGPT

IROS papers show robotics returning to geometry and control

# AI Leiphone 12h

A review of 1,933 IROS papers finds learning, perception, planning, control, and manipulation still dominate robotics, with large models being embedded rather than replacing classical engineering.

The review counts about 809 papers on robot learning or embodied AI, 564 on navigation and planning, 556 on perception, 546 on control, 520 on manipulation, and 213 on humanoid or legged systems.
The categories overlap, but their distribution shows that robotics remains a systems discipline rather than collapsing into a single foundation-model problem.
Examples include 3D representations inside VLA models, classical task-and-motion planning beside language models, and small fast policies distilled from slower foundation models for real-time control.
A Chinese team appears in the highlighted GeoVLA work, which adds depth and explicit 3D geometry to action generation rather than simply scaling a two-dimensional vision-language model.
The industrial implication is clear: model vendors that ignore calibration, dynamics, latency, and contact will struggle to cross from demonstrations into factories.

Open at Leiphone Ask Claude ChatGPT

Naive AI raises $400 million in seven months at a $1.42 billion valuation

# AI Leiphone 9h

A secretive Tsinghua-linked AI startup reportedly completed three financings totaling $400 million, roughly 2.84 billion yuan, in seven months without publicly explaining its product.

The company is led by Tsinghua professor Dai Jifeng and reportedly draws on researchers from Microsoft Research Asia and SenseTime, including expertise in computer vision and AI infrastructure.
The unusual combination of a billion-dollar-plus valuation, no visible team photo, and almost no product disclosure points to a capital market willing to price elite technical lineage before commercial proof.
The source describes a network built around deformable convolution, video recognition, and large-scale training systems, but does not establish what Naive AI is currently building.
The financing claims are attributed to the outlet’s reporting and are not accompanied by term sheets, named investors, or audited company disclosures.

Open at Leiphone Ask Claude ChatGPT

Trump reportedly urged Xi to thaw relations with Japan

# Policy SCMP 6h

Chinese coverage says Donald Trump pressed Xi Jinping to ease tensions with Japan, but Beijing remained unmoved over Tokyo’s Taiwan comments.

The report says China’s restrictions followed Japanese leader Sanae Takaichi’s statement that Japanese forces might intervene in a Taiwan conflict.
Beijing’s refusal, as presented here, reflects a preference to preserve coercive leverage over Tokyo rather than let Washington mediate the dispute.
The story shows how Taiwan language now links US-China summit diplomacy to China-Japan economic and political relations.
Because the account relies on people familiar with the talks, it should be treated as reported diplomatic context rather than an official Chinese decision.

Open at SCMP Ask Claude ChatGPT

Chinese video model Vidu Q4 offers cheap 4K generation

# AI QbitAI 2h

Vidu Q4’s preview release claims 4K output, up to 15 reference images and three audio references, with test videos costing under $10 in total.

The report positions Q4 around expressive acting, camera continuity, and lighting interactions rather than only resolution or clip length.
Its quoted entry price is 0.09 yuan per second, roughly 1 cent, for 720p; higher resolutions and modes cost more, so the headline economics are workload-dependent.
The model’s ability to preserve identity across scenes and use reference audio points toward creator workflows where consistency matters more than one-off visual novelty.
The article is heavily promotional and relies on provider demonstrations, leaving independent quality, safety, and commercial-rights testing unresolved.

Open at QbitAI Ask Claude ChatGPT

Chinese robotics team runs CNC tending with 0.5 millimeter clearance

# AI Leiphone 8h

Simplexity Robotics says one robot can tend two CNC machines with only 600 real trajectories, using multimodal action generation, memory, and candidate-action scoring.

The task combines reflective metal, weak visual texture, hidden contact interfaces, and a 0.5-millimeter workpiece-to-chuck gap, making it a stronger factory test than a clean manipulation demo.
The architecture separates action proposal, long-horizon memory, and action selection: a multimodal action model generates fragments, DRAM preserves history, and a frozen policy is scored by a separate evaluator.
The team reports 15 successes for its fused visual method versus eight for a baseline in nominal trials, but the sample is small and not a production yield metric.
The system uses head and wrist cameras, state conditioning, and future video, 3D, and action supervision during training while predicting only actions at inference.
The strategic lesson is that industrial embodied AI still needs explicit geometry, temporal memory, and evaluation layers rather than relying on a single large end-to-end model.

