China Desk — 27 Aug 2026

China-Nepal flood disaster kills at least 180 and leaves hundreds missing

# SCMP

The Chinese and international accounts differ in timing and counts, but together describe a deadly Himalayan flood after glacial collapse, with Nepal suffering the largest reported loss and China facing border disruption.

The Chinese-side figures in the international report were three dead and 265 missing, while Nepal reported 177 dead and more than 644 missing at that point.
Nepal’s government said 484 tourists were missing, including foreign nationals from at least 26 countries, making the disaster an international consular and rescue problem.
The event hit the Gyirong Port area, a strategically important border crossing whose roads, power, communications, and tourism flows can be disrupted by mountain hazards.
Scientists cited in related Chinese reporting describe glacial collapse and meltwater lubrication as possible mechanisms, while the precise trigger remains uncertain.
Changing counts across reports are expected during an active rescue operation; the important signal is the scale and cross-border nature of the disaster rather than one early tally.

Open at SCMP Ask Claude ChatGPT

Himalayan flood devastates China-Nepal trade route as toll rises

# SCMP

A flash flood and mudslide destroyed infrastructure at the Gyirong crossing, leaving hundreds missing and exposing the environmental risk behind China-Nepal trade plans.

Also: Jiemian, Leiphone, Solidot, Initium, IT Home

The crossing is strategically important to Nepal’s effort to reduce reliance on Indian transit routes, so the disaster affects both humanitarian operations and regional commerce.
Chinese authorities confirmed major casualties and missing people, while preliminary Nepalese figures reported a much higher toll on the other side of the border.
The event followed an ice-rock avalanche and debris flow rather than a conventional earthquake, complicating early-warning and infrastructure design.
Future China-Nepal roads, ports and other projects will need to account for unstable high-altitude terrain, glacial melt and compound flood pathways.
The disaster may temporarily close a trade route, but its longer-term significance is whether both governments change construction standards and hazard planning.
Casualty figures remain preliminary and vary across reports, so the human toll should not be treated as final.

Open at SCMP Ask Claude ChatGPT

MiniMax shifts from consumer apps to enterprise AI

# InfoQ China

MiniMax reports $120M of first-half revenue, while enterprise and developer services reached 63% of revenue and more than 80% of its reported $800M ARR.

The strategic change is a revenue-mix pivot: the open platform and enterprise services business grew 703% year on year, far faster than consumer products.
MiniMax says customers and developers exceeded two million, while more than 60% of first-half revenue came from overseas markets; that gives the company a broader demand base than a China-only SaaS play.
The reported ARR is an annualized run rate, not recognized revenue, so the $800M figure should not be compared directly with the $120M half-year income statement.
Token use rose twentyfold from January to July, illustrating how agents can expand inference demand faster than user counts, but profitability still depends on pricing, utilization and serving costs.
Gross profit was about $21M against roughly $300M of research spending and a $360M IFRS loss, leaving commercialization materially short of funding frontier-model development.

Open at InfoQ China Ask Claude ChatGPT

China’s internet giants make a $169B annual AI infrastructure bet

# TMTPost

Chinese internet companies are committing hundreds of billions of yuan to AI infrastructure, accepting near-term cash-flow pain in exchange for control of models, compute, and future applications.

Alibaba and Tencent spent more than 120 billion yuan (about $16.9B) in second-quarter capital expenditure, while ByteDance’s 2026 plan reached 200 billion yuan (about $28.2B).
Tencent’s 13.8 billion yuan (about $1.9B) negative free cash flow was its first reported negative result and was largely tied to 51.4 billion yuan (about $7.2B) in computing prepayments.
The companies represent four business models: Alibaba sells cloud capacity, Tencent internalizes compute into its ecosystem, Baidu sells model and token services, and ByteDance combines internal use with external infrastructure sales.
Alibaba’s AI cloud and computing-services revenue reportedly grew 45% to 48.4 billion yuan (about $6.8B), while adjusted profit fell 38%, offering the clearest early monetization but also showing investment drag.
Domestic chip substitution is being accelerated by export restrictions, with industry research cited as expecting Tencent’s domestic-chip procurement share to exceed 65%.
The key risk is utilization: if consumer AI does not produce a mass-market breakout, internalized compute at Tencent and strategic spending at ByteDance could face a long payback period.
The comparison with US cloud providers is about timing rather than simple technological superiority; China is building at high intensity while US infrastructure businesses are further along the harvest curve.

