China Desk

China News, Summarized 21 Oct 2026, 16:08 UTC 34 stories 22/23 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

Beijing proposes new export controls covering parts of China's AI sector

# ChinaTalk 2h

A draft MOFCOM and MOST catalog would expand export controls over AI-related technologies; the proposal exposes tension between global open-model adoption and state control of strategic know-how.

More
  • The draft adds 40 technologies across AI, photovoltaics and biomedicine to China's restricted or prohibited export catalog; the article focuses on the AI provisions.
  • The catalog has historically served as leverage where China holds important production or technical positions, including rare-earth processing and battery materials.
  • Potential coverage reaches beyond finished models to areas such as chip design, memory, cybersecurity, AI-content detection and minority-language technologies, according to the article's review of the draft.
  • Because this is a consultation draft, it is not yet law; the licensing system and final scope will determine whether firms face broad restrictions or narrower case-by-case controls.
  • The central government and provincial authorities may have different incentives in administering export licenses, making implementation as important as the text of the rules.
  • For open-weight labs, the strategic question is whether Beijing views model weights and related know-how as commercial exports to monetize or sensitive capabilities to retain.
  • The proposal could change the economics of Chinese labs that have used open releases to gain overseas developers, adoption and brand recognition.
Open at ChinaTalk Ask Claude ChatGPT

Chinese apps block Doubao's phone agent, exposing a fight over platform access

# Zhidx 7h

ByteDance says 155 apps reject Doubao's automated actions, showing how Chinese platforms can block agents that threaten their control of user journeys and transactions.

More
  • The blocked apps reportedly include WeChat, Taobao, Alipay, Meituan, Kimi and DeepSeek, spanning competitors, commerce and essential services.
  • The underlying conflict is whether an agent acting for a user should be treated as the user's legitimate delegate or as an unauthorized automated client subject to platform rules.
  • Platforms cite security, account integrity and terms of service; agent makers argue that users should be able to delegate actions, while platforms risk losing ad exposure, data and transaction fees.
  • A proposed personal-agent protocol seeks shared rules, but the article says major model companies have not all joined; interoperability remains a negotiation, not a technical default.
  • China's super-app ecosystem makes the stakes especially high: an agent may need access across messaging, payments, shopping and local services to complete a task.
  • The outcome will shape permission boundaries, identity, auditability and liability for agents that act across applications.
Open at Zhidx Ask Claude ChatGPT

Hong Kong independent bookstores face searches, closures and self-censorship

# Initium 17h

After national-security police searched several independent bookstores and detained staff, owners describe a changed daily reality of operating under legal and commercial pressure.

More
  • The report names multiple stores searched or affected during 2026, with several closures or planned closures and two others facing relocation.
  • Bookshop staff describe preparing for possible police action and uncertainty over whether they or colleagues could be charged.
  • The impact extends beyond individual prosecutions: stores are rethinking inventory, events and staffing, while the Hong Kong Book Fair excluded some established independent shops.
  • The article shows how national-security enforcement can reshape cultural infrastructure through uncertainty and anticipatory behavior, even without a new formal ban on every title.
  • The owners' accounts make the human and labor cost visible: small shops often depend on a few people who cannot easily absorb legal risk or lost working time.
  • Whether independent bookstores can continue depends on access to premises, events and titles, as well as the boundaries authorities set for political content.
Open at Initium Ask Claude ChatGPT

PingCAP confirms an internal breach and source-code theft

# InfoQ China 1h

PingCAP confirmed unauthorized access and copied data, while attackers claim a much larger haul; the company says China-based production databases were not affected.

