China Desk

China News, Summarized 22 Sep 2026 34 stories archived day

Today's brief

Anthropic wrote the report, and China's own regulator turned it into a case

Beijing's regulator is reading Anthropic's report

The Cyberspace Administration is now investigating DeepSeek and Moonshot. The Information reports the internet regulator has questioned staff at both labs over alleged leaks of sensitive user data to Anthropic's Claude, per Techmeme. Two weeks ago this was an American accusation against Chinese companies. It is now a Chinese enforcement matter against Chinese companies, three days before Xi lands in Washington with an AI dialogue on the agenda.

The underlying claim is a routing claim, not a training claim. Anthropic's 10 September report said Moonshot — maker of the Kimi assistant — silently forwarded customer requests to Claude and returned the answers as Kimi's, as SCMP covered it. It attributed 23m exchanges to Moonshot over three months and 12.1m to DeepSeek in fourteen July days, including surveillance footage from Chengdu cameras uploaded by a user it assessed as PLA-linked, per Vision Times.

Same week, same two labs, different room. Reuters reports DeepSeek and Moonshot are among the companies briefing the UN Security Council on AI risk during the General Assembly, via Techmeme. Representing the Chinese frontier at the UN while under investigation at home is not a contradiction Beijing appears troubled by.

And the unplugging is happening from the other end too. Xiaohongshu cut internal access to Claude's API last week — Qwen, GLM and DeepSeek still work fine, employees told Leiphone's morning brief. At Sunday's investor meeting, phones confiscated and notes taken on paper, Liang Wenfeng — the former hedge-fund manager who founded DeepSeek — said training on Huawei silicon "must succeed," with training chips expected in the fourth quarter. A 2 trillion-parameter model is in training against V4's 1.4 trillion; 8 trillion is planned.

Go deeper on this section: ClaudeChatGPT

A plenum date, and Thursday in Washington

The Fifth Plenum will sit 26–29 October. Monday's Politburo set the date and named one agenda item: a Central Committee decision on persistently advancing full and rigorous Party self-governance, with Xi Thought on Party Building at its centre, per Sinocism. That plenum will also ratify yesterday's expulsions of generals Zhang Youxia and Liu Zhenli. A discipline document is the whole plenum, not a section of it.

The summit framing is already fixed. People's Daily called for building on the "constructive strategic stability" agreed in Beijing, and the red lines were spelled out days early, per SCMP. Meanwhile the New York Times reported Taiwanese officials were told by Trump administration officials to curtail visits to Washington think tanks — which Taiwan's opposition is using to ask how rock-solid the relationship is, here.

Quietly, a five-year sentence ended. Sophia Huang Xueqin, the journalist whose reporting on sexual abuse helped galvanise China's #MeToo movement, was due out of prison on Friday, per China Digital Times. She carries a supplementary sentence of four years' deprivation of political rights. Sixty organisations asked Beijing to let her actually be free; the pattern to watch is the one Macau's closed-door trial rehearsed last week.

Go deeper on this section: ClaudeChatGPT

Xiaomi put a price on reinforcement learning

MiMo-V2.6 shipped, and the bill was public the whole time. Luo Fuli — head of Xiaomi's MiMo team, previously a core DeepSeek researcher on R1 — livestreamed the run on a dashboard with a dollar counter. Thirty RL steps each for Pro (1.02T total, 42B active) and Flash, about 750,000 trajectories, final total $3.5m (23.44m yuan), per QbitAI.

The out-of-sample number is the one to read. On DeepSWE v1.1, a held-out long-horizon software benchmark, Pro went 58.4 to 72.57 and Flash 48.7 to 65.68 — Flash gaining more for a third of the money. Pro scores 46.32 on the Artificial Analysis index, top of open weights, at $0.13 a task against $3.74 for Grok 4.7's xHigh tier. Also open-sourced: 7,000+ RL task environments and the training framework.

Alibaba answered with scale rather than efficiency. At Yunqi, Qwen3.8-Max ran 33 effective self-improvement rounds over a month with no human in the loop, lifting its index score 40 to 45; Qwen4 is training on a new architecture, with Qwen4.5 and Qwen5 aimed at 5–10 trillion parameters, per QbitAI. The Zhenwu V900 accelerator is claimed at 3x the M890, scaling to 500,000 cards, mass production first quarter 2027. Chief executive Wu Yongming put the 2032 target at 20GW of datacentre capacity.

