China Desk — 30 Aug 2026, 13:46 UTC

Shanghai puts finance veteran in line for its next mayor

# SCMP 2h

Shanghai named career finance official Zhu Zhongming party chief of the municipal government, positioning him to replace Mayor Gong Zheng and potentially become the city's next mayor.

Zhu was announced as secretary of the Shanghai government Party Leadership Group, a position that places him inside the municipal executive apparatus rather than at the top Party post.
The appointment is a personnel signal: Shanghai is elevating a finance-background official at a time when the city must manage growth, property stress and fiscal constraints.
He replaces Gong Zheng in that government-party-group role, while the city's senior-most political position remains the municipal Party secretary.
The article says the new post puts Zhu in line to become mayor, but the mayoral succession is not yet a formal appointment.
For companies operating in Shanghai, the relevant question is whether Zhu's background translates into a stronger emphasis on fiscal discipline, financial services and investment-led growth.

Open at SCMP Ask Claude ChatGPT

China says 261 foreign citizens are missing after Tibet floods

# SCMP 9h

Chinese authorities say 261 foreign nationals from 23 countries remain missing after floods near the Nepal border, while cyber police are targeting rumors around the disaster.

The missing include 104 Nepalese citizens, 49 Indians, 18 Americans and 15 Britons, according to authorities in Gyirong county.
The scale and international mix turn a regional disaster into a consular and information-management problem for China, Nepal and many third countries.
The report's pairing of casualty accounting with rumor control shows how Chinese authorities manage disasters through both administrative updates and online narrative policing.
The location near Tibet's border makes access, communication and verification especially sensitive, though the supplied report does not describe foreign rescue access or the specific rumors being targeted.
The figures are an official count of missing people, not a final death toll, and may change as search operations continue.
For companies and governments with personnel in the region, the event is also a reminder that infrastructure and crisis communications in western border areas are unusually difficult to verify.

Open at SCMP Ask Claude ChatGPT

China's autonomous-driving bill shifts some liability to automakers

# TMTPost 3h

A draft road-safety law would make manufacturers or importers responsible for violations while an automated-driving function is active, but most Chinese cars remain legally driver-supervised L2 systems.

The draft creates a national legal distinction between automated driving, which continuously performs the dynamic driving task within its design domain, and driver assistance, which merely helps a human.
When an automated-driving function is active, the manufacturer or importer would handle the traffic-violation responsibility; when it is inactive or the car is L2 assistance, the driver remains responsible.
The draft would also place the evidentiary burden on companies claiming that an incident was unrelated to the automated function, because manufacturers control the relevant logs and system data.
The unresolved boundary is the transition event: a system that hands control back shortly before a crash may create a fight over whether the failure occurred during automation or human takeover.
China's market is still overwhelmingly on the driver-responsibility side: the article cites 70.5% L2 penetration and 34.2% navigation-assistance penetration as of July 2026.
The commercial implication is that autonomy marketing, logging, handover design and liability insurance will become linked product decisions rather than separate compliance topics.

Open at TMTPost Ask Claude ChatGPT

UBTECH's factory robot business surges, but profits remain out of reach

# TMTPost 2h

UBTECH reported first-half revenue of 1.27 billion yuan (roughly $179M), up 104%, with full-size humanoid revenue up 1,445%, yet adjusted EBITDA remained negative.

The company sold 16,123 humanoid robots in the first half, while full-size embodied-robot revenue reached 590 million yuan (roughly $83M) and gross margin rose to 44.7%.
UBTECH says its stack spans servos, robot bodies, foundation models, world models, simulation and fleet intelligence, rather than simply combining third-party hardware with an external model.
Its 11.0 million real-machine data points, more than 80% from industrial conditions, are intended to create a deployment loop in which factory failures improve later control policies.
The reported customers include Airbus, BYD, Geely, Foxconn and Sany, giving the industrial story more weight than a demonstration-only humanoid launch.
The company still reported adjusted EBITDA of negative 170 million yuan (roughly negative $24M) and expects about 700 million yuan (roughly $99M) in annual research spending.
The key unknown is whether factory orders will produce repeatable margins after customization, service and hardware costs, rather than merely prove that pilots can be delivered.

Open at TMTPost Ask Claude ChatGPT

Meituan links AI investment to healthier delivery economics

# TMTPost 5h

Meituan reported second-quarter revenue of 104.6 billion yuan (roughly $14.7B), 77 billion yuan (roughly $10.8B) in research spending and a plan to embed AI in commerce rather than run a standalone token business.