China’s internet giants spend over $16.9B on AI infrastructure in Q2 TMTPost

Open at TMTPost Ask Claude ChatGPT

Chinese executives may join Xi’s US trip as trade truce extension nears

# SCMP

Chinese coverage presents a possible CEO delegation as part of extending the US-China trade truce, ahead of Xi Jinping’s expected September 24 meeting with Donald Trump.

Officials are discussing whether Chinese business leaders will join Xi’s US delegation, according to people familiar with the talks cited by the source.
The proposed extension concerns a one-year trade agreement reached in Busan last October, suggesting both sides are prioritizing managed continuity over a new comprehensive deal.
The CEO question is strategically important because Trump’s previous Beijing visit featured a large group of American executives, creating pressure for reciprocal business engagement.
Chinese participation would give companies a direct channel into negotiations that otherwise remain dominated by state officials and tariff mechanics.
The source does not identify which Chinese executives might attend, so the political signal is clearer than the commercial substance at this stage.

Open at SCMP Ask Claude ChatGPT

China’s AI buildout is becoming a power, land and finance story

# TMTPost

Chinese coverage argues that AI is now reallocating electricity, land and capital: data-service power use rose 43.3% in the first seven months as infrastructure funding expands beyond technology companies.

China’s internet data-service electricity use reached 59.3 billion kilowatt-hours in the first seven months, though that category includes cloud computing and cannot be assigned wholly to AI.
The article cites a government projection that computing-center electricity could reach roughly 6% of national consumption by 2030, up from about 1.6% in 2025.
That scale changes site selection: eastern users and companies need inference close to demand, while western regions offer cheaper land and power, creating a continuing east-west systems problem.
The supply challenge is not only GPUs. Power construction takes years while data-center delivery is being compressed toward roughly 100 days, creating a mismatch between compute deployment and grid readiness.
Chinese coverage also highlights underused capacity, with some estimates putting overall utilization below 70% and new centers below 30%, making scheduling and regional marketplaces economically important.
The financing model is broadening from venture capital and corporate cash to equity markets and institutional capital, meaning AI infrastructure risk is increasingly shared by the wider economy.

Open at TMTPost Ask Claude ChatGPT

China wants specialist doctors tied to family-doctor contracts

# TMTPost

A new health notice requires more hospital specialists to join family-doctor services by 2027, shifting the program from signing people up to making primary care usable.

The policy requires primary-care internal medicine, pediatric and traditional-medicine doctors to participate by the end of 2026, with other specialists and hospital rotations following by 2027.
It also moves from team contracts toward named individual doctors as the first responsible provider, making accountability more legible to patients and supervisors.
The underlying diagnosis is that China’s family-doctor program achieved coverage without enough usable supply, leaving residents with contracts that often produced little practical access.
Specialists are intended to strengthen triage and referral, but their limited time in community facilities may cap the policy’s effect.
The unresolved implementation variables are incentives, staffing, local capacity and whether specialist participation becomes real care rather than another administrative target.

Open at TMTPost Ask Claude ChatGPT

China’s trade-in subsidies shift from old appliances to smart products

# TMTPost

China has allocated 187.5 billion yuan (about $26.4B) for consumer trade-ins, and the 2026 program adds smart products such as glasses while narrowing support for older categories.

The Ministry of Finance says the funds have driven about 1.32 trillion yuan (roughly $186B) in sales and reached 178 million person-uses, implying more than seven yuan of sales per yuan of central funding.
Smart glasses receive a national 15% subsidy capped at 500 yuan (about $70) per item, while smart-home products are left more heavily to local experimentation.
The change from digital products to digital and smart products signals that Beijing is willing to subsidize market education for emerging devices, not only replacement demand.
The first month of smart-glasses subsidies produced only 7.2 million yuan (about $1.0M) in sales, but cited retail data show much faster growth over longer periods.
The program also narrows appliance coverage and standardizes national rules, indicating a preference for concentrated, measurable stimulus over broad support.
For manufacturers, the policy combines a demand boost with a test of whether AI-enabled devices can move beyond technology enthusiasts into ordinary households.
The important uncertainty is whether subsidy-driven trial converts into repeat usage after the discount disappears.

