China Desk

China News, Summarized 21 Oct 2026, 20:08 UTC 34 stories 21/23 sources

Today's brief

Beijing threatens Washington over DeepSeek while its biggest platforms queue for $35 billion of Nvidia cards

Beijing answered in a day, without naming a weapon

Both Washington moves drew a reply. On Tuesday the Ministry of Commerce opposed the DeepSeek order, urged Washington to revoke it, and promised "all necessary measures". It also condemned the MiraclePlus sanctions, again naming no measure. Yesterday I asked whether Beijing would answer before the Shenzhen summit. It did, with a warning and no weapon.

THE TELL: the ministry accused Washington of ignoring understandings the two presidents had reached. That aims the blame at agencies below the White House and keeps the leaders' channel clean. Henry Huiyao Wang, founder of the Beijing think tank Centre for China and Globalisation, writes that Xi and Trump agreed a formal AI dialogue and hotline, with the next meeting in November. My inference: Beijing is protecting that meeting.

Two versions of the story. The New York Times, in Techmeme's summary, pairs China's vow of countermeasures with its rejection of FBI and AI-lab accusations of illicit distillation, meaning training on rivals' outputs. Leiphone's morning digest gives it five sentences on what Homeland Security ordered, ending with Washington "claiming" a threat. No FBI, no distillation.

Go deeper on this section: ClaudeChatGPT

Nvidia cannot fill what China ordered

The order is enormous. Reuters reported, per The Next Web, that ByteDance, Alibaba and Tencent ordered over 3 million RTX PRO 5500 cards at about $11,400 each, above $35bn. Nvidia told partners they may get 5% to 10%. Output is capped near 3 million worldwide in 2027, while Chinese orders total over 6 million. Deliveries start in the first quarter.

Why this card. It is a workstation GPU built to clear US rules: the Blackwell B300 is barred and the H200 needs licences. Leiphone's sourcing adds 84GB of memory that splits into two 42GB slices for inference, theoretical compute about double the cut-down RTX PRO 6000D, and Chinese quotes of 80,000 to 120,000 yuan (~$11,000 to $17,000).

Beijing is rationing, not only retaliating. Technology.org's write-up says China's industry ministry asked buyers in late September how many cards they wanted and why, then said it intends to approve. The state threatening countermeasures over a US procurement ban is also deciding who queues for an American supplier. At 5% to 10% of 3 million, the three firms get 150,000 to 300,000 cards. That arithmetic is mine.

WORTH NOTING: Leiphone relays unconfirmed overseas chatter that Nvidia is moving GB202 chip supply away from the RTX 5090. Chinese prices for that card are already up over 50% in a month, past 50,000 yuan (~$7,000).

Go deeper on this section: ClaudeChatGPT

DeepInfer opens the door to two vendors, mostly one

The evidence is lopsided. TMTPost's analysis of DeepSeek's DeepInfer release says Ascend and Nvidia are listed as peer build targets. But published benchmarks lean on Nvidia, Ascend side-by-sides are thinner, and the drop is framed as offline inference. Cross-node cache management and load balancing are missing. That matches yesterday's read: the engine is out, the cache tier is not.

OpenAI moved within a day. Zhidx reports that OpenAI's API now offers 12-hour cache retention on GPT-5.6 and later. Cache-write pricing rises from 1.25 to 2.5 times the uncached input price. Codex and Work users get "permanent" caching. Zhidx suspects a billing waiver rather than stored tensors, and notes DeepSeek's design targets up to three days against OpenAI's 30-minute default. Cache lifetime is becoming a price lever dressed as infrastructure.

Go deeper on this section: ClaudeChatGPT

What people are chewing on

Starbucks wants a purchase for a seat. From 22 October Starbucks China requires a purchase to sit. Boyu Capital, the private-equity firm, took 60% in April; TMTPost says nothing shows it demanded this. The free-seating promise began as damage control after a 2018 Philadelphia arrest, and the US rolled it back last year. The same day the chain pledged over 300 million yuan (~$42m) for peak-hour staff and about 1,000 refits.

The third place became a rent line. My read: it was a brand premium, and once Luckin and Cotti sold coffee at a fraction of the price, the sofa had to pay for itself.

Who burns the compute? A media claim, relayed by Leiphone, says 55% of China's AI token use goes to short drama and video generation, 24% to e-commerce ads and 15% to research. It also puts nearly 200,000 yuan (~$28,000) of compute on one 12-episode drama, ten times yesterday's TMTPost production-cost figure. Doubao's PR chief called the data unreliable, saying only about 10% of ByteDance's compute goes to video generation. Industry tokens versus one company's compute: the denial settles nothing.