Management expects delivery unit economics to improve year over year but decline sequentially during the seasonal summer peak, when rider subsidies and demand costs rise.
Meituan says its core local-commerce business generated 71.5 billion yuan (roughly $10.1B), while new businesses brought in 33.1 billion yuan (roughly $4.7B) and narrowed losses.
The company describes three AI layers: building large models, applying AI at work and putting agents into products such as ordering, ride-hailing and reservations.
Its LongCat 2.0 model was reportedly trained and run end to end on a domestic-compute cluster, while the Xiaotuan assistant and merchant agent platform connect models directly to transactions.
The labor-policy context matters: a nationwide injury-insurance program for riders and new traffic-safety features raise platform costs while regulators push delivery competition away from permanent subsidies and toward service quality.
This is a more consequential AI strategy than a model launch because Meituan is testing whether agents can become a distribution and operations layer for a large local-services marketplace.

Open at TMTPost Ask Claude ChatGPT

Inference software differences can derail long-context agents

# AI QbitAI 1d

Tests on Qwen3.6-27B found that attention kernels, KV-cache precision and quantization choices changed next-token decisions over long contexts, sometimes breaking network-device tool calls.

The experiment captured more than 100,000 tokens on an RTX PRO 6000 Blackwell while holding weights, prompts and hardware constant, isolating inference-stack effects rather than model-quality differences.
Switching among FlashAttention 2, Flash Inference and Triton produced divergent logits as context length grew; the report attributes this to floating-point accumulation and kernel differences.
BF16 KV cache stayed stable in the reported test, while INT4 produced rapidly increasing top-1 flips and INT8 sometimes recovered after an error.
A concrete failure changed a Cisco interface name and caused later tool calls to issue wrong commands, showing how a tiny numerical deviation can become an operational failure through an agent loop.
The tested NVFP4 and AWQ INT4 paths also failed tool shutdown or command syntax checks, while one community INT8 configuration preserved BF16 activations and completed the workflow.
The architectural implication is that model deployment needs reproducibility and task-level validation across kernels, tensor parallelism and cache formats; identical weights do not guarantee identical agent behavior.

Open at QbitAI Ask Claude ChatGPT

A returning researcher becomes a Tsinghua Yao Class adviser

# AI QbitAI 1d

Xu Mengdi, a robotics-learning researcher who returned from Stanford to Tsinghua in 2025, has become one of four advisers for the university's elite AI-focused undergraduate cohort.

The 2026 Yao Class reportedly enrolled about 90 students, including roughly 27 top informatics competitors and other national-level academic standouts.
Xu's route runs from vehicle engineering at Tsinghua through robotics and machine learning at Johns Hopkins, Carnegie Mellon, Toyota Research Institute, MIT-IBM and Google DeepMind.
Her research emphasizes adaptation: robots should use demonstrations and experience to adjust to new tasks rather than assume that scale alone will cover the nearly infinite real-world task space.
The RoboTool work described in the article lets robots choose unconventional tools and generate executable plans, treating creative tool use as a core form of physical intelligence.
Her appointment signals how Chinese elite-computing education is importing embodied AI and safe-robotics research, not limiting the top pipeline to theoretical computer science or language models.

Open at QbitAI Ask Claude ChatGPT

OpenClaw's fad fades as AI harnesses absorb its ideas

# Analysis QbitAI 1d

A Chinese analysis argues that OpenClaw has not died but has been absorbed into easier vendor-run agents, while the market's attention shifts from installing tools to designing reliable harnesses.

The article says OpenClaw's GitHub activity remains strong, with more than 380,000 stars and 80,000 forks, even though social-media attention has moved on from the original craze.
Chinese vendors including Tencent, Baidu, NetEase, Zhipu and Volcano Engine have produced more than 30 related products, turning local deployment, messaging control and skills into packaged services.
The strategic transition is from a user assembling models, permissions and tools to a model-native harness that supplies file operations, command execution, sandboxes, long tasks and multi-agent coordination by default.
This is a useful Chinese market signal: open-source novelty can become distribution infrastructure for incumbents, while the durable product value shifts toward permissions, context management and runtime reliability.
The star counts and vendor tally come from the article and are not independently verified; the analysis does not show retention, revenue or usage comparisons between OpenClaw and the newer products.