Open at TMTPost Ask Claude ChatGPT

China’s internet giants end the AI office-product race

# TMTPost

Tencent, Alibaba and ByteDance are consolidating AI office teams into fewer centers, revealing three different bets on the future of enterprise agents.

Tencent folded QClaw into WorkBuddy, treating a validated enterprise product as the main line and a lighter consumer assistant as a supporting capability.
Alibaba shut down three parallel product lines and created a Qwen office division, choosing organizational reset over selecting one existing winner.
ByteDance split Feishu between Doubao product leadership and Volcano Engine sales, then moved TRAE and Coze into the Doubao structure.
The common signal is that rising model, distribution and customer-acquisition costs are ending the era of unconstrained internal horse-racing.
The strategic risk differs by company: Alibaba faces migration and boundary problems with DingTalk, while ByteDance must reconcile enterprise customization with a reusable model core.
These are product-organization changes, not proof that each company has unified its underlying models or compute infrastructure.
The next six months should be judged by talent retention, product velocity, customer overlap and whether sales demands contaminate the common agent platform.

Open at TMTPost Ask Claude ChatGPT

China’s foreign minister meets the US ambassador in Beijing

# Jiemian

Chinese coverage stresses continuity and crisis management in Wang Yi’s meeting with US Ambassador David Perdue, while the exchange also points to preparations for senior-level contact.

Wang framed the relationship through the official formula of constructive strategic stability, emphasizing implementation of leader-level understandings and control of disputes.
The US ambassador echoed the language of respect, reciprocity and continued communication, and said Washington wanted to prepare the next round of high-level interaction.
The meeting is routine diplomacy, but its function is practical: keep channels open and remove obstacles before more consequential leader engagement.
Beijing’s public framing prioritizes stability and cooperation without detailing the disagreements, a familiar pattern in Chinese official reporting when bilateral tensions remain active.

Open at Jiemian Ask Claude ChatGPT

PDD’s 40% R&D jump signals a move deeper into supply chains

# InfoQ China

PDD’s second-quarter revenue grew 8% while advertising rose only 3%; R&D jumped 40% to 4.6 billion yuan, roughly $648 million, for supply-chain systems and governance.

The figures show a platform whose familiar advertising engine is slowing even as transaction-service revenue grows faster.
The reported R&D budget is spread across traceability, agricultural standards, factory digitization, rural fulfillment, compliance and overseas logistics rather than a single AI program.
A lychee traceability system illustrates the shift: platform data links farms, batches, inspections, packing, shipping and after-sales responsibility.
The strategic move is from matching demand to helping shape production and fulfillment, which can improve consistency but requires operational coordination outside the platform’s traditional boundary.
PDD reported that one merchant’s lychee repeat purchases rose from about 15% to 30%; that is a local example, not proof of a platform-wide effect.
The story matters because Chinese e-commerce growth increasingly depends on controlling physical bottlenecks, not just buying traffic.

Open at InfoQ China Ask Claude ChatGPT

Xpeng upgrades its driving model from 3D space to 4D time

# New Wisdom

Xpeng says its second-generation VLA now uses 30-second temporal context, six-second prediction and streaming inference to connect perception, prediction and action.

The model adds an Infini-VLA long-sequence architecture so driving decisions can use recent motion history rather than isolated frames.
Streaming autoregressive inference is claimed to improve end-to-end response speed by 300%, allowing perception and action to proceed continuously rather than in discrete batches.
An X-Foresight world model predicts possible behavior up to six seconds ahead, while a mixture-of-experts design separates driving tasks to reduce interference.
The vehicle-level Master Agent coordinates driving, cabin, chassis and body-control agents, turning spoken intent into multi-system actions such as finding a parking place.
Xpeng says the edge model has 3.5 times more parameters than its predecessor and is being reused across production cars, robotaxis and robots.
These are company claims; the meaningful validation will be safety performance, intervention rates and generalization across markets rather than parameter count.

Open at New Wisdom Ask Claude ChatGPT

iFlytek’s results show AI moving from investment to monetization

# 36Kr

iFlytek reported 11.623 billion yuan, roughly $1.64 billion, in first-half revenue and a second-quarter cash-flow turnaround as platform and enterprise AI gained share.