Trending, manufactured. Sam's Club says the "quality collapse" videos that hit Weibo's top hot search were AI-generated. Products shown as pulled are on sale or were never stocked, it says, and thousands of near-identical posts appeared. That is the company's claim. Consumer outrage is now a farmable commodity.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • MiniMax was not a tie. Yesterday I said M3.1 Flash tied DeepSeek's V4.1-Flash-1017 on agent benchmarks. Leiphone's numbers show a split. DeepSeek leads coding, 77.0% to 73.7% on DeepSWE v1.1. MiniMax leads Automation-Bench. On a 3D-campus build MiniMax took 55.4 minutes against about 121, at roughly 6.5 yuan (~$0.92) each.
  • Agent gates are global. On 15 October I framed China's apps as deciding agents must be invited in. Zhidx says 155 apps refused Doubao's phone assistant, while Amazon, Yelp, eBay, Delta and JPMorgan block Meta's Muse. Muse reached 10.9 million downloads in 40 days. My frame was too local. The fight is over who owns the checkout.
  • Honor absorbs and raises at once. On 19 October it lifted the X80 Pro Max by 500 yuan. Today's gaming phone, the WIN2, starts at 4,999 yuan (~$704) after subsidy. The CEO says Honor absorbs memory costs, but Honor concedes it cannot say for how long.
  • The export draft widened. Sinocism's Sharp China describes proposed controls on AI, solar-related technology and machinery. I have not seen the solar or machinery text, so my "ledger, not gate" read covers only AI.
  • DeepSeek is investing in robots. On 14 October I said its round was filling with suppliers and customers. Leiphone reports, unverified, that DeepSeek is in the startup of Yu Tao, a former Tencent AI Lab researcher, and may lead Winverse AI at about $400m. In August it put 140.8m yuan (~$19.8m) into Unitree's IPO placement.
  • A 17-year-old may have left for Palo Alto. Chen Guangyu, co-first author of the Attention Residuals paper that Moonshot AI's Kimi K3 uses, left Moonshot on 15 October. An anonymous tracker says he joined Discovery Loop, the Jeff Dean startup reportedly seeking a $50bn valuation. Unconfirmed; the company won't comment. If true, 2-million-yuan starting packages are not what holds talent.

Go deeper on this section: ClaudeChatGPT

The river · 34

China proposes export controls covering AI technologies

# ChinaTalk 6h

Chinese coverage treats the proposed commerce and science ministry controls as a strategic choice: whether Beijing will trade the global reach of open AI for leverage over key technologies.

More
  • The draft proposes adding 40 technologies across AI, photovoltaics and biomedicine to China’s restricted or prohibited export catalogue; the article focuses on AI provisions.
  • The catalogue has historically targeted technologies where Chinese production dominance could create leverage, including rare-earth processing and battery materials.
  • The draft is open for comment and is not yet law, so the proposal signals direction more clearly than immediate enforcement.
  • The analysis says the AI sections reach beyond model weights to areas including chip design, memory, cybersecurity and AI-content detection.
  • Export licensing is not simply a central-government switch: the article examines how provincial and central authorities may share influence over approvals.
  • For Chinese AI labs, tighter controls could complicate the open-weight strategy that helped build adoption abroad; the effect depends on what is ultimately controlled and how licenses are administered.
Open at ChinaTalk Ask Claude ChatGPT

Hong Kong independent bookstores face a chilling national-security crackdown

# Initium 21h

Chinese-language coverage documents searches, arrests and event exclusions affecting Hong Kong bookstores, showing how national-security enforcement is reshaping everyday cultural life.

More
  • The report says national-security police searched several independent bookstores in 2026 and detained owners or staff; multiple outlets were also excluded from the Hong Kong Book Fair.
  • Some shops have closed or announced plans to close, while reopened stores face relocation and staffing pressures.
  • Booksellers describe adjusting operations amid uncertainty over what books can be sold and which events can be held.
  • The article’s account links enforcement not only to publication choices but to the survival of small cultural institutions and their public programs.
  • One bookseller says the possibility of arrest has changed ordinary working routines, including keeping a lawyer’s number and preparing for detention.
  • This is a direct picture of how political controls can alter cultural infrastructure without a single formal ban defining every boundary.
Open at Initium Ask Claude ChatGPT

Taiwan expels Chinese exchange student over a TikTok post

# Initium 11h

Taiwan revoked a Chinese student’s entry permit and imposed a five-year reentry ban over a post depicting President Lai; Chinese coverage focuses on the proportionality and legal basis of the penalty.