Open at QbitAI Ask Claude ChatGPT

Chinese and Philippine analysts revive joint energy talks in disputed waters

# Policy SCMP 7h

A Shanghai think tank report urges Beijing and Manila to resume joint oil and gas exploration in the South China Sea, alongside lower-sensitivity cooperation on the environment and maritime crime.

The proposal comes from a joint report by Chinese and Philippine analysts, not from a newly announced government agreement.
Its recommended sequencing matters: energy exploration is paired with less politically charged cooperation intended to rebuild working-level trust.
Joint development would require both sides to compartmentalize sovereignty claims, resource rights and operational control, which have repeatedly made South China Sea cooperation difficult.
The report frames economic and maritime cooperation as a pressure-release valve for a relationship increasingly shaped by patrols, alliance politics and US involvement.
There is no evidence in the supplied text that a project has restarted; the immediate signal is that Chinese policy researchers are again presenting cooperation as a viable tool.

Open at SCMP Ask Claude ChatGPT

Could the Irtysh become China's Arctic shipping route?

# Analysis SCMP 15h

China and Russia are reportedly considering the 4,248-kilometer Irtysh River as an inland cargo route to the Arctic, offering Beijing a possible alternative to vulnerable maritime chokepoints.

The river begins in China's Xinjiang, crosses Kazakhstan and joins Russia's Ob River before reaching the Arctic Ocean, making the idea inherently dependent on three-country coordination.
The strategic attraction is redundancy: a river-to-sea route could reduce exposure to contested or congested ocean shipping lanes, though it would still depend on border politics and seasonal navigability.
The proposal links western Chinese infrastructure to Russia's northern maritime ambitions, potentially giving the two countries a logistics project outside their more politically sensitive energy relationship.
The article says the idea is being weighed, not approved; the supplied text does not establish cargo volumes, engineering feasibility or a construction timetable.
Even if commercially marginal, route planning of this kind shows how sanctions, conflict and shipping insecurity are reshaping China's infrastructure imagination.

Open at SCMP Ask Claude ChatGPT

China catalogs 23,000 bronzes held in collections abroad

# Culture SCMP 5h

A 14-year Chinese research project has published 60 volumes documenting more than 23,000 ancient bronzes now held across nearly 30 countries, turning dispersed heritage into a searchable national record.

The catalog covers ritual objects, weapons and chariot fittings that left China and now sit in museums, private collections and auction houses.
Its significance is institutional as much as historical: documentation creates a basis for provenance research, public diplomacy and future repatriation arguments without requiring immediate physical recovery.
The project shows how Chinese cultural policy can use scholarship and data collection to reassemble a national narrative around objects held outside the country.
The report does not say how many objects have disputed provenance or whether the catalog has triggered specific return negotiations.

Open at SCMP Ask Claude ChatGPT

Could US pressure over Iran oil push Beijing toward a harder line?

# Policy SCMP 1h

Washington is targeting China's Iranian oil trade without naming Beijing, while Chinese analysis asks whether sanctions pressure could become a breaking point in the broader relationship.

The report says the United States sanctioned 60 people, companies and vessels linked to Iranian trade, including entities in mainland China and Hong Kong.
China is Iran's largest oil customer, making sanctions enforcement a direct test of Beijing's willingness to absorb secondary-sanctions risk for energy access and strategic ties.
The framing centers on escalation thresholds rather than a single transaction: Washington is testing coercive leverage, while Beijing must balance sanctions exposure, energy security and its relationship with Tehran.
The question is whether additional pressure produces accommodation, evasive trade routes or deeper Chinese-Iranian alignment; the excerpt does not establish Beijing's next policy move.

Open at SCMP Ask Claude ChatGPT

Chinese GPU maker reports a profit boosted by investments, not chips

# Signal TMTPost 3h

Tianshu Zhixin's revenue nearly tripled to 9.46 billion yuan (roughly $1.33B), but its reported adjusted profit relied heavily on a 7.6 billion yuan (roughly $1.07B) unrealized investment gain.

