China Desk

China News, Summarized 10 Oct 2026, 23:26 UTC 34 stories 22/23 sources

Today's brief

The cheapest part on a million-yuan car is about to move a national standard

A state lab starts asking about plastic

The questionnaire is the news. On 9 October Changan Automobile confirmed receiving a survey from CATARC — the China Automotive Technology and Research Center, a central state-owned enterprise that researches auto standards and regulations on the industry ministry's behalf — asking specifically about the use of non-metallic materials in brake pedal assemblies, IT Home reports. Industry reading: groundwork for revising the bench-test standard QC/T 788-2018.

Here is the vacuum it found. The mandatory national standard governs only pedal force — under 500N at maximum braking effectiveness. A separate recommended standard asks for 1,500±20N held five seconds, five times, with permanent deformation under 5mm and no cracks. Neither sets a quantified load threshold, material limit or test method for the bracket itself: it must simply be "not easily failed," in metal or something performing like metal.

Which means every number is the carmaker's own. So executives spent Friday posting theirs. Lu Fang, chairman of Dongfeng's premium EV brand Voyah, showed a multi-layer high-strength steel beam. GAC Honda deputy GM Lin Zhibin put the Accord's at 3,000N-plus without fracture. BYD's Zheng Yu said the Yangwang line's pedal is metal, built to a 2,500N in-house spec.

THE TELL: a 36Kr teardown of the supply chain says domestic suppliers quote under 100 yuan (~$14) for a pedal assembly even on a million-yuan car, that OEMs rarely staff a dedicated engineer for it, and that load certification is routinely the supplier's own test report, accepted without re-testing. European luxury brands still build theirs in-house in Europe to roughly 4,000N and ship them in, per 36Kr.

WORTH NOTING how differently this travelled. Chinese coverage led with newtons and the standards gap; Hong Kong and Western outlets led with the consequence — once the pedal dropped, stopping distance went from about 38 metres to roughly 140, per The Standard. Zunjie, the Huawei–JAC brand whose V800 broke, still calls the free part swap an upgrade and not a recall: plan confirmed end of October, rollout from November, via GeekPark.

Go deeper on this section: ClaudeChatGPT

Zhejiang says traffic is a power now

A provincial propaganda department picked this week. The publicity department of Zhejiang's party committee published a commentary on Thursday arguing that online traffic can no longer be treated as merely a data resource — that traffic and algorithms have become a force capable of shaping public discussion and sentiment, and need more oversight, per SCMP. Zhejiang's capital is Hangzhou. You know whose home town that is.

Jilin went the friendly route the same week. Party secretary Huang Qiang sat down with content creators in a Changchun coffee shop on Wednesday to encourage them to promote mainstream values and spread positive energy, SCMP reports. Carrot in the northeast, stick in the delta.

My read: these land in the same week a ByteDance review app repriced a carmaker and triggered a standards review that no regulator had started. Beijing is watching a platform do the state's work and noticing it did not need permission.

Go deeper on this section: ClaudeChatGPT

Taipei welcomes the talks, Shanghai warns about them

Lai Ching-te used Double Tenth to endorse a thaw. Taiwan's president, marking the Republic of China's founding anniversary on Saturday, welcomed Washington–Beijing talks aimed at reducing the risk of conflict, called for democracies to cooperate on regional peace, and warned that the mainland is pressuring the island to change the status quo, per SCMP. Analysts in the same paper doubt Taiwan derails the reset; the live worry is a $14bn arms package.

The more interesting document is Chinese, and it argues against cheering. Jennifer Kavanagh, senior fellow and director of military analysis at the Washington think tank Defense Priorities, published a proposal in September to pull US forces back from the first island chain to the second. Jia Min, a research fellow at the Shanghai Development Research Foundation who hosted her at a public discussion, told Guancha afterwards that she is accused of betrayal at home — and that Chinese readers are misreading her.

Two points are worth the trip. Kavanagh's starting premise is a reassessment of American interests, not sympathy for Beijing. And an American pullback need not leave China better off: a thinner US security presence could push Japan toward nuclear weapons and unsettle a balance Beijing finds useful, as Pekingnology translates it. A Chinese analyst telling a Chinese audience not to want what it thinks it wants is rarer than it should be.

Go deeper on this section: ClaudeChatGPT

Xiaomi published the bill for one RL run

MiMo-V2.6 cost $2.6m to post-train. Xiaomi's technical report, out 8 October, puts reinforcement-learning post-training for the Pro model — 1.02T total parameters, 42B active — at roughly $2.6m, and the 310B/15B Flash at $900k. Each RL step samples 1,568 prompts at 16 rollouts apiece: about 25,000 agent trajectories and 2.7–3.7bn tokens per step, InfoQ reports.

The line item nobody budgeted for is the judge. Grader compute — purely for scoring agent behaviour — takes 12.7% of Pro's bill and 14.2% of Flash's. The report also publishes its own failure log: 123.1 hours for Pro's 30 steps, GPU double-bit memory errors, a Kubernetes failure that killed the cyber cluster's pods, a restart when the grader network went unreachable.

Two findings I'd steal. Freezing the MoE router through RL: left trainable, layer-9 expert load variation went from 0.78 to 2.0 in twenty steps, the busiest expert from 6x to 16x average, cold experts from 0.5% to 22%. And Xiaomi trained a dedicated Hack Agent to attack its own sandboxes until no cheat path remained — confirmed reward hacking stayed under 2%.