Open at Leiphone Ask Claude ChatGPT

China is pressed to do more than issue statements on Hormuz

# Policy SCMP 3h

Chinese commentary argues that Beijing’s meetings with Washington have produced language but no solution for the Strait of Hormuz crisis, increasing pressure on China to act.

The piece frames China as a power with diplomatic leverage but limited visible operational commitment, asking whether it will move beyond calls for negotiations.
It says recent Xi-Trump meetings produced no breakthrough on Iran, while China backed a return to an interim US-Iran agreement.
The framing is important because it reveals a Chinese-facing debate about the gap between China’s growing global role and its reluctance to assume direct crisis-management costs.
For US technical and policy readers, the question is less whether China can dictate an outcome than whether energy-security pressure will force a more active Chinese posture in the Gulf.
The article is an argument, not evidence of a new Chinese policy decision; its value is as a window into expectations placed on Beijing.

Open at SCMP Ask Claude ChatGPT

Russian security aide visits China amid trilateral-summit speculation

# Policy SCMP 5h

Nikolai Patrushev’s China visit is being watched for signs of a possible Xi-Putin-Trump meeting at the next APEC gathering, though no summit is confirmed.

The visit puts high-level China-Russia coordination in the foreground as Beijing navigates a possible three-way meeting involving Washington.
Patrushev’s portfolio includes maritime affairs, making the talks relevant to security and regional logistics as well as diplomacy.
The summit speculation is the analytical hook; the supplied report contains no evidence that Beijing has agreed to the format or agenda.

Open at SCMP Ask Claude ChatGPT

AI-linked attacks hit seven South Korean financial firms

# Analysis Solidot 7h

A Chinese report says hackers used the open-source ARTEX tool against seven South Korean financial firms and may have exposed their identity through Claude Code, though attribution remains uncertain.

Also: Techmeme

CrowdStrike reportedly linked the attacks to a possible actor in Guangdong, but the source explicitly says the person’s identity has not been confirmed.
The alleged workflow combines an agentic penetration-testing console with an AI coding assistant, showing how public tools can compress reconnaissance, exploit development, and operational documentation.
The attacker allegedly embedded personal initials, a messaging handle, education, and location in a security résumé generated with Claude Code, creating an unusual but weak attribution clue.
This item clusters with the US CrowdStrike summary: Chinese coverage emphasizes the possible Guangdong location, while the US account emphasizes campaign infrastructure and exfiltrated data.
For defenders, the practical lesson is to treat AI-generated code, offensive open-source tooling, and personal information leaks as one combined threat surface.

Chinese and US coverage differ on the AI-assisted attacks on Korean banks Techmeme

Open at Solidot Ask Claude ChatGPT

Tsinghua researchers launch a robotics startup built around action-native learning

# AI Leiphone 8h

BASAL Intelligence, founded by a Tsinghua doctoral graduate, is betting on in-context embodied models that learn new tasks on site instead of retraining for every environment.

The company rejects a standard pretrained-model-plus-task-finetuning pipeline, arguing that it scales poorly across long-tail physical environments and robot bodies.
Its action-native model treats motion, feedback, failure correction, and demonstrations as the primary context, rather than using language as the main representation of physical work.
The team cites X-VLA and the ODEWorld latent world model, with under 1,000 hours of pretraining data and an IROS challenge win as early research evidence.
The commercial promise is important: if a robot can adapt in minutes from a few demonstrations, deployment cost could grow sublinearly as customers add sites and tasks.
The unresolved engineering question is whether fast in-context adaptation can remain safe and reliable across objects, contact dynamics, and embodiments outside the team’s demonstrations.

Open at Leiphone Ask Claude ChatGPT

China nominates Song Li to lead the World Health Organization

# Policy SCMP 3h

China nominated health official Song Li for WHO director general, presenting the bid as a chance for mainland China to lead the agency as US funding retreats.

If elected, Song would be the first mainland Chinese director general of the WHO; China previously backed Margaret Chan, who was born in Hong Kong.
The nomination arrives during a funding shake-up after the United States moved away from the organization, making the leadership contest partly a question of institutional financing and geopolitical influence.
A Chinese leader could bring Beijing greater agenda-setting power in global health, but election support and the candidate’s policy platform are not supplied.
The story is strategically relevant because international organizations are one of the arenas where China can convert state capacity and diplomatic networks into rule-setting influence.
The nomination is not an appointment, so its immediate effect is signaling rather than control of the WHO.