GPU product revenue reached 9.16 billion yuan (roughly $1.29B), with inference-chip revenue up 651.8% to 6.54 billion yuan (roughly $921M) as demand shifted toward serving agents.
Gross margin fell from 50.1% to 17.2%, far below the roughly 55% peer average cited in the report, showing that growth is being bought at aggressive pricing or cost.
The company says it delivered about 52,000 chips by mid-2025 and held a 0.3% share of China's general-purpose GPU market, ranking fifth under the cited market definition.
The reported first-half profit included an unrealized fair-value gain of 7.6 billion yuan (roughly $1.07B) on an investment bought for about 1 billion yuan (roughly $141M).
That gain was not removed from the company's self-defined adjusted-profit calculation, making the headline turnaround a poor proxy for recurring chip economics.
The reported talks with ByteDance about supplying at least 50,000 inference chips remain unconfirmed, so they should not yet be treated as a customer win.

Open at TMTPost Ask Claude ChatGPT

Uzbekistan confirms purchase of China's J-10CE fighter

# Policy SCMP 3h

Uzbekistan has confirmed it acquired China's J-10CE, giving Beijing a second foreign customer and a foothold in a Central Asian fighter market long dominated by Russian aircraft.

Video released by the Uzbek president's office showed at least six aircraft during independence-day events, confirming a deal previously reported but not officially acknowledged.
The aircraft is an export variant of China's J-10 fighter, and the report says earlier accounts pointed to an order of 24 jets.
The strategic significance is market access: a Chinese combat-aircraft sale in Central Asia tests whether Beijing can convert defense-industrial credibility into influence in a region historically tied to Russia.
The report does not establish the contract value, delivery schedule or whether the purchase changes Uzbekistan's wider force structure.

Open at SCMP Ask Claude ChatGPT

Alibaba's Qoder moves coding agents beyond the IDE

# AI QbitAI 1d

Alibaba's Qoder desktop client lets non-programmers describe a task, then uses planning, code generation, browser control and testing to build and verify applications across existing repositories.

The reported execution chain is explicit: understand the goal, produce a plan, edit code, run it, validate the result and revise when the environment exposes a failure.
Qoder uses computer and browser interaction to test generated software in a real environment, making it more than a code-completion interface or a one-shot website generator.
The harder demonstration is repository maintenance: the agent must infer an existing project's stack, rendering flow and camera logic before adding a 3D travel mode without rewriting the whole application.
Alibaba says Qoder supports more than 40 connectors, 70 plugins and 20,000 skills, positioning the harness as an integration layer for project systems and internal tools.
The product is technically notable because it treats context and verification as first-class infrastructure, though the article is a promotional demonstration and gives no independent reliability or customer-retention data.

Open at QbitAI Ask Claude ChatGPT

Sunwoda's battery revenue surges while profit quality deteriorates

# Signal TMTPost 10h

Sunwoda grew first-half revenue 41.5% to 38.179 billion yuan (roughly $5.38B), but net profit fell 29.6% and adjusted profit plunged as hybrid batteries and data-center storage expanded.

Second-quarter net profit reached 489 million yuan (roughly $69M), up 328% sequentially, yet adjusted net profit was only 97 million yuan (roughly $14M) for the half.
Sunwoda says its hybrid-electric-vehicle battery installations ranked first globally in the second quarter, entering a segment historically led by Japanese and Korean suppliers.
Its 684Ah and 588Ah large-format storage cells entered volume supply, with data-center orders benefiting from the power demands of AI infrastructure.
The headline profit is distorted by the accounting presentation of foreign-exchange losses and hedging gains, so reported net income is a poor measure of recurring battery margins.
The company is using hybrid batteries and AI-data-center storage to differentiate from China's largest battery makers, but rising receivables, leverage and negative operating cash flow raise financing and collection risk.
The decisive follow-up metrics are whether storage shipments continue rising and whether adjusted profit margins recover without relying on valuation gains or external funding.

Open at TMTPost Ask Claude ChatGPT

War game puts Chinese coastguard weapons use at 0.3%

# Policy SCMP 23h

A Chinese coastguard research exercise reportedly modeled tense encounters with small vessels and estimated a very low probability of weapons use, highlighting escalation control as a design problem.

The scenario has a coastguard ship protecting fishing boats while fast craft probe around it, forcing operators to distinguish nuisance behavior from a coordinated trap.
The reported 0.3% figure comes from a war-game model, not a forecast of real-world behavior or an empirical probability.
The value of the study is institutional: it suggests Chinese maritime researchers are treating non-lethal signaling, decision latency and escalation management as measurable operational variables.
The supplied text does not identify the model assumptions, scenario set or how the probability was calculated, so the number should not be overinterpreted.