More
  • PingCAP confirmed that stolen service-account credentials were used to access some internal systems and copy data and source code; the company has not verified the attackers' claimed 10.6 TB haul.
  • The attackers claim the data includes more than 100,000 API keys, customer SQL records and hundreds of private repositories, but those figures are unverified and disputed.
  • PingCAP says some exported API-key records were encrypted and an outside test found a sample unusable; it has not ruled out exposure of other working credentials.
  • The company says the breach affected control-plane and business information, including user and organization details, SQL editor text, support records and commercial data.
  • PingCAP says it found no evidence that customer data stored inside TiDB or the database kernel was compromised; that does not erase the sensitivity of query history and control-plane metadata.
  • The incident shows why SaaS database security must include identity and cloud-control-plane access, not just the database engine and customer data plane.
  • The eventual impact depends on forensic verification of the attackers' claims and customer-by-customer assessment, which the company says is ongoing.
Open at InfoQ China Ask Claude ChatGPT

China backs real-world training sites to close the robot data gap

# Leiphone 9h

A joint government initiative is promoting embodied-AI training on actual production and service sites, aiming to create data pipelines that connect training, deployment and failure feedback.

More
  • A cited industry report estimates that broadly capable embodied models need tens of millions of hours of real interaction data, while available high-quality robot data is far smaller.
  • China had more than 70 embodied-AI training sites open by mid-2026, but many face repetitive tasks, weak data exchange and limited sustainable business models.
  • The government initiative makes real workstations the training and testing environment, with users defining tasks and deployment data feeding back into model development.
  • The article describes a Shanghai ecosystem linking data collection, model training, test facilities and production deployments, including a humanoid robot doing battery-cell handling.
  • A factory example reports nearly 80,000 operations and 98% uptime, but the excerpt does not define the measurement period or provide independent validation.
  • The infrastructure question is whether shared standards and public test facilities can reduce each company's cost of collecting data and validating robots in new settings.
  • Real-site training still has to address safety, responsibility, data formats and the gap between a successful demonstration and reliable production work.
Open at Leiphone Ask Claude ChatGPT

Chery lays out an unusually broad in-house plan for chips, models and autonomy

# GeekPark 6h

Chery is positioning itself as a technology company, unveiling automotive chips, an enterprise model and world-model-based driving plans while extending into farm machinery and motorcycles.

More
  • The announced chip roadmap spans central compute, silicon-carbide power devices and control chips; the C100 is claimed to deliver 1,200 TOPS per chip and is targeted for vehicles in 2027.
  • Chery says its CheryMind model draws on 700 TB of enterprise knowledge, 200 PB of engineering data and extensive driving and robot-action records, and is already linked to hundreds of internal systems.
  • Its driving stack pairs a world model with models for scenario generation, road understanding and decision-making; the company says it has generated 700,000 training scenarios, including high-risk cases.
  • A 100-billion-parameter model running in the car is a company target, not a demonstrated deployment; the C100 and its power, memory and thermal constraints remain to be validated.
  • Chery's adjacent bets include electric tractors, solar-assisted off-road vehicles and range-extender motorcycles, reflecting an attempt to reuse vehicle-electronics capabilities across products.
  • The strategic test is execution: chip yields, software quality, safety validation and cost will determine whether the broad in-house portfolio becomes leverage or a collection of announcements.
Open at GeekPark Ask Claude ChatGPT

A Chinese robotics startup puts its predictive world model on robots

# AI QbitAI 8h

Qianjue says its Polibrain model predicts task-relevant physical state changes rather than generating every future frame, and is being adapted across several robot forms.

Also: New Wisdom

More
  • The system represents the environment in a task-relevant latent state and predicts how actions change that state, aiming to reduce physical trial and error.
  • Its design de-emphasizes pixel-perfect future-video generation in favor of information such as object position, contact relationships and likely action outcomes.
  • A feedback loop compares predicted and observed outcomes so a robot can reassess after a failed grasp or changed scene.
  • The company says it is testing the model on quadruped, humanoid, wheeled and dual-arm robots, but the report does not provide standardized cross-body performance results.
  • The underlying research focuses on identifiable latent representations and polynomial representations intended to separate task-relevant variables from irrelevant scene details.
  • Claims of earlier theoretical work than comparable LeCun research are the company's framing; the practical test is reliable transfer across environments and robot embodiments.
Open at QbitAI Ask Claude ChatGPT

Li Auto unveils a battery, motor and charging system with lifetime fire guarantees

# Analysis QbitAI 15h

Li Auto's new in-house electric platform combines fast-charging cells, power electronics and a safety-monitoring system, backed by unusually broad battery-fire guarantees.