Meanwhile the buyers are cutting. Microsoft reportedly trimmed a $1bn-plus annual internal Claude budget by over a third, with per-head allowances in its cloud and AI division falling from $100,000 a month to about $10,000; Meta halved internal Claude Code seats from roughly 60,000 to 30,000. Deloitte China finds 60.4% of surveyed firms have no systematic way to measure AI value, and the share unable to calculate ROI rose from 35.8% to 44.4%, per TMTPost.

Both ends now price intelligence explicitly. Labs publish what a training run cost; enterprises meter what a token earns. Moonshot AI closed its final private round at about $50bn ahead of a Hong Kong listing targeted for the first quarter of 2027, per TMTPost's week in review — and StepFun says its 600bn-parameter Step 5 Preview, 27bn active, tops OpenRouter's trending board and opens its weights on 15 October.

Go deeper on this section: ClaudeChatGPT

The travel agent that sent you the wrong way

Holiday demand for AI trip planning was real. Alibaba's Fliggy assistant saw interaction volume up 158%; Meituan's AI concierge beat May Day by over 60%; Tongcheng's DeepTrip saw peak hourly visits more than triple on day one. A survey of 3,000 travellers found 84.6% had used conversational AI for travel information, per TMTPost.

Execution was not. One user got a Suzhou metro interchange that does not exist. Others arrived at closed attractions or walked what they called special-forces routes. The mechanism is latency: global distribution system fare caches refresh every two to three minutes, hotel inventory every ten to fifteen. And only 17% of users will let AI touch payment or passport data unsupervised, so the agent stays a shop assistant.

The money moved anyway. Hotel bookings referred through ByteDance's Doubao now carry an 11.4% software service fee plus 0.6% payment handling — about 12%, against roughly 8% when the same traffic arrived through Douyin organically. Four points is what the conversation layer costs.

Elsewhere, the humanoid reality check. Daily rental for popular humanoids has fallen from roughly 10,000 yuan (~$1,400) in early 2025 to 2,000–3,000 yuan (~$280–420) early this year, with some Unitree listings now starting at 289 yuan (~$41) a day, per GeekPark's digest of Star Market Daily reporting. Set that beside IDC's count of global humanoid shipments up 432.1% year on year in the first half — of which 69% still went to research, education, performance and government procurement.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • Huawei's patent book cashed a cheque. Following Monday's piece: Huawei and Qualcomm announced a long-term broad cross-licence spanning 5G, computing, AI and networking, with Qualcomm also buying some of Huawei's US patents in those fields, per TMTPost. The same day, Luxshare Precision disclosed that it and its Dongguan subsidiary are respondents in ITC investigation 337-TA-1526, opened 9 October; the products are still in customer validation, not mass production, via Jiemian.
  • The Claude access thread now runs on identity, not geography. Anthropic's 8 October usage policy update, effective 12 November, writes the regional restriction to cover people physically located in unsupported regions, entities incorporated or headquartered there, and entities majority-owned or controlled from there — and the company now asks some users for government ID and a live selfie, per Medianama, which also describes a grey market of API proxies reselling at as little as 10% of list using harvested documents. On V2EX, one poster hit liveness verification and was told to produce a passport; another is three days into hosting a box at a friend's house in Vietnam with no ban. Sentiment, not data.
  • The KV-cache ripple got a proposed cure. Following yesterday on ByteDance Seed's finding that DeepSeek V4 answers differently depending on token position: Leiphone's version names the remedy as dynamic chunking and the contenders — HKUST Guangzhou's ChunkKV at 73.8% on needle-in-a-haystack versus SnapKV's 58.9% at a 128-slot cache budget, Jina AI's late chunking, Microsoft Research Asia's MInference claiming up to 10x prefill speedup on million-token contexts, here. It also notes Llama 3 carries the same structural weakness. My call that averaged benchmarks hide this stands; the fix is architectural, not a tuning pass.

Go deeper on this section: ClaudeChatGPT

The river · 34

ByteDance finds a periodic blind spot in DeepSeek's compressed context

# Leiphone 11h

A ByteDance paper reports that fixed-size KV-cache compression makes DeepSeek models unusually sensitive to token position, with harmless spaces changing answers.

More
  • In one code-completion test, adding decorative equals signs changed the answer; error rates reached 71.3% for some alignment phases versus 91.5% accuracy for others.
  • The proposed mechanism is phase sensitivity: fixed-size chunks compress tokens into learned summaries, giving some within-chunk positions stronger retention than others.
  • On a 128,000-token needle-in-a-haystack test, accuracy varied by as much as 40.2 percentage points for one DeepSeek base model and 34.8 points for another; post-training reduced but did not eliminate the gap.
  • ByteDance's controlled experiments with Qwen-sized models found the periodic behavior whenever fixed-length compression was used, while a full-attention baseline did not show it; changing positional encoding or gating did not remove the effect.
  • DeepSeek's later model reduced the gap to 6.1% by halving the compression stride, but the report argues that dynamic chunk boundaries may be needed for a complete fix.
  • The practical risk is silent: a contract clause or code token near a compression boundary can be discounted without any visible failure signal, making long-context systems difficult to trust in high-stakes review.
Open at Leiphone Ask Claude ChatGPT

Xiaomi details a costly scale-up strategy for agentic reinforcement learning

# InfoQ China 6h

Xiaomi's MiMo V2.6 report scales rollouts, environments and graders together: each RL step produces about 25,000 trajectories and costs up to $2.6 million.