Revenue rose 6.5%, while net loss narrowed to 204 million yuan, roughly $29 million, against R&D spending of 3.007 billion yuan, roughly $424 million.
Open-platform revenue grew 36% to 3.705 billion yuan, roughly $522 million, becoming the company’s largest business line with a 20.2% gross margin.
API and model-as-a-service revenue rose about 70%, supporting a transition from project delivery toward recurring platform and token usage.
The company says it has more than 11.5 million AI developers, including more than 3.2 million large-model developers, giving it a distribution base for agents and model services.
Education, healthcare and automotive are the important verticals, while hardware supply constraints hurt its AI learning-device business.
The improvement is not yet a clean profitability story: large R&D and sales spending remain necessary, and second-quarter cash flow turning positive is only an early quality signal.

Open at 36Kr Ask Claude ChatGPT

DeepSeek reportedly seeks $7.4B at a $74B valuation

# Techmeme

US coverage presents DeepSeek’s reported financing as a business-model signal, while Chinese coverage stresses rapid revenue growth, high API margins, and the cost of building domestic AI infrastructure.

The Wall Street Journal report surfaced by Techmeme says DeepSeek may raise $7.4B at a $74B valuation; these figures remain source-attributed investor reporting.
Chinese reporting says January-to-July revenue reached 475 million yuan (about $67M), roughly ten times 2025 full-year revenue, while net loss was 715 million yuan (about $101M).
The same Chinese reports say overall gross margin was 44.6% and API gross margin 82.9%, but the company has not publicly confirmed the figures.
Infrastructure spending reportedly reached 11 billion yuan (about $1.5B), up from 1.2 billion yuan (about $169M) in 2025, showing how low-cost model pricing can coexist with heavy capital needs.
The contrast matters because Western coverage asks whether DeepSeek can support an IPO, while Chinese coverage frames efficiency and domestic model adoption as the strategic asset.
DeepSeek’s recent API price increases suggest the company is testing whether high usage can become sustainable cash flow without abandoning its cost advantage.
The financing story matters more if margins survive scale-up; it matters less if growth depends mainly on subsidized compute or unusually favorable access to infrastructure.

Open at Techmeme Ask Claude ChatGPT

Hands-on tests find GLM-5.3-Flash cheap enough for serious workflows

# AI Zhidx

A Chinese hands-on review says GLM-5.3-Flash can edit video, build websites and prototype games while cutting model prices by up to 97.5%.

Also: Leiphone

The model accepts text, images, video and files and was tested on visual question answering, document recovery, web replication, video editing and code generation.
The reviewer used it to extract a highlight reel from a 48-minute talk, preserving clips, subtitles and intermediate artifacts rather than returning only a summary.
In coding tests it inspected screenshots of a generated game and iterated on behavior, showing the value of multimodal feedback inside an agent loop.
The model is priced at 0.8 yuan per million input tokens, roughly $0.11, and 2.8 yuan per million output tokens, roughly $0.39, before temporary discounts.
The test is useful because it measures workflow completion rather than benchmark scores, but it remains a single publication’s hands-on evaluation.
Its combination of low price, multimodality and domestic-chip serving is the competitive threat to watch, not any one game recreation demo.

GLM-5.3-Flash claims frontier performance on domestic Chinese chips Leiphone

Open at Zhidx Ask Claude ChatGPT

MiniMax shifts from consumer apps toward enterprise model demand

# Signal QbitAI

MiniMax says first-half revenue reached $116.6 million, up 283%, while business revenue rose to 63.4% of the total and token use hit 20 times January’s level.

The company’s annual recurring revenue reportedly exceeded $800 million in August, up from $150 million in February, though the figure is management-disclosed.
Enterprise and developer services generated $73.9 million in the first half, up 703%, and supplied roughly three-quarters of incremental revenue.
Consumer products still grew, so the shift reflects faster business demand rather than a reported collapse in consumer usage.
The token surge suggests agents are increasing workload intensity: one automated task can generate far more inference than a human chat session.
MiniMax says infrastructure efficiency lifted gross margin to 17.9%, while R&D still reached $296.9 million, showing that commercialization has not removed the capital burden.
The strategic question is whether API and enterprise demand remains durable after novelty-driven adoption and open-model experimentation fade.