More
  • Li Auto says its 5C lithium-iron-phosphate cell has 2.6% higher energy density than its 2C cell, while its winter peak power rises by more than 40%.
  • The company attributes improvements to graded-pore anodes, particle-size tuning, thermal management and an adaptive battery-state algorithm that it says keeps estimation error under 3%.
  • A quality-problem battery fire that totals a vehicle would qualify for a replacement car of the same configuration; crash- or underside-damage fires are covered through a separate compensation commitment.
  • The company says its battery production system has 9,960 control points and that vehicle, charger and cloud data feed a warning platform capable of alerting on some risks up to 72 hours ahead.
  • These are manufacturer claims and warranty terms; real-world frequency, exclusions and claims handling will determine whether the guarantee changes consumer confidence.
  • The launch shows an automaker attempting to own more of the cell, power-electronics and charging stack, but does not by itself establish that vertical integration beats specialist suppliers.
Open at QbitAI Ask Claude ChatGPT

Taiwan expels a Chinese exchange student over a TikTok post

# Culture Initium 7h

Taiwan revoked a mainland student's entry permit and imposed a five-year re-entry ban over a politically charged video, prompting debate about proportionality and speech.

More
  • The student had posted a staged image of President Lai Ching-te with pro-unification hashtags; he later removed it and apologized, saying he did not intend physical harm.
  • Taiwan's immigration authorities said the post included language advocating the elimination of Taiwan's sovereignty and ordered his departure.
  • The case turns on how immigration law applies to political expression by a visiting student, and whether the maximum five-year restriction was proportionate.
  • The episode shows how online nationalist content can quickly become a cross-strait legal and diplomatic flashpoint, even when the post is framed as a joke or performance.
Open at Initium Ask Claude ChatGPT

A CATL factory visit shows why battery-making is more than cell chemistry

# Analysis QbitAI 11h

A visit to CATL's facilities illustrates how production consistency, process controls and pack integration shape the gap between a battery design and a reliable vehicle product.

More
  • The report says CATL employs more than 50,000 people in Ningde and over 230,000 globally, and has held the top position in global installed battery volume for several years.
  • The factory describes more than 7,000 quality-control points and AI visual inspection, targeting defect levels measured in parts per billion; these are company claims.
  • Battery packs require tightly controlled cell capacity and internal resistance because variation can limit usable pack energy and worsen with age.
  • The report describes sodium-ion work for cold-weather use and multiple cell chemistries at different stages, from mass production to validation.
  • Automakers can design their own packs, but qualification, production yield and cell-to-pack integration create substantial engineering work for a new entrant.
  • The visit pushes back against the idea that automakers can quickly replace CATL: building cells at scale and integrating them into safe vehicles are distinct capabilities.
Open at QbitAI Ask Claude ChatGPT

A Tsinghua-affiliated startup tests a multi-agent system for hospital care

# AI QbitAI 4h

Zijing AI Hospital is testing specialist AI agents that can contribute to consultations and cross-department case discussions while leaving clinical decisions to doctors.

More
  • The system can surface a specialist agent when a case contains relevant information outside the attending doctor's specialty, such as asthma history affecting an endocrinology consultation.
  • It offers both non-disruptive alerts during a visit and a more explicit multidisciplinary review mode for complex cases.
  • Rather than hiding disagreement, the system is designed to show conflicting agent opinions so clinicians can reassess inputs and make the final call.
  • The company says agents can learn from doctors' practice, but the excerpt provides no independent clinical validation, safety study or outcomes data.
  • The approach treats orchestration and workflow integration as central: a hospital system must fit existing clinical processes, not merely produce plausible medical text.
Open at QbitAI Ask Claude ChatGPT

Chery wants to sell Chinese families a recreational motorcycle

# Analysis GeekPark 1h

Chery is entering premium leisure motorcycles with hybrid and electric models, betting that Chinese demand is shifting from commuter bikes toward recreation.