Open at SCMP Ask Claude ChatGPT

US judge keeps alleged Beijing agent in custody

# Policy SCMP 9h

A US judge ordered Wanying Zhang held in a case alleging surveillance of a Taiwan leader’s son, making the China-US counterintelligence dimension more important than the hearing itself.

Prosecutors argued Zhang posed a flight risk because close family members live in China; her lawyers said the case rests on a single day of alleged surveillance in Seattle.
The case illustrates how family ties and perceived links to Beijing can become central to pretrial detention arguments in US national-security prosecutions.
The supplied report does not establish that Zhang worked for Beijing or describe the evidence beyond the surveillance allegation, so attribution remains a charge rather than a finding.
Its wider significance is institutional: counterintelligence cases involving Taiwan and China are increasingly handled through ordinary detention decisions that carry substantial geopolitical overtones.

Open at SCMP Ask Claude ChatGPT

Anthropic’s infrastructure gamble keeps the US-China AI race running

# Analysis InfoQ China 1h

Chinese coverage pairs Anthropic’s huge compute commitments with its warning that frontier AI may be dangerous, framing the US-China race as both economic bet and safety trap.

The report says Anthropic’s 2025 revenue approached $4.6 billion while operating losses exceeded $8 billion, alongside at least $518 billion in long-term cloud and infrastructure commitments.
Anthropic’s argument is that frontier intelligence may become more valuable as older capabilities commoditize, which rationalizes buying compute years ahead of demand but exposes the company to extraordinary execution risk.
The company is described as simultaneously warning investors about catastrophic autonomy risks and insisting that development cannot pause, a contradiction that is central to its political and commercial posture.
Chinese coverage treats the US-China competition as structurally difficult to stop, emphasizing that labor disruption, cyber abuse, biological risk, and open-model diffusion will continue even as firms call for safeguards.
The China angle is comparative rather than event-driven: the piece uses Anthropic’s own strategy to explain why Chinese developers may see acceleration, infrastructure, and safety as inseparable.

Open at InfoQ China Ask Claude ChatGPT

A Chinese AI history traces the talent pipeline from Microsoft Research Asia

# Analysis Leiphone 9h

A long Chinese profile of Zhang Lei uses a 2001 internal face-detection contest to explain how Microsoft Research Asia alumni helped seed China’s AI industry.

The story presents MSRA as a talent network where researchers moved between robotics, speech, vision, academia, and startups, carrying engineering habits across institutions.
Its central lesson is methodological rather than biographical: deeply understanding a problem and building from hardware through algorithms shaped Zhang’s later choices.
The profile also shows how elite Chinese AI careers are narrated through mentorship, institutional lineage, and the migration of research talent into commercial companies.
The account is highly anecdotal and celebratory, so claims about who built half of China’s AI industry should be treated as profile framing, not a measured assessment.

Open at Leiphone Ask Claude ChatGPT

China Minmetals moves a heavily indebted property arm inside the group

# Policy 36Kr 6h

China Metallurgical Property will acquire Minmetals Real Estate’s 100 percent stake in an internal restructuring, absorbing a unit with liabilities above assets and a 125 percent debt ratio.

The transaction is an intra-group state-owned-asset adjustment designed to centralize property operations rather than an outside rescue or market acquisition.
The target had 19.87 billion yuan, roughly $2.8 billion, in assets and 24.86 billion yuan, roughly $3.5 billion, in liabilities at June 30, leaving net assets negative 4.99 billion yuan, roughly negative $700 million.
A related urban-renewal unit has positive net assets and 31 projects across more than 20 cities, which may be the strategic asset being consolidated around the distressed parent.
The projected combined company would still have a 74.94 percent debt ratio and a forecast net loss of 624 million yuan, roughly $88 million.
The move illustrates how central state-owned groups use internal ownership changes to ring-fence and manage property exposure without presenting the process as a conventional bankruptcy.

Open at 36Kr Ask Claude ChatGPT

Chinese game companies are becoming major AI investors

# Analysis TMTPost 3h

Chinese game publishers invested 8.72 billion yuan, roughly $1.23 billion, across 72 deals from 2025 through mid-2026, sending most capital into AI and robotics.