Open at SCMP Ask Claude ChatGPT

Kuaishou pitches niche communities as its next creator moat

# Culture Leiphone 3h

At its annual creator conference, Kuaishou said 270 million users posted content in the past year and argued that local, specialist communities produce better long-term monetization than generic reach.

Kuaishou reported 632 interest communities and 4,112 tags, with specialist creators earning an average 2.2 times the monthly income of broad-traffic creators.
The platform is offering a 500 million yuan (roughly $70M) cash program and 50 billion traffic units for professional creators, alongside a separate program for community-oriented accounts.
The strategy treats regional dialects, sports, niche hobbies and local culture as durable network assets rather than low-value content categories.
Kuaishou plans a modular content platform combining distribution, live interaction, commerce and AI tools, suggesting it wants to sell infrastructure to media and entertainment partners as well as audience access.

Open at Leiphone Ask Claude ChatGPT

Geekplus's warehouse robots grow while its humanoid bet stays small

# Signal TMTPost 2h

Geekplus grew first-half revenue 25.3% to 1.28 billion yuan (roughly $180M), but its warehouse AMR business still supplies 93.6% of sales while humanoids have not reached scale delivery.

New orders rose 35.5% to 2.38 billion yuan (roughly $335M), while gross margin edged up to 35.8%; the core warehouse-fulfillment business generated 1.20 billion yuan (roughly $169M).
More than 75% of revenue came from outside mainland China, where gross margin was 46.2%, making overseas execution central to the company's economics.
The balance sheet shows the cost of project growth: operating cash outflow reached 290 million yuan (roughly $41M), while receivables, inventory and prepayments all increased.
Geekplus has introduced the Gino 1 humanoid and Gravity embodied-intelligence framework, but the report says it has not disclosed meaningful revenue or shipment figures for that business.
The market is therefore asking two separate questions: can the mature AMR business generate consistent profit and cash, and can humanoids become a second revenue curve rather than a valuation narrative?

Open at TMTPost Ask Claude ChatGPT

Human-wildlife conflict rises in China's panda reserve

# Culture SCMP 11h

A new study finds that farmers around China's Giant Panda National Park reported more crop, property, livestock and human losses from wildlife between 2015 and 2024.

The park covers about 22,000 square kilometers, so conservation success has expanded the zone in which protected wildlife and working communities interact.
The study calls human-wildlife conflict the park's most prominent administrative challenge, shifting conservation from habitat protection toward compensation, prevention and local governance.
The data show that a flagship national conservation project can create distributional costs for nearby farmers even when its ecological goals are popular.
The report does not quantify the losses or specify which species caused them, limiting conclusions about whether pandas themselves or other wildlife are the main driver.

Open at SCMP Ask Claude ChatGPT

Huawei adds semantic image search and super-resolution to HarmonyOS

# AI InfoQ China 1d

HarmonyOS 7's Core Vision Kit adds beta APIs for natural-language image search and on-device image enhancement, giving small teams a way to test vision features without building the full stack.

Semantic search maps descriptions such as a meeting whiteboard or blue technical cover to indexed images, while super-resolution reconstructs detail in low-quality pictures.
The APIs shift architecture work from building embeddings, vector search and model serving to managing indexing, permissions, result presentation and device compatibility.
On-device processing can reduce uploads, cloud storage and latency, but the beta documentation still requires real-device testing and explicit handling of user data.
The product lesson is validation order: developers can test whether users want the feature before committing to a custom vision-infrastructure project.
The report says the APIs are beta, and the article does not establish quality across the HarmonyOS device fleet.

Open at InfoQ China Ask Claude ChatGPT

SF Intra-city turns delivery AI and unmanned vehicles into a margin plan

# Analysis TMTPost 3h

SF Intra-city reported first-half revenue of 11.74 billion yuan (roughly $1.65B) and net profit of 350 million yuan (roughly $49M), while using AI dispatch and unmanned vehicles to reduce labor intensity.