More
  • Chery's Zero Dimension unit says China's motorcycle market is export-heavy and low-margin, while domestic sales of larger recreational bikes have been growing faster than commuter models.
  • Its first concepts include an electric adventure scooter with liquid-cooled fast charging and a range-extender cruiser; the company says its four-mode hybrid platform spans electric, range-extender, hybrid and direct-drive operation.
  • The products borrow automotive battery, thermal-management and electronic-architecture expertise, but the company's executive says adapting that technology to actual riders has been a major challenge.
  • The report describes a team deliberately combining car-industry engineers, motorcycle veterans and software/AI hires, while requiring early team members to ride motorcycles.
  • Chery's pitch is larger than vehicle sales: it wants to build a recreational riding ecosystem around accessories, modification and touring, a more ambitious claim than its disclosed product plans yet demonstrate.
Open at GeekPark Ask Claude ChatGPT

Tencent opens its Yuanbao search index as an API for agents

# AI Zhidx 3h

Tencent is offering its AI search capability through APIs, MCP and skills, turning internal search infrastructure into a service for developers building agents.

More
  • The report says Yuanbao search has a web index at the scale of hundreds of billions of entries and has been used inside Tencent products such as WorkBuddy and ima.
  • Its API returns conventional web results and structured vertical data such as weather and share prices, which can spare agents a separate scrape-and-parse step.
  • The article's own tests show an agent using search for policy research and academic-paper discovery, but these are demonstrations rather than independent accuracy benchmarks.
  • Opening the service lets Tencent monetize and extend infrastructure developed for its own AI applications, while shifting source quality and uptime dependence to Tencent.
  • For Chinese developers, the offering is another route to current web information without building and maintaining their own crawling, cleaning and indexing pipeline.
Open at Zhidx Ask Claude ChatGPT

China's offshore-trust tax deadline tightens scrutiny of overseas wealth

# Policy SCMP 3h

The end of a disclosure and settlement grace period signals Beijing's intent to bring more offshore trust liabilities and capital flows within reach of tax authorities.

More
  • The report places the deadline in a wider tightening of scrutiny around offshore assets, cryptocurrency, stablecoins and real-world asset structures.
  • For wealthy families, the practical issue is not only tax payment but the reduced ability to keep international holdings outside government visibility.
  • The article warns that tighter controls could encourage some users toward digital-dollar instruments, an argument about possible behavior rather than a demonstrated shift.
  • The longer-term signal is an expanding state effort to connect cross-border wealth, tax reporting and capital-control enforcement.
Open at SCMP Ask Claude ChatGPT

Shanghai pilot gives AI agents digital identities inside organizations

# Policy Leiphone 9h

A Shanghai platform is issuing identity codes for agents, part of a wider effort to define who an agent represents, what it can access and who remains accountable.

More
  • The initiative treats each agent as a separately identifiable actor, enabling organizations to manage its identity, capabilities and cross-system access.
  • China's proposed agent guidance calls for registration and discovery services and for differentiated rules around tool permissions and autonomous decision-making.
  • The article describes companies assigning agents bounded roles, with human review for planning and final business decisions.
  • Identity alone does not establish safe delegation: permissions, logs, approval points and a named responsible person are needed across a chain of agents.
  • As agents move from personal tools into business workflows, the governance problem shifts from model output quality to delegation, auditability and liability.
  • The Shanghai platform is a local implementation effort, while the referenced standards and policies indicate a wider national push for interoperable agent governance.
Open at Leiphone Ask Claude ChatGPT

Taiwan's test-flown attack drone is designed to saturate PLA defenses

# Analysis SCMP 2h

Analysts say a Taiwan-made, Shahed-style drone could overwhelm Chinese air defenses with low-cost numbers, though its wartime use remains uncertain.