More
  • The student, visiting from the University of British Columbia, posted a staged image of Lai with pro-unification hashtags, according to the report.
  • Taiwan’s immigration authority said the post included text advocating the elimination of Taiwan’s sovereignty and ordered the student to leave.
  • The student removed the post and apologized, saying he had no concrete plan to cause physical harm.
  • The case has prompted debate over how authorities decide that speech warrants forced departure and the maximum five-year exclusion period.
  • The incident sits at the intersection of cross-strait political expression, immigration discretion and public perceptions of security; the excerpt does not provide a court ruling on the decision.
Open at Initium Ask Claude ChatGPT

PingCAP confirms breach of internal systems and source code

# InfoQ China 5h

PingCAP confirms unauthorized access and data copying, while the attackers’ claimed scale remains unverified; the company says it has found no impact to Chinese customers’ production databases.

More
  • PingCAP says an attacker used stolen service-account credentials to access some internal systems and copy data and source code; the company dates the access period to September and October.
  • The attacker claims a 10.6TB haul including tenant records, API keys, SQL histories and private repositories, but those figures have not been independently verified.
  • PingCAP says exported records included user and organization details, project associations, SQL-editor text and some commercial data.
  • The dispute over API-key exposure remains material: the attacker alleges plaintext keys, while PingCAP says the relevant exported table contained encrypted keys.
  • The company says an external review found no evidence of attacker commands or persistent access in the production systems reviewed, and no evidence that data stored in TiDB was affected.
  • Even without direct access to customer databases, a cloud control plane can expose query history, account metadata and support information that may aid follow-on attacks.
Open at InfoQ China Ask Claude ChatGPT

Chinese apps block ByteDance’s Doubao phone assistant from automation

# Analysis Zhidx 11h

Chinese coverage frames app restrictions as a fight over who controls user actions and platform access: 155 apps reportedly reject Doubao’s automation, echoing conflicts around global personal agents.

More
  • The report says 155 apps rejected automated operations by Doubao’s phone assistant, including major services from Tencent and Alibaba as well as competing AI products.
  • Platforms cite rules, security and account risks; the article argues that control of advertising exposure, user data and transaction flows is also at stake.
  • An agent that bypasses an app’s interface can disrupt the business model built around discovery, recommendations and checkout, even when the user authorizes the action.
  • The story connects China’s app restrictions to disputes involving Meta, Amazon and other foreign services, suggesting a shared platform-versus-agent conflict rather than a uniquely Chinese technical issue.
  • The piece describes restrictions and industry arguments, not a settled legal ruling on whether users may delegate online activity to an agent.
Open at Zhidx Ask Claude ChatGPT

Li Auto unveils its in-house battery and electric-drive system

# Analysis QbitAI 19h

Li Auto detailed a vertically integrated EV stack spanning cells, motors, power electronics and charging, and offered lifetime protection for qualifying battery-fire losses.

More
  • The company says its 5C cells improve charging speed without sacrificing energy density, using electrode and pack-design changes; these are manufacturer claims.
  • A battery-state algorithm is reported to keep charge-estimation error within 3%, while a winter power-control method is claimed to raise peak output by more than 40%.
  • The company’s six-part safety design combines impact and water protection, thermal barriers, manufacturing checks and cloud risk monitoring.
  • Li Auto says a fire caused by a defect in its own battery that totals the vehicle would qualify for a replacement vehicle; collision-related fire coverage is structured differently.
  • The announcement shows the strategic appeal of controlling battery, inverter, motor and charging behavior together, but test conditions and warranty exclusions matter to the real-world value.
Open at QbitAI Ask Claude ChatGPT

China’s offshore-trust tax reporting window closes as capital scrutiny tightens

# Policy SCMP 7h

Chinese coverage frames the offshore-trust tax deadline as a sign Beijing is asserting tighter control over overseas wealth, with possible spillover into demand for digital dollars.

More
  • The report says a grace period for disclosing and settling certain offshore-trust tax liabilities ends on October 22.
  • It presents the move as part of broader scrutiny of capital leaving China and of offshore assets held by wealthy families.
  • The article connects the tax measure to tighter attention on cryptocurrency, stablecoins and real-world assets, but the excerpt does not specify the rules governing each.
  • The possibility of a digital-dollar shift is the headline’s projected consequence, not a reported outcome or measured change in usage.
  • The policy’s actual reach will depend on enforcement and on how affected families restructure assets; the provided excerpt does not describe either in detail.
Open at SCMP Ask Claude ChatGPT

China’s embodied-AI push is building training around real workstations

# Analysis Leiphone 13h

China is shifting embodied-AI training from demo venues toward real factory and service tasks, where data collection, validation and deployment feedback can form a reusable pipeline.