Open at QbitAI Ask Claude ChatGPT

ByteDance’s Doubao Work makes enterprise context the agent advantage

# AI Zhidx

A hands-on Chinese review finds Doubao Work can research, write documents, build presentations and act across Feishu, using organizational permissions and context as its main edge.

The reviewer says it produced a roughly 25,000-word research document with 47 traceable sources and a 14-slide presentation, though these are test results from one publication.
The agent can inspect Feishu chats, documents and tables, then turn project status into a stored weekly report and context-aware reminders.
Its skill system packages search, document access, table queries and messaging into reusable enterprise capabilities rather than one-off prompts.
The architecture inherits Feishu permissions and audit boundaries, addressing the central enterprise-agent problem of what the system may read and change.
The product’s advantage is not necessarily a smarter base model; it is privileged access to organizational memory and structured workflows.
The test does not establish reliability across arbitrary companies, and the review is closely tied to the product launch.

Open at Zhidx Ask Claude ChatGPT

Qwen Office says its new Flash model doubles speed and cuts tokens 75%

# AI QbitAI

Alibaba’s Qwen Office launched a model-tuned Qwen3.8-Flash mode that it says doubles task speed and cuts average token consumption by 75%.

Also: Leiphone

The model was tuned jointly for multi-step planning, tool selection and context compression rather than exposed as a generic base-model upgrade.
Alibaba says 95% of routine office tasks can use the lower-cost standard mode, reserving the advanced mode for harder work.
This is an application-level optimization claim: harness design, model routing and task-specific training matter as much as raw benchmark intelligence.
The source is authorized promotional material and supplies no independent latency test, token accounting or quality tradeoff.
If the gains generalize, office agents can become economically viable by reducing wasted reasoning and context rather than simply lowering model prices.

Qwen Office launches Qwen3.8-Flash as its new standard mode Leiphone

Open at QbitAI Ask Claude ChatGPT

Feihe puts 436 agents into dairy-factory operations

# AI Leiphone

Chinese dairy maker Feihe says it has built 436 agents across maintenance, process control and workforce management, using a common platform and factory data.

The company reports more than 240,000 active users over three months and nearly 23 billion token consumption across more than ten factories.
Its equipment-management agent cut average fault-recovery time from 104 hours to 47 and repeated-fault rates from 34% to 8%, according to Feihe.
A membrane-cleaning workflow reportedly saves more than 500,000 yuan, roughly $70,000, in annual purchasing and 144,000 tons of water.
A workforce agent reduced process nodes from 15 to five and reportedly saved more than 2 million yuan, roughly $282,000, in development and maintenance costs.
The operating model matters more than the agent count: frontline workers propose use cases while a central platform handles models, skills, permissions and governance.
These are company-reported results, but they show a credible path from agents as demonstrations to agents embedded in industrial control and administrative loops.

Open at Leiphone Ask Claude ChatGPT

Kunlun’s embodied world model ranks second in WorldArena

# AI Leiphone

Chinese robotics startup Kunlun says its GeWu world model won the image-quality category and ranked second overall in the WorldArena 2.0 evaluation.

The model scored 65.91 overall and 69.58 on image quality among 77 entries, according to the company’s account of the benchmark.
Its architecture separates intent, intervention and consequence into three parameter-independent transformer towers.
A one-way causal-attention pattern prevents the action tower from seeing future visual consequences while allowing the consequence tower to condition on intent and action.
The company reports that removing these pathways caused the model to imitate familiar video patterns instead of predicting physically consistent outcomes.
The result is technically interesting because it treats a world model as a simulator of action consequences rather than merely a video renderer.
The ranking is reported by the vendor and should be checked against the benchmark’s published protocol, task mix and reproducibility before being treated as a field-wide lead.

Open at Leiphone Ask Claude ChatGPT

Four ministries target vehicle reliability, hype, and autonomous-driving claims

# Policy Jiemian

A joint Chinese campaign will inspect vehicle quality and curb exaggerated marketing, while testing automated-driving systems, cloud security, AI safeguards, and accident logs.