More
  • The report says the drone has been flight-tested and that its maker claims it could reach military targets on the mainland; those capabilities are not independently established here.
  • The concept draws on Ukraine and Middle East combat lessons: inexpensive one-way systems can force defenders to expend more costly interceptors and manage many simultaneous targets.
  • Analysts differ on whether such a drone would primarily target PLA forces or play another role in a conflict, so the operational concept is unsettled.
  • The significance depends on production scale, survivability, navigation and the PLA's ability to detect and defeat massed low-cost systems, none of which is quantified in the brief.
Open at SCMP Ask Claude ChatGPT

China's daily AI token use exceeds 500 trillion, government says

# Signal Leiphone 9h

A ministry official said daily token calls in China topped 500 trillion by June, as industry debate shifts from chip counts toward the cost and quality of useful inference.

More
  • Token counts are becoming a billing and demand metric, but the article stresses that tokenization, caching, modalities and output quality differ across models, so raw counts are not directly comparable.
  • DeepSeek has experimented with time-of-day pricing; the cited cached-input rate is 0.02 yuan per million tokens (about $0.003), specific to a model, time window and cache condition.
  • The article argues that operators should track throughput, first-token latency, output speed, availability, energy and task quality alongside token volume.
  • It cites a chip-model collaboration that reportedly reduced some model adaptation work from one or two months to days, showing how software readiness affects usable compute supply.
  • The central business shift is from counting installed accelerators to measuring how reliably a system turns compute into useful, paid inference.
  • For robotics, tokens are only one part of task cost; control latency, power, success rate and hardware amortization also matter.
Open at Leiphone Ask Claude ChatGPT

ByteDance, Alibaba and Tencent reportedly order more Nvidia workstation GPUs than can ship

# Signal Solidot 1d

Chinese internet firms are said to have ordered millions of RTX PRO 5500 cards, far beyond Nvidia's projected supply, underscoring demand for domestic AI compute.

More
  • The report says the three companies ordered more than 3 million cards worth over $35 billion, while total Chinese orders exceeded 6 million.
  • Nvidia reportedly warned partners that customers could receive only 5% to 10% of requested quantities, with global production capped around 3 million cards in 2027.
  • The orders show Chinese firms seeking workstation-class GPUs for AI workloads, but the report does not establish how many units are approved for delivery or when.
  • If the supply estimates hold, the gap could intensify competition for compute and encourage further software work on alternative accelerators.
  • The numbers are attributed to a report about company orders, not confirmed purchase contracts; export approvals and final allocations remain decisive.
Open at Solidot Ask Claude ChatGPT

Alibaba Cloud holds 45% of China's automotive public-cloud market

# Signal Leiphone 6h

IDC estimates Alibaba Cloud captured 45.1% of China's automotive public-cloud infrastructure market in the first half, with revenue up 111% year over year.

More
  • The market totaled about 8.34 billion yuan (roughly $1.2 billion) in the first half, according to the cited IDC tracking report.
  • The result indicates a concentrated cloud supplier position in Chinese automotive infrastructure, as connected-car and AI workloads expand.
  • The accompanying claim that Alibaba's models and in-house chips underpin its lead is company positioning; the report provides no breakdown of customer or workload share.
Open at Leiphone Ask Claude ChatGPT

Beijing's Taiwan affairs chief meets a US policy delegation

# Policy SCMP 5h

Song Tao met visiting scholars from the National Committee on American Foreign Policy and reiterated Beijing's stated goal of peaceful cross-strait development and its political red lines.

More
  • The meeting is notable because the report describes it as a rare engagement between the Taiwan Affairs Office director and a US think-tank delegation.
  • Song placed the discussion in the context of recent China-US leader-level meetings and emphasized Beijing's position on Taiwan.
  • The format gives Beijing a channel to communicate its preferred framing directly to US policy interlocutors, without changing the underlying disagreement.
  • The brief does not report any concrete policy agreement or shift in cross-strait conditions.
Open at SCMP Ask Claude ChatGPT

Starbucks China will require a purchase to use seats

# Culture TMTPost 6h

Starbucks China is ending general free seating, a visible retreat from the chain's third-place identity as local coffee rivals and tighter unit economics reshape the business.