More
  • A research report cited by the article says more than 70 embodied-AI training sites were operating by June, but high-quality real-robot data remains scarce relative to the millions of hours thought necessary.
  • The central technical bottleneck is the gap between simulation or lab training and new physical environments, where robots may have to relearn tasks.
  • A government program launched by the industry ministry and state-assets regulator uses real workstations as training and validation settings, linking data collection directly to deployment.
  • Shanghai examples span robot data factories, shared prototyping and testing facilities, and production-line trials; the article says one robot completed nearly 80,000 battery-line operations with 98% uptime.
  • The useful infrastructure is not just floor space: it is a loop that captures failures, labels interaction data, updates policies and tests them back on physical equipment.
  • The reported deployments are early evidence, not proof of general-purpose transfer across factories, robot bodies or task types.
Open at Leiphone Ask Claude ChatGPT

Shanghai pilots identity codes for workplace AI agents

# Policy Leiphone 13h

A Shanghai platform is assigning identifiable credentials to agents, as Chinese policy discussions move from agent capability toward identity, delegated authority, auditability and responsibility.

More
  • The Shanghai launch involved 30 organizations and is presented as a local implementation of national agent-interoperability standards and policy guidance.
  • The policy direction described in the article includes registration, discovery, capability declarations and differentiated rules for tool access and autonomous decisions.
  • The account argues that identity alone is insufficient: an organization still has to specify who delegates work, which actions require human approval and who owns outcomes.
  • Examples from Chinese companies show agents performing defined jobs, such as production planning or business reporting, with human review retained at decision points.
  • As agents call tools and hand tasks to other agents, logs, permission boundaries and approval checkpoints become essential to tracing errors or unauthorized actions.
  • This is a governance and organizational design question as much as an identity technology: the rules will determine whether agents can operate across institutional systems.
Open at Leiphone Ask Claude ChatGPT

Virginia Tech accelerates closure of Xidian joint program amid Pentagon pressure

# Policy SCMP 29m

Virginia Tech is ending its undergraduate partnership with China’s Xidian University earlier than planned, part of a wider US university retrenchment under Pentagon pressure.

More
  • The university is bringing forward the program’s planned end from 2029, according to a spokesperson and a letter reviewed by the SCMP.
  • The article places the decision amid more than 80 termination-related actions across 30 institutions, pointing to a broader tightening of US academic ties.
  • For China, the issue is not only research access: joint education programs are also channels for sustained institutional and people-to-people exchange.
  • The source attributes the shift to Pentagon pressure, but does not establish that every university action is driven by the same federal requirement.
Open at SCMP Ask Claude ChatGPT

Chinese economist argues public investment should lead demand revival

# Policy Pekingnology 3h

A Chinese economist makes the case that infrastructure, public services and housing investment can counter weak demand, challenging the view that more public borrowing must worsen debt.

More
  • The article says the Politburo called for stronger countercyclical measures and domestic demand after growth slowed in the second quarter, then describes NDRC efforts to accelerate investment.
  • The proposed “six networks” include computing power, next-generation telecommunications and logistics, alongside utilities and power infrastructure.
  • Zhang Bin argues that investment in public services, affordable housing and upkeep of existing facilities could raise incomes and spending, not merely expand capacity.
  • His counterargument on debt is that weak growth can itself worsen debt ratios, while productive borrowing may improve repayment capacity.
  • This is an argument by a CASS economist, not a reported change in fiscal policy or a commitment to a specific spending package.
Open at Pekingnology Ask Claude ChatGPT

Chery targets China’s premium leisure-motorcycle market

# Analysis GeekPark 5h

Chery is entering motorcycles with an ambition to sell a lifestyle and technology platform, betting on higher-end leisure riders rather than shrinking commuter demand.

More
  • The company’s unit says China’s motorcycle exports are large but lower-margin than leading foreign brands, while domestic demand is shifting toward larger leisure models.
  • Its initial products include a fast-charging electric scooter and a range-extended touring motorcycle, with automotive battery, thermal-management and electrical-system know-how adapted for two wheels.
  • The company says it is combining automotive engineers with motorcycle-industry veterans after learning that four-wheel expertise alone does not translate directly to rider needs.
  • Its proposed AI features prioritize navigation and safety warnings; managers say they rejected a cornering-angle leaderboard because it could encourage dangerous riding.
  • The strategy is broader than vehicle sales: the company wants to build accessories, customization and touring services around recreational riding, though the article gives no evidence that this ecosystem is established.
Open at GeekPark Ask Claude ChatGPT

Chery lays out a broad in-house chip, model and autonomy roadmap

# Analysis GeekPark 10h

Chery presented plans spanning vehicle chips, enterprise and product models, driver assistance, robotaxis and electric farm equipment, casting itself as a technology group rather than only a carmaker.