The company says its city logistics system forecasts demand, combines orders and dynamically matches riders, producing a reported 22-minute average delivery time for trips under three kilometers.
Its AI agents can accept natural-language delivery requests through WeChat and Alipay, turning conversational intent into a physical delivery workflow without forcing users through the full app funnel.
By midyear, the unmanned-vehicle network covered 124 cities, with more than 1,000 vehicles and over 60,000 average monthly trips across hubs, campuses, retail and food-delivery routes.
The strategic pitch is not immediate rider replacement: AI first improves rider utilization and retention, while standardized route segments create places where autonomous vehicles can gradually substitute for repetitive work.
The company reported delivery-service revenue of 7.19 billion yuan (roughly $1.01B), gross profit of 820 million yuan (roughly $115M) and gross margin of 7.0%.
The model depends on regulation, route standardization and worker stability; without those, AI may optimize a labor-intensive network without materially changing its cost curve.

Open at TMTPost Ask Claude ChatGPT

Kaspersky finds malware risk in DoFun Android car systems

# Analysis IT Home 1h

Kaspersky says attackers can exploit an update service in DoFun Android infotainment systems, potentially turning connected cars into remotely controlled proxy nodes.

The reported chain abuses the TWCore service used for updates and app installation, then deploys a downloader and additional payloads through a file called JarService.
Researchers say the malware can collect device details, receive commands, open webpages and execute further code, including proxy-network workloads.
The exposure is unusually broad because DoFun says its software runs in more than 30 million vehicles worldwide, although the report does not independently verify that installed base.
The issue is a supply-chain problem: the vulnerable boundary is a vendor update service embedded by automakers, not Android in general.
Kaspersky says patches have been released, but the report does not identify affected vehicle brands or quantify confirmed infections.
For automakers, the strategic lesson is that infotainment update infrastructure needs the same isolation and signing discipline as a safety-relevant embedded system.

Open at IT Home Ask Claude ChatGPT

Apple study finds Chinese reasoning training nearly matches English

# AI New Wisdom 13h

Apple and the Hasso Plattner Institute report that Chinese-language reinforcement-learning training trailed English by only 1.1 percentage points in tests across nine models and 11 languages.

The experiments compared English and target-language reasoning rewards across model families including Qwen3, Gemma 3 and SmolLM3, with models ranging from 1B to 8B parameters.
Chinese, French, German and Japanese showed gaps near or below 2 points, while several lower-resource languages had larger losses, linking training economics to available data.
The cross-language results suggest Chinese can be both a useful training language and a recipient of transferred reasoning ability, weakening the assumption that non-English models must reason in English.
The study also found dangerous model-language combinations: some Qwen3 configurations suffered major regressions on unseen English tasks or hard mathematics after single-language training.
The evidence is limited to open models below 8B, programmatic math and logic tasks, single-seed experiments and limited hyperparameter tuning, so production behavior remains unproven.
For Chinese AI teams, the practical implication is that target-language reasoning can be attempted without automatically accepting a large English-performance tax, provided regressions are tested by model-language pair.

Open at New Wisdom Ask Claude ChatGPT

ByteDance delays Doubao 2.2 to improve coding and agent skills

# AI IT Home 3m

ByteDance is reportedly delaying Doubao 2.2 so Seed can make a larger coding and tool-use jump, signaling a preference for slower in-house training over quick model refreshes.

The delay is attributed to a longer pretraining and post-training cycle aimed at coding, tool calling and agent capabilities, rather than to a product or marketing issue.
The model sits between ByteDance's current generation and a reportedly much larger next model, making it a bridge release that must improve coding without waiting for the next architecture.
Chinese coverage frames the decision as a deliberate rejection of model distillation and short-term benchmark chasing, echoing management's stated willingness to lag temporarily while strengthening core research.
The strategic test is whether Seed can turn that extra training time into a visible coding advantage over domestic rivals, rather than merely producing another incremental release.
The report does not establish a new release date or independently verify the claimed competitive targets.

Open at IT Home Ask Claude ChatGPT

China-linked accounts accused of stoking US data-center anger

# Policy Solidot 22h

X and SpaceX allies say about 200 China-linked accounts amplified US opposition to data centers, while a former Twitter executive argues that local power and cost concerns are genuine.

The alleged posts highlighted grid pressure, higher electricity bills and claims that data-center operators profit at public expense; the source does not establish who controlled the accounts.
The counterargument is politically important: Republican and Democratic governors in Texas and New York have both restricted data-center development, suggesting domestic grievances cannot be dismissed as foreign influence.
The dispute shows how infrastructure opposition is becoming securitized, with technology companies able to frame community resistance as information warfare rather than answer questions about taxes, power and water.
The item does not provide an independent attribution study, reach data or evidence that the accounts changed public opinion.