More
  • The policy takes effect nationwide from October 22, with seating generally tied to a purchase, according to the report.
  • The article traces the policy's history to Starbucks' response to a 2018 US racial-discrimination incident, then notes that the company later reversed free access in North America.
  • In China, the move comes amid intense price competition from local chains and pressure on Starbucks' premium positioning and store costs.
  • The report links the change to Boyu Capital's majority stake in Starbucks China, but notes that Starbucks had already restricted free seating in the US before the deal.
  • The shift puts a commercial price on the third-place promise: customers may read it as sensible cost control or as evidence that the brand's hospitality is conditional.
Open at TMTPost Ask Claude ChatGPT

Tencent and Ant back a personal-assistant startup led by a former ByteDance executive

# AI Zhidx 12h

Today AI says it has raised a cumulative $200 million for a cross-device assistant that remembers user context and can connect to work apps.

More
  • The startup was founded in March 2026 by Qi Junyuan, formerly a Feishu product executive and head of Doubao's PC product.
  • Its assistant runs across mobile and desktop platforms and can connect to services including WeChat, Tencent Meeting, Baidu Netdisk and Kingsoft Docs.
  • The product pitch is persistent memory and proactive help across work and personal tasks, rather than one-off chat; the report does not provide retention or task-completion results.
  • The $200 million is a reported cumulative financing total, with no valuation or current-round details in the supplied article.
  • The startup's premise creates a difficult trust tradeoff: usefulness depends on broad context access, while that same access raises privacy and platform-permission concerns.
Open at Zhidx Ask Claude ChatGPT

DeepSeek's inference code aims to make Ascend a first-class target

# AI TMTPost 5h

A DeepSeek inference package reportedly targets Huawei Ascend and Nvidia GPUs from one codebase, potentially lowering the cost of switching Chinese inference workloads between chip suppliers.

Also: GeekPark

More
  • The report says the project is designed so a workload can target both hardware families without maintaining separate forks, reducing software lock-in at the serving layer.
  • Inference is a more direct cost-per-token and latency contest than training, so portability could help operators route around chip availability and policy constraints.
  • The report cautions that public benchmark evidence is thinner for Ascend than for Nvidia; compiling on both platforms does not establish comparable performance under production load.
  • The package is described as an offline-inference release, leaving control-plane needs such as caching, load balancing and multi-node operation unresolved.
  • The strategic value depends on whether dual-target support survives real traffic and narrows the engineering burden of deploying domestic accelerators.
Open at TMTPost Ask Claude ChatGPT

MiniMax's M3.1 Flash targets complex agent tasks at lower cost

# AI Leiphone 11h

MiniMax says its new Flash model leads selected agent benchmarks and completed one hands-on task faster than a DeepSeek alternative at similar API cost.

More
  • The company reports a score of 43 on Artificial Analysis, up from 29 for its previous M3 model, while weighted average task cost fell from $0.51 to $0.22.
  • MiniMax's model is reported first on Automation-Bench and tied with Opus 5 on BankerToolBench, while DeepSeek remains ahead on the cited software-engineering benchmark.
  • In the article's test, MiniMax Code with M3.1 Flash took 55 minutes to build an interactive 3D campus; OpenCode with DeepSeek took about 121 minutes, with each costing roughly 6.5 yuan (about $0.90).
  • The benchmark claims and single task test are not independently replicated, and comparisons depend on harness, prompts and scoring.
  • The release shows China's low-cost model competition moving beyond token prices toward reliable multi-step tool use.
Open at Leiphone Ask Claude ChatGPT

China's aging debate puts human care alongside AI and smart-home monitoring

# Culture 36Kr 11h

A Beijing forum on AI and aging argues that monitoring and assistive devices can help, but cannot replace human contact, while older adults also face a digital divide.

More
  • Participants described older people learning AI tools and using smart devices for home safety, while emphasizing that technology cannot substitute for trust and care relationships.
  • The article cites survey and community-program material from a nonprofit, including technology classes and home-safety interventions, but does not evaluate their impact independently.
  • Speakers argue that older adults want not only to receive services but also to contribute and be recognized, challenging a purely service-recipient view of aging.
  • China's demographic pressure makes care automation attractive, but the central social question is how to support caregivers without reducing older people to monitored endpoints.
  • The story connects the digital divide to family time and patience, not simply device affordability.
Open at 36Kr Ask Claude ChatGPT

Wang Yi and Djibouti sign security cooperation agreements

# Policy Jiemian 39m

China's foreign minister met Djibouti's foreign minister and signed agreements on security cooperation and maritime stability around the Horn of Africa and Red Sea.