More
  • The company says its Ziwei chip family will cover central compute, silicon-carbide power and control functions; its C100 central chip is planned for 2027.
  • Chery says its CheryMind model draws on company engineering and operations data and is used across hundreds of internal systems and workflows.
  • Its autonomy stack includes a world model and task-specific models, with the company claiming tens of thousands of generated driving scenarios and a long-term global robotaxi target.
  • The report also describes electric tractors and a range-extender motorcycle, reflecting a strategy to reuse vehicle electrification and software capabilities across adjacent products.
  • Most performance, adoption and launch claims are company statements; the C100’s eventual deployment will be an important test of the roadmap.
Open at GeekPark Ask Claude ChatGPT

Tencent opens its Yuanbao search service to developers

# AI Zhidx 7h

Tencent is exposing its web-search infrastructure through API, MCP and skill integrations, including structured vertical data that agents can consume without scraping pages.

More
  • Tencent says the index covers hundreds of billions of items and has been used inside products including WorkBuddy and ima; those scale claims come from the company.
  • The API returns ordinary web results as well as structured sources for queries such as weather and stock prices, potentially avoiding a separate crawl-and-parse step.
  • That interface is useful for agents because retrieval errors or stale information can propagate into later tool calls and actions.
  • The article reports hands-on tests for news, policy, academic and developer searches, but does not provide independent accuracy or freshness benchmarks.
Open at Zhidx Ask Claude ChatGPT

MiniMax M3.1 Flash claims strong agent results at lower cost

# AI Leiphone 15h

MiniMax says its new Flash model leads selected agent benchmarks and completed one test task faster than a DeepSeek setup at similar API cost, intensifying China’s low-cost model competition.

More
  • The company reports first place on Automation-Bench and a tie with Opus 5 on BankerToolBench, while DeepSeek remains ahead on the cited software-engineering benchmark.
  • In one reported 3D-campus task, MiniMax Code with M3.1 Flash finished in about 55 minutes versus roughly 121 minutes for an OpenCode and DeepSeek configuration; both model bills were about 6.5 yuan (roughly $0.90).
  • The report says the model’s benchmark score improved as its task-weighted cost fell, but these comparisons depend on the benchmark setup and do not establish general superiority.
  • MiniMax lists pricing of 1 yuan per million input tokens, 4 yuan per million output tokens and 0.1 yuan per million cached input tokens.
  • The competitive picture is task-dependent: DeepSeek is stronger on coding in the cited comparison, while MiniMax is positioning around multi-step enterprise agents.
Open at Leiphone Ask Claude ChatGPT

Taiwan’s low-cost attack drone could test PLA air defenses

# Analysis SCMP 6h

A Taiwan-made drone modeled on Iran’s Shahed is being positioned as a low-cost way to saturate Chinese air defenses, though analysts differ on how it might be used.

More
  • The manufacturer says the flight-tested system could reach military targets on the mainland; that capability claim is not independently confirmed in the excerpt.
  • Analysts see potential value in inexpensive massed drones, drawing on lessons from Ukraine and the Middle East.
  • The article says experts disagree on deployment concepts in a conflict with Beijing, so the strategic role is not settled.
  • The relevant measure is likely not one airframe’s sophistication but whether Taiwan could field, coordinate and sustain enough systems to complicate defenses.
Open at SCMP Ask Claude ChatGPT

DeepSeek’s open inference code aims to make Huawei and Nvidia targets peers

# AI TMTPost 9h

DeepSeek’s inference package is designed to support Huawei Ascend and Nvidia GPUs in one codebase, potentially making serving hardware easier to switch as supply and cost change.

Also: GeekPark

More
  • The release targets offline inference and lists both Ascend processors and Nvidia data-center GPUs as build targets, according to the report.
  • If the same serving stack can run on both, teams could choose hardware based on availability, power and cost rather than maintaining separate software forks.
  • Inference is a plausible early opening for alternative accelerators because models are already trained and the economics hinge on latency, throughput and utilization.
  • The engineering claim is not yet fully proven: public benchmarks described in the article lean on Nvidia hardware, with less comparable Ascend data.
  • A dual build target is only a starting point; production fleets also need caching, load balancing, failure handling and a complete control plane.
Open at TMTPost Ask Claude ChatGPT

China refits an older destroyer as the navy rotates out legacy ships

# Analysis SCMP 10h

A Type 051C destroyer has reportedly appeared in sea trials after a refit that may replace Russian air-defense systems with domestic ones, amid a wider naval fleet transition.