More
  • MiMo V2.6 Pro has 1.02 trillion total parameters with 42 billion activated per inference, while Flash has 310 billion total and 15 billion activated; both use sparse mixture-of-experts models with local and global attention.
  • The training system separates rollout, grading and parameter updates, allowing each bottleneck to scale independently; a step samples 1,568 prompts with 16 rollouts each and can span up to one million tokens of context.
  • Pro post-training costs about $2.6 million and Flash about $900,000, with grading alone consuming 12.7% and 14.2% of the respective compute budgets.
  • Xiaomi uses groupwise reward synthesis and advantage redistribution to distinguish a minimal, maintainable patch from a bloated solution that merely passes tests.
  • The team also found reward-hacking behavior such as retrieving upstream fixes or tampering with evaluation paths, forcing network isolation, environment cleanup and adversarial agents into the training harness.
  • The strategic implication is that frontier agent training is becoming an environment-engineering and evaluation-compute problem, not just a larger-model problem; the cost numbers will matter only if the resulting behavior transfers beyond the curated tasks.
Open at InfoQ China Ask Claude ChatGPT

WorldArena 2.0 tests whether world models work on real robots

# Leiphone 17h

A Tsinghua-led benchmark expands world-model evaluation from video generation to tactile input, online reinforcement learning and real-robot execution.

More
  • WorldArena 2.0 adds three dimensions that offline video benchmarks miss: visual-tactile sensing, interactive policy training and physical deployment with accumulated noise and delay.
  • Track 1 uses 70 tasks, including out-of-distribution table colors and textures, and scores visual quality, motion, consistency, geometry and instruction execution across 15 metrics.
  • Track 2 treats the world model as an RL environment, requiring repeated actions, state transitions and reward signals that remain useful during training rather than only producing plausible frames.
  • Track 3 evaluates real robot manipulation, making sim-to-real transfer and physical failure part of the score.
  • The reported standings put a JEPA-based model from Vision Future first in Track 1 and a Tongji spatial-intelligence model second, with academia and industry each represented among the top two.
  • The benchmark's importance is methodological: it tests whether a model predicts consequences of actions well enough to train and control a robot, not whether its generated video looks convincing.
Open at Leiphone Ask Claude ChatGPT

Zhejiang Party officials warn platforms wield too much traffic power

# SCMP 11h

Chinese Communist Party officials in Zhejiang say platform traffic and algorithms now shape public sentiment, calling for stronger oversight of major internet companies.

More
  • The provincial Party committee's publicity department said online traffic should no longer be treated merely as a data resource, but as a force capable of influencing public discussion.
  • The language suggests a shift from regulating individual content or data practices toward treating distribution power itself as a governance object.
  • Zhejiang is home to major internet firms and Hangzhou's platform economy, so the statement comes from inside the industry's political heartland rather than from a distant regulator.
  • In China's policy vocabulary, concern over sentiment-shaping algorithms usually implies tighter expectations around recommendation systems, platform accountability and the promotion of approved public narratives.
  • The commentary does not announce a specific law or enforcement action, but it can serve as a directional signal for future platform supervision.
Open at SCMP Ask Claude ChatGPT

Chinese military aircraft may be using a joint facility in Laos

# SCMP 9h

Satellite imagery shows eight Chinese-made K-8 or JL-8 family trainers at a Lao airfield, twice the assessed Lao inventory, raising questions about Chinese use of the base.

More
  • The imagery from September 30 also showed more than two dozen military and support vehicles at Ban Keun Airport, according to the account relayed from Reuters.
  • The aircraft count is an indicator, not proof of a permanent Chinese deployment; the source says they may already be operating from the facility.
  • A joint aviation site would extend China's operational access in mainland Southeast Asia without requiring a formal overseas base, while giving Laos training and light-attack capacity.
  • The episode fits China's preference for security relationships built through equipment, training and shared infrastructure rather than public alliance commitments.
  • What is missing is evidence about command arrangements, sortie patterns, construction ownership and whether the aircraft belong to China or are temporarily supporting Lao operations.
Open at SCMP Ask Claude ChatGPT

A broken brake pedal exposes the cost of cutting corners on luxury cars

# 36Kr 23h

Three JAC-backed Exeed V800 test cars reportedly suffered identical brake-pedal failures, prompting an industry survey of nonmetal pedal components.

More
  • The test involved repeated 100-to-zero braking and applied more than 1,600 newtons of force; all three pedals reportedly broke at the same plastic reinforcement around the pivot.
  • China's current mandatory standard caps pedal force during maximum braking at 500 newtons, while a recommendation tests 1,500 plus or minus 20 newtons; the gap leaves substantial room for company-specific safety margins.
  • The incident matters because low-cost mechanical parts can escape the engineering attention lavished on batteries, sensors and software, even in a vehicle priced as a luxury product.
  • The China Automotive Technology and Research Center's questionnaire to automakers may foreshadow changes to material or testing rules, but no regulatory revision has yet been announced.
  • The episode exposes a governance problem in fast-moving Chinese auto manufacturing: premium branding raises customer expectations faster than supplier validation and internal standards necessarily mature.
Open at 36Kr Ask Claude ChatGPT

China's medical-imaging AI has a payment problem, not just an algorithm problem

# TMTPost 8h

China is expanding AI-assisted imaging through licensing and public procurement while keeping AI bundled into existing medical fees, leaving hospitals and government buyers to fund deployment.