Also: IT Home

The action runs nationally for one year and covers cars and motorcycles, with special attention to production consistency, reliability, durability, and validation of new technology.
Manufacturers must examine their own processes and suppliers, report defects, and initiate recalls when problems are found.
Regulators can publicly name companies and testing bodies, suspend product approvals or certification, block registrations, and impose fines for serious violations.
The policy explicitly treats marketing claims as part of product safety, an important distinction for driver-assistance systems whose capability boundaries are often communicated commercially.
The campaign also formalizes pressure on third-party testing organizations to resist non-rational competition and maintain objective, traceable evaluations.

Open at Jiemian Ask Claude ChatGPT

Wang Yi urges Washington to clear obstacles before Xi-Trump talks

# Policy SCMP

China’s foreign minister called for fewer obstacles to high-level exchanges and a positive agenda before an expected Xi-Trump meeting, signaling diplomacy aimed at stabilizing ties while managing disputes.

Wang made the appeal to US ambassador David Perdue in Beijing, according to China’s foreign ministry.
The official framing emphasizes implementing existing leader-level consensus, managing differences, and keeping relations stable, healthy, and sustainable.
The statement is preparatory diplomacy rather than evidence that contentious issues such as trade, technology, or security have been resolved.
Its significance lies in the Chinese preference for treating summit logistics and political atmosphere as part of the strategic negotiation itself.

Open at SCMP Ask Claude ChatGPT

China's tax crackdown and Unitree's stock slide top a policy roundup

# Analysis Sinocism

A China-focused newsletter connects tax enforcement, a surprise anti-corruption session, Evergrande's founder, and Unitree's stock decline to Beijing's effort to stabilize revenue and manage markets.

The roundup says the central government is using tax crackdowns to help cover local-government shortfalls, linking fiscal stress to more visible enforcement.
It also highlights a surprise plenary session of the Central Commission for Discipline Inspection and the life sentence of Evergrande founder Hui Ka-yan, reinforcing a climate of continued political and financial discipline.
The discussion of Unitree's sharp post-IPO decline asks what the Party wants from the stock market after encouraging strategic technology listings, rather than treating the price move as an isolated market event.
A cabbage scandal in Hebei is included as a reminder that food quality and consumer trust remain politically salient even amid high-tech policy.
The newsletter is commentary, not primary reporting, but its value is synthesis: fiscal extraction, anti-corruption, property cleanup, and technology-market management are being read as parts of one governance posture.

Open at Sinocism Ask Claude ChatGPT

Why Kimi and GLM converge on the same linear-attention setting

# AI Leiphone

Chinese researchers note that Kimi K3 and GLM-5.3-Flash both use a linear-attention forgetting bound of minus five, illustrating a narrow stability regime for long context.

Linear attention compresses history into a fixed-size state instead of retaining a full KV cache, reducing memory growth for very long sequences.
Its weakness is recurrent numerical instability: repeated state updates can underflow, explode or produce NaNs in half-precision training.
The reported minus-five gate bound preserves a small residual memory signal while avoiding the overload associated with retaining too much history.
Kimi appears to use the setting as part of a native hybrid architecture, while GLM combines related ideas with sparse attention and other components.
Similar public configuration values do not prove copying; independent teams can converge on the same narrow engineering optimum.
The broader signal is that Chinese model competition is moving below headline benchmarks into kernels, state representations and memory economics.

Open at Leiphone Ask Claude ChatGPT

Leju’s industrial VLA uses a middle-training layer for robots

# AI Leiphone

Chinese robotics maker Leju says its KUAVO VLA model used more than 600 hours of same-platform robot data and doubled efficiency in selected industrial tasks.

The model is positioned between a general VLA foundation model and task-specific post-training, storing common industrial capabilities before individual skills are learned.
Training on one robot configuration reduces variation in action space, mechanics and control behavior, making later adaptation less expensive on that platform.
The company reports a 48.27% task-success rate and 74.51% process score across 25 tasks, ahead of the comparison models listed in the article.
The design emphasizes process quality, smoothness and repeatability rather than only whether the final task succeeded.
This is a strong argument for a middle-training layer in embodied AI, but the reported benchmark is company-run and uses a limited task set.
The commercial payoff depends on whether skills transfer across factories and workflows without recreating the same data-collection burden.