More
  • The ship is the lead vessel of its class, and social-media images appear to show domestic air-defense equipment replacing Russian-made hardware.
  • The identification and equipment changes have not been independently confirmed, according to the report.
  • The refit could indicate that the PLA Navy is extending the useful life of some older hulls while newer ships enter service.
  • The significance depends on what systems were actually installed and whether the refit materially improves the ship’s combat capability.
Open at SCMP Ask Claude ChatGPT

Starbucks China’s seating rule tests the value of its “third place”

# Culture TMTPost 10h

Starbucks China is moving toward requiring purchases for seating, reopening a debate over whether its community-space promise can survive price competition and tighter store economics.

More
  • The company confirmed a nationwide “purchase to sit” policy, according to the article, as local coffee chains compete at far lower price points.
  • The article argues that the policy cannot be explained only by the new Chinese joint-venture owner: Starbucks had already reversed its no-purchase seating rule in North America.
  • That earlier open-door policy followed a 2018 racial-discrimination scandal, showing how an operating rule became part of the brand’s moral identity.
  • In China, the shift touches a consumer expectation built around spending time in stores, not just buying coffee.
  • Starbucks and its new partner have a stated expansion ambition, but the report does not establish that the seating rule will improve sales or customer loyalty.
Open at TMTPost Ask Claude ChatGPT

China’s aging debate asks where AI care ends and human presence begins

# Culture 36Kr 15h

A discussion involving older people and care specialists presents AI as a support for aging in place, while warning that relationships, patience and dignity cannot be automated.

More
  • Participants describe older adults learning to use AI and smart-home devices, but also encountering misinformation and the digital divide.
  • One speaker argues that technology can improve monitoring and convenience but cannot replace the time and relationships that determine care quality.
  • The discussion also rejects treating older people only as recipients of services; participants want opportunities for agency and recognition.
  • The article reports community programs and device deployments, but does not evaluate their long-term effects on safety or care outcomes.
Open at 36Kr Ask Claude ChatGPT

Chinese startup pitches AI for children as a creative copilot, not a tutor

# Culture Leiphone 17h

CapaLeap is positioning its children’s AI products around guided exploration and making things, responding to concern that answer-first chatbots can displace independent thought.

More
  • The article cites surveys in which many Chinese students use AI for homework, while parents report concerns about dependence and reduced independent thinking; the excerpt does not detail survey methods.
  • The company says its agent should ask questions, break down tasks and help children create, rather than simply supply answers or act as an emotional companion.
  • Its architecture combines a shared agent and memory layer with age-specific devices, intended to carry interests and projects across products.
  • That design makes persistent memory central for personalization, but also raises questions about children’s privacy, consent and age-appropriate interactions that the article does not address.
  • The strategy deliberately sits outside the crowded test-score and learning-tablet market; whether families will pay for open-ended creative learning remains unproven.
Open at Leiphone Ask Claude ChatGPT

Hikvision applies sensing and small-sample AI to industrial inspection

# Analysis Leiphone 17h

Hikvision is extending its security-camera expertise into industrial sensing, pairing diverse inspection hardware with models intended to reduce the data burden of factory deployments.

More
  • The company’s exhibit spans X-ray inspection, acoustic sensing and high-temperature imaging, illustrating an approach based on multiple sensor types rather than cameras alone.
  • Hikvision says its industrial inspection model can be trained with dozens of labeled images instead of thousands, and in some settings with only a handful of samples.
  • The report also cites a vehicle-parts ultrasound application with claimed inspection-speed gains above 60% and defect accuracy of at least 99.5%.
  • These are company-reported examples; the key test is whether small-sample performance holds across product changes and less controlled factory environments.
Open at Leiphone Ask Claude ChatGPT

Chinese analysts urge Europe to treat Beijing as an Arctic partner

# Analysis SCMP 7h

After the EU warned of China’s growing Arctic alignment with Russia, Chinese analysts argued for cooperation on shipping rather than framing Beijing as a security threat.

More
  • The EU policy, as summarized by the article, identifies Russia as its main Arctic threat and raises concern about China’s ties with Moscow.
  • The Chinese analysts’ response emphasizes shared commercial interests in Arctic shipping and rejects linking cooperation automatically to security alignment.
  • That framing reveals Beijing’s effort to present its Arctic ambitions as economic and pragmatic, while the EU frames them through strategic competition.
  • The excerpt does not establish what specific joint shipping projects are feasible or how European governments will respond.
Open at SCMP Ask Claude ChatGPT

China’s AI industry is trying to make tokens a usable measure of work

# Signal Leiphone 13h

The article argues that token counts are becoming a pricing interface for AI, but real economics depend on model quality, hardware utilization and task success rather than token volume alone.