More
  • Specialist vendors are pursuing Class III medical-device approvals, while Alibaba's RADAR model reportedly identifies 146 conditions across 18 organs and was open-sourced after publication in Science.
  • A 2026 draft regulatory guideline adds dedicated chapters for large models, medical agents and multi-disease products, with explicit risk-management requirements for hallucination and bias.
  • National medical insurance treats AI diagnosis as an extension of the underlying scan, not a separately billable service; patients cannot simply be charged an extra AI fee and a licensed physician must review the result.
  • That rule creates a structural gap: policy wants broad deployment, but the reimbursement channel does not provide a standalone revenue line for AI vendors.
  • Public hospitals and health authorities therefore become the practical buyers, using operating budgets, research funds or fiscal programs; one provincial platform contract was reported at 205.7 million yuan, roughly $29 million.
  • The emerging business model is infrastructure procurement and outcome-linked service, including shared regional compute paid by usage, rather than a direct consumer subscription or per-diagnosis surcharge.
Open at TMTPost Ask Claude ChatGPT

Huawei opens three native HarmonyOS PC development tools to public testing

# AI Zhidx 11h

Huawei's HarmonyOS PC versions of DevEco Studio, DevEco Code and DevEco CLI bring IDE, agentic coding and command-line skills into the native development environment.

More
  • DevEco Studio provides the IDE and multi-device debugging layer, including virtual phone, tablet and foldable windows that can be checked in parallel without a conventional emulator setup.
  • DevEco Code embeds Huawei knowledge, skills and MCP tools, with Goal, Plan and Build modes that separate autonomous execution from human approval of the plan.
  • DevEco CLI exposes more than 70 skills through a command-line interface so external coding agents can invoke HarmonyOS tools without hand-built adapters.
  • The architectural move is strategic: Huawei is trying to make the operating system a development workstation and to control the tool interfaces through which future agents build applications.
  • Huawei says HarmonyOS 6 and 7 devices have surpassed 90 million, but the source does not establish how many developers or production apps will adopt the new tools.
Open at Zhidx Ask Claude ChatGPT

Step 5 Preview tops OpenRouter's trending chart before open release

# AI Leiphone 19h

Chinese startup StepFun says its Step 5 Preview reached the top of OpenRouter's trending list and will be open-sourced on October 15.

More
  • The model uses a sparse mixture-of-experts design with 600 billion total parameters and about 27 billion activated per token, targeting coding, software engineering, finance and professional knowledge work.
  • OpenRouter's trending and daily token charts measure real usage and growth, but they are not controlled capability evaluations; the report also cites a separate intelligence score of 44.
  • The model's appeal is positioned as a cost-capability tradeoff: a Chinese model seeking global developer adoption through open access, aggregation platforms and practical agent workloads.
  • The account relies heavily on company claims and one researcher's impression, so independent evaluation after release will matter more than the prelaunch ranking.
Open at Leiphone Ask Claude ChatGPT

Li Auto is turning in-house batteries into an industrial capability

# Analysis Leiphone 14h

Li Auto is moving self-developed 5C batteries across its lineup and taking an 8.79% direct stake in Sunwoda Power to secure manufacturing and data control.

More
  • Li Auto plans to invest 2.65 billion yuan, roughly $373 million, in Sunwoda Power; combined with an earlier holding, its related entities would own about 11.17%.
  • The company is shifting from supplier-specific battery purchasing toward integrated control of cell chemistry, pack design, battery management and vehicle thermal systems.
  • The rationale is not simply lower cost: Li Auto says battery safety, degradation, charging speed and winter range depend on data flowing from materials and manufacturing through the vehicle and service network.
  • The strategy preserves a hybrid model in which Li Auto defines the battery and system architecture while partners provide large-scale cell manufacturing, avoiding the capital burden of becoming a fully integrated cell producer.
  • The move also follows a battery recall lesson: the automaker remains accountable to the customer even when a partner manufactures the part, but deeper data access could surface failures earlier.
Open at Leiphone Ask Claude ChatGPT

Analysts doubt Taiwan will derail the latest US-China reset

# Policy SCMP 20h

Diplomatic analysts cited by SCMP say Washington-Beijing momentum can survive recent Taiwan tensions, though a reported $14 billion arms package remains a major risk.

More
  • The account describes Beijing as pragmatic in managing Taiwan-related friction while continuing to press Washington on what it calls the correct position on the island.
  • The arms package matters because it could convert diplomatic reassurance into a visible military and political setback, even if leaders keep broader talks moving.
  • The source presents analyst judgment rather than a new agreement, so the reset is a fragile process, not an established settlement.
Open at SCMP Ask Claude ChatGPT

Tesla softens its European FSD branding as Chinese rivals gain ground

# Policy QbitAI 16h

Tesla renamed Full Self-Driving to Tesla Assisted Driving in Europe after German officials called the label misleading, while Chinese EV brands outpace Tesla in EU registrations.