Open at Leiphone Ask Claude ChatGPT

Chinese coverage frames Meta’s child-safety settlement as a product reset

# Policy GeekPark

Chinese analysis stresses Meta’s forced limits on recommendation, scrolling, and teen access, while US discussion frames the same settlement as Meta’s bid to lead child safety without admitting wrongdoing.

Also: Techmeme

The article says Meta may pay up to $18B over ten years, with most directed to states and youth-safety programs, though the final structure is complex.
Its most important point is that the settlement mandates product changes: time limits, night blocking, reduced notifications, private defaults, age detection, and a non-algorithmic feed.
Chinese coverage treats these constraints as an attack on the growth engine itself, while the US aggregator emphasizes Meta’s positioning and refusal to concede liability.
The proposed terms would matter beyond Meta if similar obligations reach TikTok and YouTube, because unilateral restrictions could simply shift teen usage to competitors.
The case is a precedent for regulating addictive design through enforceable product architecture rather than only fines or disclosure rules.

US coverage casts Meta’s settlement as a child-safety leadership play Techmeme

Open at GeekPark Ask Claude ChatGPT

China's 6G policy is converging with Qualcomm's AI-network vision

# Policy Zhidx

China says it will accelerate 6G research while Qualcomm presents networks that combine connectivity, sensing, and compute, making AI-native infrastructure a shared direction.

Qualcomm's vision treats 6G as more than a faster data pipe: networks should sense objects, classify motion, allocate resources, and decide where computation occurs.
The company demonstrated full-duplex sensing of vehicles and drones, large MIMO prototypes, and a terminal-to-edge-to-data-center compute continuum.
Chinese officials say the next five-year period will prioritize 6G core technology, trials, standards, and ecosystem development, with commercial deployment targeted around 2029 in the company's account.
The China dimension is substantive because Qualcomm expects Chinese companies to participate heavily in standards while China is also building domestic alternatives and defining its own industrial priorities.
The hard systems problem is coordination across radio, sensing, edge compute, devices, and AI workloads; the marketing vision is much easier than achieving useful latency, energy, and reliability.

Open at Zhidx Ask Claude ChatGPT

Extreme weather is unlikely to derail China’s nuclear expansion

# Policy SCMP

Chinese nuclear expansion is less exposed to river-water constraints than some European fleets, according to an industry specialist, though heat remains a design risk.

The analysis contrasts China’s continued build-out with European reactors forced to reduce output as rivers become warmer and shallower.
China’s newer nuclear sites and broader cooling choices may reduce exposure to the specific river conditions affecting France and Romania.
That does not remove climate risk: heat waves, drought, storms and water competition still affect plant siting, cooling design and grid operations.
The strategic implication is that climate adaptation may reinforce China’s nuclear push by making firm low-carbon generation more valuable, rather than stopping it.
The source offers an industry view rather than a plant-by-plant safety assessment, so it does not establish that China’s fleet is immune to extreme weather.

Open at SCMP Ask Claude ChatGPT

Siemens says industrial agents need decades of plant knowledge

# AI QbitAI

Siemens is pairing industrial software and AI agents with its Xcelerator ecosystem, including deployments in China, but the story is mainly a global product strategy.

Its Eigen Engineering Agent can read electrical-design files, identify conflicts and generate PLC labels and projects from hardware topology and natural-language descriptions.
The company reports two- to five-fold workflow speed improvements and deployments across more than 100 companies, including Chinese pilots.
Its Intelligence Center X is an orchestration layer connecting PLM, ERP, MES, CRM and operational technology rather than another foundation model.
The argument is that industrial agents need equipment protocols, safety constraints, permissions and traceability; a generic chatbot cannot safely bridge IT and OT.
The China dimension is substantive through local sales and deployments, but the source provides no China-specific adoption, regulatory or performance comparison.

Open at QbitAI Ask Claude ChatGPT

Chinese academic calls for maritime readiness in a future US rivalry

# Policy SCMP

A prominent Chinese scholar argues that China needs a national consensus on sea power and should channel overseas investment returns into defense as maritime competition with the US intensifies.

The argument starts from China’s dependence on sea-borne trade and the United States’ existing maritime advantage.
It represents strategic commentary, not a government policy announcement, but its language reveals how some Chinese academics frame commercial shipping and naval power as one problem.
The policy direction implied is a broader fusion of economic security, overseas investment, trade-route protection, and defense spending.