More
  • An industry official is quoted as saying China’s average daily token calls had passed 500 trillion by June 2026; the article does not provide an independent measurement method.
  • Token prices can vary with model, time of day, cache status and modality, so cost per million tokens is not a reliable cross-model comparison by itself.
  • The report describes a shift in compute evaluation from installed accelerators toward delivered inference throughput, latency, availability and cost per successful task.
  • It gives vendor-reported examples of faster model adaptation to domestic chips and lower inference cost from cross-cluster scheduling, but results are workload-specific.
  • For robotics, tokenized visual input is only one part of cost: power, action reliability, inference delay and equipment utilization also determine the economics.
Open at Leiphone Ask Claude ChatGPT

China’s delivery-robot race still has to prove it can replace labor

# Analysis TMTPost 10h

A major funding round for Neolix contrasts with Meituan’s operational focus, while the article argues that autonomous vehicles struggle most with the human-heavy last mile.

More
  • Neolix announced a nearly $400 million round, while the report says Meituan postponed promotion reviews to align them with a new vehicle launch.
  • The article distinguishes standardized trunk or shuttle routes from doorstep delivery, where gates, elevators, customer availability and handoffs still require people.
  • A vehicle can reduce labor on a predictable route yet add a transfer or waiting step that erases savings in dense neighborhoods.
  • The business case may be clearer in depot-to-area trunk transport than in replacing couriers end to end.
  • The account relies partly on unnamed industry sources and does not provide unit economics for either operator.
Open at TMTPost Ask Claude ChatGPT

Chinese mathematician invites researchers to tackle 1,000 open problems

# Culture SCMP 8h

Nanjing University mathematician Sun Zhiwei has assembled more than 1,000 number-theory problems, arguing that asking strong questions remains valuable in an age of AI-generated answers.

More
  • The article presents Sun’s problem collection as a long-running contribution to the research community, not as a benchmark designed to rank AI systems.
  • His argument centers on mathematical curiosity and problem formulation as capabilities that matter even when models can solve difficult tasks.
  • The excerpt does not describe the problems’ difficulty distribution or any AI evaluation results.
Open at SCMP Ask Claude ChatGPT

Chinese robotics team plans a public livestream of embodied-AI training

# AI Leiphone 10h

World-model developer Shengshu says it will livestream two weeks of Motubrain 2.0 training, exposing experiments and failures before the embodied model is open-sourced.

Also: New Wisdom

More
  • The planned stream covers pretraining, task-specific post-training, agent task execution and experiments in recursive self-improvement.
  • Unlike a robot demo, the format is intended to show how capabilities change during training and testing, though livestream visibility is not a substitute for reproducible evaluation.
  • The team says Motubrain 2.0 will be open-sourced, but the report does not specify a release date or license.
Open at Leiphone Ask Claude ChatGPT

OpenAI expands prompt-cache retention as DeepSeek pushes longer windows

# AI Zhidx 13h

OpenAI is offering a 12-hour API cache option, while its claim of persistent cache for Codex users remains unclear; Chinese coverage explicitly compares the change with DeepSeek’s longer target.

More
  • The article says the API option applies to newer models and raises cache-write pricing while leaving cache-read pricing unchanged.
  • It cautions that cache eligibility, actual cache hits and request latency are separate properties; a longer retention window does not guarantee reuse or lower latency.
  • The report contrasts OpenAI’s change with a DeepSeek inference framework designed for a cache-retention window of up to three days.
  • The competitive significance depends on real hit rates and serving costs, which the article does not provide for a like-for-like workload.
Open at Zhidx Ask Claude ChatGPT

Chinese startup raises more than $14 million for a physical-space engine

# AI 36Kr 17h

Ether Technology says its spatial-intelligence infrastructure will turn changing real environments into reusable, machine-readable assets for robots and other physical AI systems.

More
  • The company raised more than 100 million yuan (roughly $14 million) across seed and angel rounds, with Huawei’s Hubble fund and several robotics and technology firms among the investors.
  • Its proposed stack fuses visual, location, inertial and force-touch data, then aligns them in space and time to represent object state, geometry and interaction.
  • The company says its assets will encode collision relationships and physical constraints, going beyond a static 3D scan toward a reusable representation of how a space behaves.
  • It is prioritizing hand-interaction data because manipulation captures contact, object changes and task feedback that can train robot policies.
  • The investment thesis depends on whether spatial assets can transfer across devices and tasks; the report describes a roadmap, not a demonstrated general-purpose engine.
Open at 36Kr Ask Claude ChatGPT

Beijing’s Taiwan affairs chief meets a US academic delegation

# Policy SCMP 9h

Song Tao met US scholars in an uncommon direct exchange, reaffirming Beijing’s cross-strait red lines while presenting peaceful development as its preferred course.