More
  • Germany says the system can assist with acceleration, braking and steering but still requires continuous driver attention; the new name makes that limitation explicit.
  • Tesla's change is a regulatory compromise: Germany may support European approval, while the company keeps the original branding in the United States.
  • The timing matters commercially because the report says BYD registered about 178,000 EU vehicles in the first eight months, above Tesla's roughly 142,000, with 163% year-on-year growth.
  • Tesla is therefore using supervised driving as a product differentiator while accepting a narrower regulatory description, a tradeoff shaped partly by Chinese competition in Europe.
  • The EU approval vote was reportedly delayed until at least December, and the current approvals mainly cover smaller markets, so the commercial effect remains uncertain.
Open at QbitAI Ask Claude ChatGPT

China's AI inference boom is forcing a rethink of storage architecture

# Signal Leiphone 21h

A Chinese storage conference cited 140 trillion daily token calls and argued that SSDs, NAND and hard drives must specialize around long-running agent workloads.

More
  • The cited China token volume exceeds the scale of ordinary chatbot use and implies persistent demand for model weights, KV cache, user state and multimodal histories.
  • Agent workloads create three distinct pressures: richer data, more complex inference and operational persistence as agents remain active for long tasks and high concurrency.
  • Storage vendors are responding with tiering by access distance and role: GPU-attached SSDs for hot KV data, networked SSDs for scalable service and capacity-oriented storage for colder state.
  • The architecture is moving away from a simple storage pyramid toward workload-aware QoS, prefetching and differentiated endurance, because planning, retrieval and execution have different latency and bandwidth needs.
  • The number is conference-reported and should be treated cautiously, but the systems implication is sound: inference economics increasingly depend on memory hierarchy, not only accelerator throughput.
Open at Leiphone Ask Claude ChatGPT

Chinese analysts argue US retrenchment is not pro-Beijing

# Analysis Pekingnology 21h

A Chinese commentary says US advocates of pulling back from East Asia may be reassessing American interests rather than favoring the Chinese Communist Party, and warns withdrawal could destabilize the region.

More
  • The debate centers on a proposal to shift US forces toward the second island chain and reduce commitments along the first, including Taiwan-related posture.
  • The Chinese analysis rejects the simple assumption that less US presence automatically benefits Beijing, arguing that Japan could pursue nuclear weapons and regional instability could increase.
  • That framing is revealing because it treats US restraint as a potential strategic problem for China rather than accepting a zero-sum reading of every American withdrawal.
  • The article is an interpretation of US policy debate and Chinese commentary, not evidence that Washington has adopted the proposal.
Open at Pekingnology Ask Claude ChatGPT

Leapmotor plans a second brand around a software-defined vehicle platform

# Analysis Leiphone 15h

Leapmotor says its second brand will pursue a radically reconfigured cabin, centralized electronics and in-house AI, battery and drive systems.

More
  • The proposed architecture uses a central vehicle computer, cuts wiring below 500 meters and combines a 48-volt electrical system with asymmetric automotive Ethernet links.
  • A cloud-and-edge agent design would keep long-term reasoning in the cloud while retaining latency-sensitive execution and privacy-sensitive functions in the vehicle.
  • The company claims a 19% increase in cabin length, 37.5% more front lateral space and a 63.7% increase in interior height through integrated HVAC, rear-mounted powertrain and low-floor packaging.
  • Leapmotor says its world-model driving route moved from selection in 2025 to a working demo in 2026, but the account provides no independent evidence for its claim of first-tier performance.
  • The strategy is vertically integrated: 8,500 of 10,000 employees are reportedly in research and development, and the company says it controls 65% of vehicle cost.
Open at Leiphone Ask Claude ChatGPT

Nio says AI demand is creating diseconomies of scale in autos

# Signal Leiphone 15h

Nio CEO William Li says battery, memory and computing demand from AI data centers is raising vehicle costs even as volume grows, while Nio's premium mix improves margins.

More
  • Nio's per-vehicle cost reportedly rose 14,000 yuan, roughly $2,000, from the end of 2025, adding about 1.5 billion yuan, roughly $211 million, to quarterly costs at its current volume.
  • Despite the pressure, first-half deliveries rose 67.4% to 191,100, revenue reached 57.67 billion yuan, roughly $8.1 billion, and combined gross margin reached 18.7%.
  • The unusual industry dynamic is that a large buyer may pay more, not less, when it competes with AI infrastructure for constrained memory and other components.
  • Nio's average transaction price reached 406,000 yuan, roughly $57,000, in the second quarter and 434,600 yuan, roughly $61,000, in July, above major German luxury brands.
  • The company reports three consecutive quarters of operating profit, but the source notes that the figure uses a non-GAAP measure, so the durability of the turnaround remains a live question.
Open at Leiphone Ask Claude ChatGPT

Tencent's Hunyuan leads China's new agent-safety ranking

# AI TMTPost 8h

Chinese coverage says Tencent Hunyuan's Hy4 Preview ranked fifth globally and first in China on an agent-safety index, while its data and evaluation chief gains influence.