More
  • Wang urged Djibouti to use China's zero-tariff policy and its geographic position to deepen economic cooperation and attract Chinese investment.
  • The meeting linked bilateral ties to the strategic importance of the Red Sea and Gulf of Aden, where shipping security is under pressure.
  • The two sides issued a joint statement on regional peace and shipping safety and signed a memorandum on cooperation under China's Global Security Initiative.
  • The readout also sought Djibouti's continued support for Beijing's positions on issues it calls core interests, including the one-China principle.
Open at Jiemian Ask Claude ChatGPT

China policy watch: export controls, the Fifth Plenum and pressure on Japan

# Analysis Sinocism 7h

A China-focused newsletter surveys proposed export controls, economic signals and the approaching Fifth Plenum, alongside rare-earth tensions with Japan.

More
  • The roundup flags a draft package of controls touching AI, solar technologies and machinery, with a separate analysis of Japan's access to rare earths.
  • It also notes debate over China's third-quarter growth figures, market weakness and what the Fifth Plenum may prioritize.
  • The newsletter includes diplomatic and foreign-policy items, but the excerpt does not provide enough detail to assess each independently.
  • Its value is as an agenda map of what China-watchers were tracking, not original reporting on every listed development.
Open at Sinocism Ask Claude ChatGPT

European tech sovereignty is about viable alternatives, not isolation

# Analysis InfoQ China 8m

A European open-source panel warns that sovereignty requires real switching skills and options; a Chinese comparison highlights the hardware dependencies software openness cannot solve.

More
  • Panelists framed open source as preserving choice, not requiring governments to abandon global cloud providers outright.
  • The discussion distinguishes access to source code from the ability to operate, secure and migrate a system in practice.
  • A speaker contrasted Europe's focus on software choice with China's attention to access to chips and optical systems, pointing to a harder supply-chain dependency.
  • Kubernetes illustrates portability across self-hosted and commercial infrastructure, but moving between proprietary platforms can still demand substantial expertise.
  • The discussion's practical test is whether an alternative can actually be used and supported, not whether it exists on paper.
Open at InfoQ China Ask Claude ChatGPT

China's premier visits the SCO; Beijing watches the Fifth Plenum and Japan

# Analysis Sinocism 17h

A China-focused newsletter flags Li Qiang's regional diplomacy, a new national data group, Fifth Plenum speculation and rare-earth pressure on Japan.

More
  • The newsletter says Li Qiang attended the Shanghai Cooperation Organisation heads-of-government meeting in Dushanbe before traveling to Russia for a regular premier-level meeting.
  • It lists the inauguration of a National Data Group and the upcoming Fifth Plenum as domestic issues to watch, but the excerpt does not explain the new group's mandate.
  • The roundup also tracks pressure on Japan around rare-earth access, placing industrial supply chains alongside diplomatic relations.
  • This is a curated watchlist with links and commentary, not a single reported event; its value is the mix of political and industrial signals it surfaces.
Open at Sinocism Ask Claude ChatGPT

Motubrain opens a live window into embodied-AI training

# AI Leiphone 6h

Chinese startup Shengshu says it will livestream the training process for its Motubrain 2.0 embodied model, exposing experiments before its planned open-source release.

Also: QbitAI, New Wisdom

More
  • The planned stream covers pretraining, task-specific post-training, agent execution and experiments in recursive self-improvement.
  • The team says it wants viewers to see training, evaluation and iteration rather than only polished robot demonstrations or final benchmark results.
  • For embodied AI, the open process could reveal how teams test transfer across tasks and robot bodies, though the report provides no technical results yet.
  • The work is still exploratory; the livestream itself is not evidence that the model generalizes reliably in real-world deployment.
Open at Leiphone Ask Claude ChatGPT

Chinese analysts urge Europe to treat the Arctic as a place to cooperate

# Analysis SCMP 3h

After the EU warned of China's growing Arctic alignment with Russia, Chinese analysts argued that Europe should work with Beijing on shipping rather than frame China as a security threat.