More
  • The meeting joined the director of China’s Taiwan Affairs Office with a delegation from the National Committee on American Foreign Policy.
  • The official framing links cross-strait policy to broader leader-level diplomacy between Beijing and Washington.
  • The meeting offers a channel for signaling positions to US policy specialists, but the excerpt gives no evidence of a change in policy or a negotiating breakthrough.
Open at SCMP Ask Claude ChatGPT

Chinese sellers find Latin America requires more than cheap products

# Analysis TMTPost 7h

Chinese e-commerce and consumer brands are finding that Latin American growth depends on adapting products, service and financing to local habits, not simply shipping low-cost goods.

More
  • The article says online retail growth in Mexico and Brazil attracts exporters, but consumers compare prices closely and expect products to work in local conditions.
  • Examples in the text describe adapting security-camera bundles and personal-care products to housing, usage and electrical requirements.
  • The commercial lesson is localization: service capacity, transparent delivery and market-specific product choices can matter as much as low prices.
  • The excerpt is incomplete and does not provide a full account of the companies or markets surveyed.
Open at TMTPost Ask Claude ChatGPT

Chinese microdramas are experimenting with more credible workplace stories

# Culture TMTPost 9h

A Chinese AI-generated medical microdrama has drawn attention for putting professional knowledge and work routines at the center, rather than using a job merely as romance scenery.

More
  • The article describes a series whose emergency-room plot uses triage, trauma assessment and clinical terminology to create its dramatic momentum.
  • Viewers with medical training reportedly discuss the scenes and explain procedures in comment threads, turning audience response into an informal accuracy check.
  • The piece contrasts this style with workplace dramas criticized for implausible salaries, roles and professional behavior.
  • The argument is about changing audience expectations and short-form storytelling; the excerpt does not independently assess the medical accuracy of the show.
Open at TMTPost Ask Claude ChatGPT

ManageEngine brings more than 30 IT agents to China

# AI Zhidx 15h

ManageEngine launched a locally adapted suite of IT-operations agents for Chinese customers, emphasizing on-premises context, domestic infrastructure compatibility and no extra token charge.

More
  • The company says the agents cover observability, service management, endpoint security, identity and security-event management, with more planned by year-end.
  • The offering is designed to reuse context from existing ManageEngine products and support domestic databases, operating systems and middleware.
  • The company plans to invest 425 million yuan (roughly $60 million) in China over five years, according to its announcement.
  • Claims of reduced resolution time and automated ticket handling are vendor projections, not independently measured results.
Open at Zhidx Ask Claude ChatGPT

Australian Open is expanding its China audience beyond live matches

# Culture 36Kr 12h

The tournament is adding a China-wide one-point championship and lifestyle programming as younger fans seek participation and platform-specific content, not only match broadcasts.

More
  • The Australian Open plans qualifying events in five Chinese cities, with the China winner advancing to a final in Melbourne.
  • Its representative says the event’s China audience has grown around Chinese players’ success and that younger fans engage through merchandise and experiences as well as broadcasts.
  • The organization adapts content for platforms such as Weibo, WeChat and Xiaohongshu, reflecting a fragmented media environment and localized audience behavior.
  • Audience and viewing figures in the interview are supplied by the organizer, not independently audited in the report.
Open at 36Kr Ask Claude ChatGPT
Source ledger
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 2/20
QbitAI 200 2/10
36Kr 200 5/30
Solidot ERR 0/0 ConnectionError: HTTPSConnectionPool(host='www.solidot.org', port=443): Max retries exceeded with url: /index.rss (Caused by NewConnectionError("HTTPSConnection(host='www.solidot.org', port=443): Fail
Leiphone 200 12/20
GeekPark 200 4/30
IT Home 200 12/60 +36 over cap
SCMP 200 10/50 +6 over cap
Initium 200 6/15
Jiemian 200 8/30 +14 over cap
China Digital Times 200 0/15
The Wire China 200 0/50
Techmeme 200 0/15
Hacker News 200 0/20
V2EX 200 8/40 +30 over cap
Zhidx 200 10/20 +3 over cap
InfoQ China 200 8/20 +12 over cap
New Wisdom 200 8/15 +4 over cap
TMTPost 200 8/15 +6 over cap