Open at SCMP Ask Claude ChatGPT

Univisual grows from phone camera modules into robot vision

# AI 36Kr

Chinese supplier Univisual reached about 2 billion yuan (roughly $282M) in annual revenue and now derives nearly a third from robot business, supplying vision systems to major robot makers.

The company moved from commodity phone modules into cleaning robots, then humanoids and AI hardware, using manufacturing and NPI capabilities as a bridge into higher-value systems.
Its customers include Ecovacs, Unitree, and other robot and hardware companies; the article says a US electric-scooter order exceeded $20M, the company’s largest to date.
The technical shift is from standard modules to integrated optics, algorithms, SoC adaptation, and mass production for navigation, obstacle avoidance, and edge deployment.
Univisual’s strategy is to serve many startups and convert prototypes into repeatable product lines, using a broad customer funnel rather than depending on one anchor account.
The company recently completed a funding round in the hundreds of millions of yuan, but the source does not disclose valuation or investor amounts.

Open at 36Kr Ask Claude ChatGPT

OceanBase leads China’s distributed database market with 20.2%

# Signal Leiphone

A new market study puts China’s distributed transactional database market at about 5.95 billion yuan, roughly $838 million, with OceanBase leading at 20.2%.

The study defines the category by horizontal scalability, cross-node transaction consistency, strong consistency, high availability and zero recovery-point loss.
Alibaba Cloud, Tencent Cloud and Huawei Cloud followed OceanBase with 17.9%, 16.5% and 14.3% shares respectively.
The market grew from about 1.69 billion yuan in 2021 to 5.52 billion yuan in 2025, a reported 34.5% compound annual growth rate.
Adoption is moving from internet and payment back ends into financial cores, communications and other infrastructure sectors where migration risk matters more than benchmark speed.
OceanBase is also extending toward HTAP, multimodal data and agent-facing services, pushing the category toward an AI data-platform role.
The figures come from a consulting report and should be read as a market-definition claim, not an audited industry census.

Open at Leiphone Ask Claude ChatGPT

Tsinghua team outlines an end-to-end path to cheaper model inference

# AI InfoQ China

A Tsinghua team describes an inference engine that combines operator tuning, KV-cache management, quantization, heterogeneous scheduling and adaptive parallelism for Chinese chips.

The presentation separates prefill, which is compute-heavy, from decode, which is dominated by memory access and makes different demands on hardware.
Its optimization stack spans kernel-level operators, heterogeneous KV-cache handling, mixed-precision compilation, CPU-GPU scheduling and workload-aware parallelism.
The open-source Chitu engine is presented as a way to carry those optimizations across domestic accelerator ecosystems rather than assuming a single NVIDIA stack.
The strategic point is that inference economics depend on the whole software-hardware path; faster individual chips do not solve memory movement, irregular model shapes or changing request loads.
The article is a technical presentation rather than an independent benchmark report, so the claimed gains should be treated as research and engineering claims pending comparable testing.

Open at InfoQ China Ask Claude ChatGPT

DeepSeek’s price hike is also a report on China’s AI power shortage

# Analysis TMTPost

DeepSeek raised peak API prices and added cheaper off-peak windows, prompting Chinese analysts to read the move as both demand rationing and a test of sustainable model economics.

The pricing design turns time into an allocation mechanism: latency-sensitive requests pay more while batch jobs are pushed toward nights and weekends.
Chinese market coverage interpreted the first increase as evidence of scarce inference capacity, then treated the weekend discount as a scheduling tool rather than a simple reversal.
The article links the pressure to uneven geography, with inference demand concentrated in the east while much training capacity and power sit farther west.
The deeper business signal is a possible end to China’s prolonged API price war, as model providers try to fund training and infrastructure instead of buying share with ever-cheaper tokens.

DeepSeek’s higher prices give rival Chinese models room to grow TMTPost

Open at TMTPost Ask Claude ChatGPT

The flood that erased Gyirong has a sequel, and Beijing is trying to compute when it arrives

Gyirong is gone and a lake is filling upstream

Xi lands in Washington on 24 September

Four ministries come for "pseudo-innovation"

GLM-5.3-Flash served 62 trillion tokens on Chinese chips

Shawshank sells out a thousand-seat hall

Threads we are pulling