More
  • The EU's Arctic policy identifies Russia as the primary threat and raises concern over China's strategic ambitions and closer ties with Moscow.
  • The Chinese analysts' counter-frame emphasizes economic and shipping cooperation, rather than the EU's security framing.
  • The disagreement reflects competing views of Arctic access: commercial transit and development on one side, strategic presence and Russia ties on the other.
  • The excerpt supplies no concrete EU-China project or policy response, so the argument's practical effect remains unclear.
Open at SCMP Ask Claude ChatGPT

A Feishu-Doubao contest tests students on real company work

# Culture QbitAI 13h

A Beijing final put student teams on business problems using Feishu and Doubao, reflecting a shift in Chinese employers toward demonstrated AI-enabled work rather than tool familiarity.

Also: New Wisdom

More
  • The organizers say 12 teams tackled company-defined tasks, and a report based on more than 4,300 teams and 108 firms was released alongside the event.
  • Examples included production scheduling, radar-plus-AI incident workflows and a short-drama production pipeline; one team claimed it could produce a 120-minute drama in three days for 3,000 yuan (about $420).
  • The competition included company interviews and job offers, turning a contest into a recruiting channel and a demonstration of practical skills.
  • The event is sponsored by the product vendors, so claims about productivity and the representativeness of the selected projects should be treated cautiously.
  • The hiring signal is that firms increasingly value problem definition, business context and verification of outputs alongside fluency with AI tools.
Open at QbitAI Ask Claude ChatGPT

China's delivery-robot race is shifting from replacing riders to moving freight

# Analysis TMTPost 6h

Neolix raised nearly $400 million as Meituan rethinks autonomous delivery; the strongest near-term use case may be standardized trunk routes, not doorstep delivery.

More
  • The article says Neolix's new round was nearly $400 million (about 2.8 billion yuan), with domestic and strategic investors participating.
  • Meituan reportedly delayed promotion reviews so they could reflect delivery of a new vehicle project, a sign that business results are becoming more important than technical milestones.
  • The core operational mismatch is last-mile complexity: buildings, gates, elevators and customer coordination are work human couriers handle flexibly, while vehicles may require handoffs that erase labor savings.
  • The article describes Meituan narrowing toward depot-to-area and business-to-business trunk transport, where routes and handoffs are more standardized.
  • Whether the sector is a durable business depends on utilization and labor avoided per route, not fundraising or autonomous-driving capability alone.
Open at TMTPost Ask Claude ChatGPT

A Chinese startup positions its AI learning device around making, not answers

# Culture Leiphone 13h

CapaLeap is pitching an agent for children that prompts exploration and supports creation rather than completing homework, amid concerns about AI dependence and reduced independent thinking.

More
  • The company distinguishes itself from China's concentrated learning-tablet market by targeting curiosity, creative work and guided questioning rather than test scores.
  • Its proposed architecture combines a shared agent platform, long-term memory across age-specific devices and a community for sharing and remixing projects.
  • The article cites surveys suggesting widespread student use of AI for homework and parental concern about dependence, but those figures do not prove the product's alternative will improve learning.
  • The design challenge is to preserve child agency without turning a companion agent into another source of dependency or surveillance.
  • The product's educational claims remain a positioning strategy; no comparative learning outcomes are provided.
Open at Leiphone 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 0/20
Ginger River Review 200 0/20
Sinocism 200 2/20
QbitAI 200 7/10
36Kr 200 5/30
Solidot 200 11/20
Leiphone 200 13/20
GeekPark 200 4/30
IT Home 200 12/60 +35 over cap
SCMP 200 10/50 +8 over cap
Initium 200 6/15
Jiemian 200 8/30 +13 over cap
China Digital Times 200 0/15
The Wire China 200 0/50
Techmeme 200 0/15
Hacker News 200 1/20
V2EX 200 8/50 +38 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