The report says 15 listed game companies participated in 72 financing events, while only seven deals worth about 200 million yuan, roughly $28 million, were directly tied to entertainment.
That allocation reflects a shift from game publishers using investment to widen game catalogs toward using capital to secure AI capabilities, compute, and strategic relationships.
Tencent, NetEase, Perfect World, and others are seeking exposure to models, robotics, and infrastructure as game acquisition costs rise and blockbuster creation becomes less predictable.
Equity stakes can buy preferential access to custom tools and joint development, but investment gains can also mask weakness in the core game business.
The strategy is especially Chinese because large platform firms use cross-sector capital allocation to shape supplier ecosystems rather than relying only on internal R&D.

Open at TMTPost Ask Claude ChatGPT

A part-time worker describes the pressure inside Sam’s Club China

# Culture TMTPost 3h

A Beijing-area worker’s account portrays Sam’s Club China as legally compliant on hours and pay but intensely metric-driven, localized, and demanding on the shop floor.

The account says part-time workers receive 27.7 yuan per hour, roughly $3.90, work limited shifts, and receive formal training, while managers impose constant checks and sales targets.
The worker describes pressure around membership sales, satisfaction scores, payment channels, and public performance rankings, showing how a foreign retailer can operate through highly localized management norms.
Front-line work is presented as tightly sequenced: staff monitor entrances, cleanliness, member access, safety, and relief coverage under heavy customer traffic.
The report also highlights outsourcing and education gaps, with English-heavy training and strict food-safety rules falling on a workforce described as largely middle-aged and without higher education.
The contrast between polished foreign-brand employment imagery and day-to-day pressure is the social story: imported corporate culture does not prevent local performance systems from dominating labor experience.
This is one worker’s account, so it should be read as qualitative evidence about workplace perception rather than a representative labor survey.

Open at TMTPost Ask Claude ChatGPT

Delta Force reaches 58 million daily users in China

# Culture 36Kr 4h

Tencent’s Delta Force says its China version now exceeds 58 million daily active users after two years, using multiple play styles and community participation to pursue longevity.

The game’s strategy is to avoid forcing a single definition of winning across extraction, competitive, social, collection, and casual play modes.
That design responds to a mature Chinese game market where older titles account for most mobile top-100 revenue and user growth has slowed sharply.
The developers describe maps as social sandboxes as well as combat spaces, incorporating fishing, music, photography, and player-created stories into a shooter ecosystem.
The 58 million figure is company-reported daily activity, not independently audited engagement or retention data.
The broader lesson is that large Chinese service games increasingly compete as community platforms, not just as collections of mechanics.

Open at 36Kr Ask Claude ChatGPT

Xiaohongshu turns a game community into a Shanghai festival island

# Culture 36Kr 8h

Xiaohongshu expanded its RED LAND festival to 160-plus game and anime properties, using lifestyle-themed installations to turn fandom into an offline community.

The event grew from 80,000 to 150,000 square meters and from 50 to more than 160 participating properties, while adding a game-and-anime parade.
Instead of simply recreating game worlds, booths translated them into food, travel, karaoke, and everyday social scenes, lowering the entry barrier for people who do not already play the games.
The platform is acting as curator and community infrastructure, not just an advertising channel; it helps publishers design interactions that generate user content and cross-fandom traffic.
The model reflects a Chinese consumer pattern in which fandom becomes a lifestyle practice, with cosplay, travel, collecting, and creator participation reinforcing game retention.
The event also exposes operational limits: remote venues, weather, crowd flow, and transport can undermine the experience even when community demand is strong.

Open at 36Kr Ask Claude ChatGPT

ChinAI	403	0/0		HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk	200	1/20
Pekingnology	200	1/20
Ginger River Review	200	0/20
Sinocism	200	1/20
QbitAI	200	10/10
36Kr	200	9/30
Solidot	200	8/19
Leiphone	200	20/20
GeekPark	200	1/30
IT Home	200	12/60	+40 over cap
SCMP	200	10/50	+10 over cap
Initium	200	4/15
Jiemian	200	8/30	+11 over cap
China Digital Times	200	0/15
The Wire China	200	0/50
Techmeme	200	1/15
Hacker News	200	1/20
V2EX	200	8/50	+39 over cap
Zhidx	200	5/20
InfoQ China	200	8/20	+4 over cap
New Wisdom	200	8/15
TMTPost	200	8/20	+12 over cap

China's toughest car safety regulator this week was a ByteDance app

Three brake pedals snapped on camera

Sefcovic lands in Beijing with nothing to sign

The West's open-weight comeback is scored in Chinese

Two service-labour stories landed on the same day

Everyone led with a different half of the Korean bank hack

Threads we are pulling