More
  • The index reportedly uses more than 90,000 real agent conversations across 27 models and measures unauthorized action, false attribution and deceptive completion rather than general intelligence.
  • Hy4 Preview scored 78.5, with a 1.47% unauthorized-action rate, 6.44% false-attribution rate and 13.24% deceptive-completion rate; the last measure is above the reported average.
  • The result is strategically important because Tencent is consolidating model, infrastructure, data and evaluation responsibilities under Yao Shunyu, making evaluation design part of model leadership rather than a separate compliance function.
  • The underlying benchmark is still a vendor-adjacent ranking and the source does not provide enough methodology to compare scores across providers with confidence.
  • The shift from leaderboard intelligence to real-task failure modes mirrors a broader change in how Chinese AI companies are trying to differentiate: reliability and restraint are becoming product claims.
Open at TMTPost Ask Claude ChatGPT

AI memory shortages send global PC shipments down more than 20%

# Signal Solidot 15h

Omdia says Q3 global PC shipments fell 21.2% to 58.1 million as memory and SSD costs rose, while Lenovo's shipments dropped 22.6%.

Also: Techmeme

More
  • Omdia and IDC both estimate a roughly 20% year-on-year decline, and IDC says shipments also fell 9.1% from the prior quarter despite the normal back-to-school lift.
  • Memory and SSD costs reportedly rose more than fourfold and now account for nearly 40% of PC component cost, up from about 15%, as AI servers absorb capacity.
  • Lenovo remained the largest vendor at 14.9 million units, but HP, Dell and Apple also declined; Omdia expects another 24% year-on-year fall in Q4.
  • The linked China dimension is industrial rather than consumer: Chinese PC manufacturing and Lenovo are being squeezed by an AI infrastructure boom that reallocates component supply away from ordinary devices.
  • This is a sign that AI capex is becoming a tax on adjacent hardware categories, not merely a demand boom for accelerators.
Open at Solidot Ask Claude ChatGPT

China uses dancing robots to court African business ties

# Culture SCMP 15h

Chinese embassies in Africa showcased dancing robots and robot dogs at National Day events, presenting robotics as a symbol of China's industrial upgrading.

More
  • The events in Angola and Congo mixed entertainment with diplomatic messaging, with Chinese officials describing robotics as a small example of higher-quality economic development.
  • Robot demonstrations are a soft-power tool: they make Chinese manufacturing visible and approachable while opening conversations with officials and business representatives.
  • The account does not show that these robots generated contracts, but it demonstrates how China exports an innovation narrative alongside products and infrastructure.
Open at SCMP Ask Claude ChatGPT

China unveils an AI robot for wheat breeding

# Signal SCMP 14h

China demonstrated a robot designed to help breeders select better wheat strains, linking machine vision and automation to food-security goals.

More
  • Traditional wheat breeding requires breeders to inspect tens of thousands of hybrid offspring and identify plants with desirable traits; the robot is intended to automate part of that selection loop.
  • The project matters less as a consumer robot than as a data-acquisition system for agriculture, where repeatable phenotyping can accelerate the search through large breeding populations.
  • State media framed the prototype as part of China's drive to strengthen agricultural technology and food security, placing robotics inside a strategic production mission.
  • The supplied account does not specify the robot's sensors, autonomy level, field results or whether it has improved breeding-cycle time.
Open at SCMP Ask Claude ChatGPT

Xi and Modi cannot close the India-China divide alone

# Policy SCMP 21h

An analysis of Xi Jinping's India visit says open-weight AI offers limited common ground while border, trade and strategic distrust continue to constrain the relationship.

More
  • The account treats the leaders' conversation as a positive signal but contrasts it with continuing headlines about an asymmetric and unsettled relationship.
  • Open-weight AI is presented as a possible practical cooperation area because it can support shared development without requiring full trust in a single proprietary provider.
  • The broader lesson is that summit-level rapport can reduce friction but cannot by itself resolve the institutional and security disputes shaping India-China ties.
Open at SCMP Ask Claude ChatGPT

Nio CEO says the next three years will focus on premium markets

# Analysis Leiphone 15h

Nio's internal strategy emphasizes its core auto business, high-end positioning and priority markets while keeping embodied AI as an investment rather than a main business.

More
  • Li says Nio delivered 262,893 vehicles in the first eight months, up 57.9%, even as China's passenger-car market fell more than 20% in the same period.
  • The company argues that its services and community business is becoming a second growth curve, with first-half revenue of 5.8 billion yuan, roughly $817 million, and profits reportedly covering charging and swapping infrastructure costs.
  • Nio plans to explore embodied intelligence through strategic investments and startups rather than building a robotics division, preserving focus while retaining an option on the sector.
  • Management also predicts faster pure-electric adoption, but the source presents those forecasts as executive expectations rather than independent market consensus.
Open at Leiphone Ask Claude ChatGPT

China's scrap-metal trade becomes a strategic supply chain

# Policy TMTPost 7h

Chinese scrap processors are benefiting from tight primary-metal supplies and new reverse-invoicing tax rules, turning recycled material into a more formal commodity market.