Open at Solidot Ask Claude ChatGPT

Beijing's HICOOL summit markets an ecosystem for hard-tech startups

# Analysis New Wisdom 13h

HICOOL's 2026 summit brought 10,209 projects from 141 countries and regions to Beijing, pairing prize money with industrial partners, policy access and robotics infrastructure.

The summit awarded 200 projects, with 12 first prizes worth 2 million yuan each (roughly $282,000) and a total prize pool above 110 million yuan (roughly $15.5M).
Featured projects covered embodied intelligence, edge inference, RISC-V, quantum information and biotechnology, showing how Beijing is bundling hardware, software and scientific commercialization.
The event's most revealing move was a robotics autonomous-evolution academy led by reinforcement-learning pioneer Richard Sutton, extending an earlier robotics kindergarten into a longer training and industrialization pipeline.
HICOOL says past winners have produced five listed companies, 18 unicorns and 219 specialized technology firms, while winner financing has exceeded 71.9 billion yuan (roughly $10.1B).
The summit functions as more than a contest: it links foreign founders and domestic capital to laboratories, manufacturing partners and local government programs.
That infrastructure model is a competitive advantage if it converts demonstrations into customers; the supplied article does not independently verify the historical success figures.

Open at New Wisdom Ask Claude ChatGPT

China accelerates REIT approvals as the secondary market sinks

# Signal Jiemian 44m

China approved 20 public REITs in 2026, including 10 new listings worth 35.3 billion yuan (roughly $5.0B), even as the sector's total-return index hit a yearly low.

Five REITs received registration approvals over two days, spanning highways, heating networks, hotels, shopping centers, rental housing and industrial parks.
The supply push is turning existing real estate and infrastructure into securities, widening the government's toolkit for recycling capital from completed assets.
The timing is the important signal: approvals accelerated while the public REIT total-return index fell to 925.23, suggesting that administrative supply and investor appetite are moving in opposite directions.
The sector's challenge is not only issuance but valuation and cash-flow credibility; more products will matter if investors believe the underlying assets can deliver stable distributions.
The report says 20 products were approved in 2026, with 10 new listings totaling 35.3 billion yuan (roughly $5.0B), excluding expansions.

Open at Jiemian Ask Claude ChatGPT

China's mRNA developers enter cancer vaccines' second act

# Analysis TMTPost 10h

An interview with Abogen Bio founder Ying Bo uses Moderna's personalized cancer-vaccine trial to explain why Chinese mRNA firms are betting on delivery platforms, oncology and manufacturing depth.

Moderna and Merck's personalized mRNA vaccine reportedly met recurrence-related endpoints in a phase-three melanoma trial, giving the platform a major validation beyond infectious-disease vaccines.
The treatment logic is individualized: sequence a patient's tumor, select targets and use mRNA to instruct the immune system to recognize cancer-specific signals.
Ying's background spans RNA delivery, lipid nanoparticles and chemistry, manufacturing and controls, highlighting that commercialization depends on delivery and production systems as much as on target selection.
Abogen's Chinese perspective is not simply a copy of Moderna's route; the interview frames domestic opportunity around building an mRNA platform where local expertise and capacity remain uneven.
AI may help rank tumor targets and design candidates, but the supplied discussion emphasizes clinical validation, delivery and manufacturing as the harder bottlenecks.
The trial result is a global scientific event, while the China-specific value is how a domestic company interprets it and positions for the next platform race.

Open at TMTPost Ask Claude ChatGPT

AI companies are turning cheap hardware into subscription gateways

# Analysis GeekPark 2h

A Chinese technology essay argues that software companies are building low-cost physical devices to make AI more visible, customizable and habit-forming outside the phone and browser.

Examples include Hugging Face's $399 open-source walking duck, Plaud's $249.99 cellular recording earbuds and ByteDance's wearable AI badge experiment.
The architectural shift is from a self-contained appliance to a thin physical shell around models, tools and subscriptions; the charging case or companion device can become the real computer.
Open hardware and user modification are being used as distribution strategies, especially for developers who can extend a partly finished product rather than wait for a fixed consumer appliance.
The Chinese framing treats cheap, playful hardware as an ecosystem entry ticket and suggests that software companies may be better positioned than traditional hardware firms to define the category.