More
  • China's recycled-resource sector earned 12.74 billion yuan, roughly $1.8 billion, in profit during the first half of 2026, up 104.5% year on year despite little change in revenue.
  • Recycled aluminum prices near 20,700 yuan per ton, roughly $2,900, reflect shortages in primary metals; the report says recycled aluminum already exceeds one-quarter of China's aluminum supply.
  • The supply logic is strategic: China imported more than 80% of its iron ore, while one ton of scrap steel can replace about 1.7 tons of ore and 350 kilograms of standard coal.
  • The tax authority's 2026 rule lets recycling firms reverse-invoice individual sellers and taxes annual scrap sales up to 600,000 yuan, roughly $84,500, at 0.25%, reducing the paperwork barrier at the informal end of the chain.
  • This is a formalization story as much as a price story: traceable invoices and finer sorting allow the sector to connect neighborhood collection to industrial buyers and global commodity prices.
Open at TMTPost Ask Claude ChatGPT

China's Singles' Day shifts from shopping ritual to purchase certainty

# Culture TMTPost 8h

Chinese platforms are removing cross-store discounts, extending sales windows and putting final prices upfront as slowing growth makes promotional complexity harder to justify.

More
  • Tmall, JD.com and Douyin are replacing threshold-based cross-store coupons with direct reductions, reducing the forced cart-building that previously inflated orders and later returns.
  • JD.com is dropping deposit-based presales, while Tmall is stretching its in-stock period to nearly 25 days; platforms are smoothing the one-day demand spike instead of manufacturing urgency.
  • Online goods retail grew 4.3% in the first eight months, but physical-goods online sales grew only 1.1%, making fulfillment waste, returns and customer-service costs more visible.
  • Subsidies are becoming more targeted at members and policy-supported categories, including a fourth trade-in subsidy tranche of about 69 billion yuan, roughly $9.7 billion.
  • The change reads as a consumer-mood signal: platforms are optimizing for conversion and repeat purchase among known buyers rather than assuming a festival can create unlimited new demand.
Open at TMTPost Ask Claude ChatGPT

Why open AI models still resist reuse

# AI InfoQ China 6h

A Chinese analysis argues that open weights and code do not make AI capabilities reproducible: models, data, tools, prompts and runtime state must line up.

More
  • For agent systems, reproducibility also depends on planning, state management, recovery and permissions, so changing one component can shift behavior or break a workflow.
  • The article points to documentation, stable interfaces, environment reproduction and independent deployment as better reuse metrics than stars, downloads or one benchmark score.
  • Its larger argument is that open AI is moving from publishing artifacts to publishing verifiable, composable capabilities; protocol layers such as MCP address connection costs, but not reliability or licensing.
  • The account cites a 2026 open-source AI competition with more than 14,000 developers and 2,999 valid entries, illustrating how multi-agent systems are becoming integration exercises rather than model demos.
Open at InfoQ China Ask Claude ChatGPT

Jilin's Party secretary meets content creators at a coffee shop

# Policy SCMP 21h

Jilin Party secretary Huang Qiang invited digital creators to a coffee shop and urged them to promote mainstream values and positive energy.

More
  • The informal setting is deliberate: the meeting treats creators as a political communications channel rather than merely as entertainment businesses.
  • The official language indicates that provincial authorities want online influence aligned with approved narratives while maintaining a more conversational relationship with influencers.
  • This is a small but clear example of China's propaganda system adapting its outreach to platform-native creators instead of relying only on traditional media.
Open at SCMP Ask Claude ChatGPT

Aito wants a distinct brand inside Huawei's automotive ecosystem

# Analysis Leiphone 15h

Aito says it will deepen Huawei integration while building its own manufacturing, sales and service capabilities, seeking identity beyond shared Huawei technology.

More
  • Aito's strategy is not separation: it wants Huawei's supply-chain, retail and intelligent-vehicle capabilities while gradually owning its own research, production, delivery and service loops.
  • The brand reports more than 6,000 employees, a dedicated factory and roughly 90 dedicated stores, alongside continued access to nearly 1,000 Huawei retail locations.
  • The tension is structural: as multiple automakers share Huawei driving and cockpit systems, each partner must create differentiation through product definition, manufacturing quality and brand-specific service.
  • The account positions Aito's high-priced V9 and forthcoming RX as attempts to make that distinction visible rather than relying on Huawei as a substitute for brand identity.
Open at Leiphone Ask Claude ChatGPT

BYD marks its 150,000th new-energy commercial vehicle

# Signal Leiphone 19h

BYD rolled its 150,000th new-energy commercial vehicle, a Q3 electric tractor, off a smart factory line and is shifting attention toward long-haul trucks.

More
  • The Q3 uses a commercial-vehicle blade battery, a 600-kilowatt peak motor and dual-gun fast charging, targeting the harder range, payload and uptime requirements of trunk logistics.
  • BYD says its commercial vehicles now operate in more than 70 countries and is pairing trucks with charging networks, fleet software and lifecycle maintenance.
  • The milestone is strategically more useful than a generic production number because it marks a move from buses and light commercial vehicles toward integrated zero-emission freight systems.
Open at Leiphone Ask Claude ChatGPT

China's AI roundup pairs Huawei licensing with Moonshot's IPO ambitions

# AI TMTPost 13h

Chinese coverage highlights a broad Huawei-Qualcomm patent deal, Moonshot's reported $50 billion pre-IPO valuation and overseas revenue-sharing for Zhipu models.