Open at GeekPark Ask Claude ChatGPT

Chongqing schools may ban smartphones under new youth law

# Policy IT Home 59m

A Chongqing law taking effect September 1 allows schools to ban smartphones and other smart devices on campus, extending a national push to limit minors' phone use.

The local regulation allows schools to prohibit students from bringing smart devices to school or using them on campus, while tightening control over student internet use.
Similar restrictions already exist in Fujian, Guangzhou and Zhengzhou, alongside a 2021 national education ministry rule that generally bars personal phones from classrooms.
The policy treats the school as a managed digital environment, shifting the debate from parental supervision to institutional control of devices and connectivity.
For technology companies and educators, the important question is whether these restrictions change how learning platforms, assignments and identity systems can rely on phones.

Open at IT Home Ask Claude ChatGPT

Hangzhou tests a faster, elder-friendly model for renovating old homes

# Culture Jiemian 34m

A 55-square-meter apartment in Hangzhou's Xihu district is being used as a model for modular, accessible and connected renovation of aging housing stock.

The renovation uses factory-made carbon-crystal wall panels and clip-together installation, cutting reported construction time from about 60 days to 35.
Accessibility features include bathroom supports, a folding shower seat and anti-slip flooring, while concealed sensors monitor gas, leaks and falls.
Emergency buttons and alerts to relatives show how renovation is being tied to home-based elder care rather than treated as cosmetic refurbishment.
The project is positioned as a repeatable method for fragmented, small-unit renovations in old neighborhoods, where dust, noise and neighbor disruption make conventional construction difficult.

Open at Jiemian Ask Claude ChatGPT

Chinese broker asset management rebounds, led by large firms

# Signal Jiemian 46m

Thirty-seven listed Chinese brokerages reported 8.73 trillion yuan (roughly $1.23T) in entrusted assets, up 7.39% from the start of 2026, with more than 70% posting revenue growth.

CITIC Securities remained the largest private asset manager at 1.98 trillion yuan (roughly $279B), while China Securities and Huatai ranked next by entrusted assets.
Growth was not confined to the top tier: 16 of 27 expanding firms grew assets by more than 10%, including several smaller brokerages that more than doubled from a low base.
The pattern suggests a two-track market in which large institutions add licenses and scale while smaller firms seek narrow, specialized niches.
The figures describe assets and operating momentum, not necessarily durable fee profitability; the report does not establish how much of the growth came from market performance versus net inflows.

Open at Jiemian Ask Claude ChatGPT

Chinese robot makers still depend on Nvidia's chips and software

# Analysis Techmeme 8h

A Techmeme roundup highlights reporting that China's robot industry remains dependent on Nvidia silicon and software, an important constraint beneath its rapid physical-AI expansion.

The item is an aggregator summary, so the claim should be treated as a pointer to Wall Street Journal reporting rather than independently verified here.
The China angle is substantive: domestic robot makers may control mechanical integration and deployment while relying on US compute hardware and software ecosystems.
That dependence makes export controls and alternative domestic stacks strategically relevant to robotics, not only to chatbot training.
The cited roundup also says Nvidia's physical-AI business generates about $10B annually, but that figure is a US company claim summarized by the aggregator.

Open at Techmeme Ask Claude ChatGPT

ChinAI	403	0/0		HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk	200	0/20
Pekingnology	200	0/20
Ginger River Review	200	0/20
Sinocism	200	0/20
QbitAI	200	6/10
36Kr	200	0/30
Solidot	200	6/20
Leiphone	200	1/20
GeekPark	200	3/30
IT Home	200	12/60	+38 over cap
SCMP	200	10/50	+2 over cap
Initium	200	1/15
Jiemian	200	8/30	+19 over cap
China Digital Times	200	0/15
The Wire China	200	0/50
Techmeme	200	2/15
Hacker News	200	0/20
V2EX	200	8/50	+15 over cap
Zhidx	200	0/20
InfoQ China	200	2/20
New Wisdom	200	8/15	+1 over cap
TMTPost	200	8/19	+10 over cap

China just wrote its self-driving liability rules for a fleet that barely exists

Beijing hands the speeding ticket to the carmaker

Shanghai's next mayor comes from Finance

A GPU maker's profit came from a stock

Sutton is now running a robot school in Beijing

What developers spent the weekend complaining about

Nvidia's robot cut, and 200 accounts

Threads we are pulling