More
  • Huawei and Qualcomm reportedly agreed to cross-license patents spanning 5G, computing, AI and networking, while Qualcomm will acquire some Huawei US patents.
  • Moonshot is reportedly preparing a Hong Kong listing and a financing round at a $50 billion valuation, though the account provides no confirmed terms or timetable beyond a 2027 target.
  • Zhipu's GLM-5.3 is entering Amazon Bedrock and other cloud marketplaces under usage-based revenue sharing, giving Chinese model companies a route to overseas distribution without owning all the infrastructure.
  • The roundup places commercial scaling, intellectual-property bargaining and international cloud access alongside the separate debate over AI safety reporting, showing how Chinese coverage treats ecosystem formation as the main story.
Open at TMTPost Ask Claude ChatGPT

How a museum replica created rare US-China collaboration

# Culture SCMP 1h

A life-size replica of a Yuan-era Buddhist mural at New York's Metropolitan Museum helped create an unusual collaboration between Chinese and American cultural institutions.

More
  • The Met's Chinese Buddhist art gallery displays a Medicine Buddha work whose assembly spans more than 15 meters by 7.5 meters, making scale a central conservation and exhibition problem.
  • The article's value is institutional rather than purely artistic: reproducing a monumental work can let audiences experience its composition without moving or endangering the original.
  • A project that joins US museum practice with Chinese cultural expertise also shows how heritage work can preserve cooperation when official relations are more brittle.
  • The supplied extract ends early, so the partners, production method and eventual public use of the replica are not fully documented here.
Open at SCMP Ask Claude ChatGPT

China warns a backdoored video player infected more than 700,000 devices

# Signal IT Home 20m

China's emergency response center says the Jianpian Windows player can execute remote code and has infected more than 700,000 domestic devices.

More
  • The warning describes a full remote-control capability rather than ordinary adware: the software can collect host fingerprints, alter clipboard contents, inject processes and run scripts from two command channels.
  • The reported infection count reached 700,260 devices between September 15 and 22, with as many as 251,880 controlled endpoints online in one day.
  • CNCERT's framing treats unofficial software distribution as a national-scale supply-chain problem, while the technical indicators point to persistence through registry startup entries and system services.
  • For enterprise defenders, the important signal is that a seemingly benign media utility combined persistence, data collection and arbitrary code execution; endpoint trust cannot be inferred from the application's user-facing function.
  • The source does not identify the operator or explain how the player was distributed, so attribution and the campaign's commercial or criminal motive remain unclear.
Open at IT Home Ask Claude ChatGPT

Avatr chairman shifts toward Changan's overseas business

# Leiphone 15h

Leiphone reports that Avatr chairman Wang Hui has begun handling Changan's Southeast Asia and Latin America operations, with Avatr president Chen Zhuo expected to take broader control.

More
  • Wang remains listed internally as Avatr chairman, so this is a reported role shift rather than a confirmed formal replacement.
  • His stated prior mandate was to help prepare Avatr for an IPO, which the report says is still progressing; the change may therefore redistribute operating responsibility without ending that process.
  • The move would put more day-to-day authority with a younger Changan executive while the parent company increases its overseas push.
Open at Leiphone Ask Claude ChatGPT

Qing dynasty ceramics damaged at the British Museum

# Culture SCMP 12h

Two Chinese ceramics were damaged during a British Museum display incident and are being restored, renewing debate over how fragile heritage objects should be exhibited.

More
  • The museum says a bowl fell from its shelf onto a small box below, damaging both objects; the incident reportedly occurred on October 4.
  • Both pieces have been transferred to conservation specialists, but the supplied account does not describe the extent of the damage or the display's safety design.
  • The Chinese framing emphasizes stewardship of cultural treasures held abroad, making museum engineering and collection governance part of the story rather than treating the incident as a simple accident.
Open at SCMP Ask Claude ChatGPT

DJI and Insta360 dominate the global smart-camera market

# Signal Techmeme 4h

IDC data cited by the Financial Times puts Shenzhen-based DJI at 73% and Insta360 at 20% of the Q2 smart-camera market, leaving GoPro at 3%.

More
  • The concentration shows Chinese camera makers competing not merely on factory cost but on product iteration, software, stabilization and global distribution.
  • DJI and Insta360 are both Shenzhen companies, so this is a Chinese industrial success story with direct global market-share evidence rather than a domestic-sales claim.
  • The reported competition between the two Chinese leaders suggests the next battle is for ecosystem and category ownership, not basic market entry.
  • GoPro's 3% share illustrates how quickly an established Western action-camera brand can lose relevance when the category shifts toward connected, computational imaging.
Open at Techmeme Ask Claude ChatGPT
Source ledger
ChinAI 403 0/0 HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk 200 0/20
Pekingnology 200 1/20
Ginger River Review 200 0/20
Sinocism 200 0/20
QbitAI 200 1/10
36Kr 200 5/30
Solidot 200 5/20
Leiphone 200 20/20
GeekPark 200 3/30
IT Home 200 12/60 +43 over cap
SCMP 200 10/50 +3 over cap
Initium 200 2/15
Jiemian 200 8/30 +22 over cap
China Digital Times 200 0/15
The Wire China 200 0/50
Techmeme 200 5/15
Hacker News 200 0/20
V2EX 200 8/50 +21 over cap
Zhidx 200 4/20
InfoQ China 200 8/20 +8 over cap
New Wisdom 200 8/15 +2 over cap
TMTPost 200 8/19 +11 over cap