Then the sober number of the day, from inside Alibaba. Cainiao, its logistics arm, drove AI's share of committed code from about 10% to over 90% in seven months — and requirement-delivery cycles got only 10% faster, R&D director Guo Fengzhao told AICon, via InfoQ China. Coding is roughly 30% of the cycle; the rest still waits on a human to confirm a stage and open the next tool.

Go deeper on this section: ClaudeChatGPT

Selling scrap became an extraction shooter

Young people in Chinese cities have discovered scrap arbitrage. ATRenew has over 60,000 smart recycling machines across 40 cities; Beijing alone shows 9,493 devices on the map, Shenzhen 5,400, Jinan 4,000. Everything pays a flat 0.6 yuan a kilo — about four cents a pound — and the writer's first haul of two boxes and two buckets earned 0.32 yuan, per a Hedgehog Commune piece in TMTPost.

The framing is what makes it interesting. The piece reads the loop as a loot-and-extract shooter: scan your flat for materials, pack them densely, avoid the neighbourhood's elderly scavengers whose bins you are now emptying, reach the machine, cash out. ATRenew staff told the reporter the machines are not very profitable — revenue is the resale spread plus government subsidies on recycling equipment. Sentiment, not a business model.

Separately, Brandy Melville took away the fitting rooms. The brand's Beijing and Chengdu stores have closed them and stopped cash refunds — exchange or equal-value voucher only, within 30 days — with staff saying head office notified them suddenly and it may be permanent, per Jiemian. It sells one size, roughly XS–S. The homemade height-weight chart that circulated in 2020 paired 5'5" with 104 pounds; every entry sits below a BMI of 18.5.

Go deeper on this section: ClaudeChatGPT

Two Chinese outlets, one Xiaomi model

QbitAI wrote the money and the science. Its piece leads on the burn rate, Luo Fuli's claim that the engineering exceeded what she worked on at DeepSeek-R1, and a materials-science demo where MiMo designed metal-organic frameworks for capturing forever chemicals — a Peking University researcher quoted calling the output equivalent to a trained postdoc.

Zhidx opened a terminal instead. Same outlet family, same day: it ran the models in OpenCode and reported Flash building a Xiaomi EV website with correct specs and photographs of Mercedes, BMW and Audi, then Pro spending 64 minutes on a racing game with an unclear track, here. Western reaction, meanwhile, skipped the leaderboard — the Hacker News thread's top comment praised the disclosure, not the score.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • Zhipu did both things, one day after I said it had done neither. Following Thursday through Monday on ZCode uploading customer workspaces: Zhipu has open-sourced the client and brought in outside review, with the China Academy of Information and Communications Technology and NSFOCUS confirming the Alibaba Cloud storage bucket and all its objects were deleted and that v3.14.0 removed the snapshot and upload paths, via 36Kr. Monday I called the zero-retention checkbox neither open-sourcing nor audit. Both landed within 24 hours.
  • And Xiaomi's coding CLI has the same shape of problem. A V2EX user's static analysis of the official MiMo CLI v0.1.14 Windows build — code not present in the public repository, injected at build time — says it posts repository URL, commit, branch, install UUID and machine specs to a Xiaomi tracking endpoint by default, with a collectCodebase routine capable of bundling up to 2,000 tracked files that is present but never called, here. One environment variable turns the reporting off. Claim, not finding.
  • Tesla is auditing the robot supply chain. Following Friday and Sunday on China's robotics money: the humanoid team spent last week inspecting Tuopu, Sanhua and Joyson, all of which already hold orders for joint modules, actuators and precision parts, and all of which are existing Tesla car suppliers, per Leiphone's brief.
  • The R&D league table came out. Huawei and Tencent rank first and second among private firms, on 192.3bn yuan (~$27bn) at 21.8% of revenue and 85.75bn (~$12bn) respectively; the top 500 booked 44.93 trillion yuan (~$6.3tn) of revenue, entry threshold 25.6bn (~$3.6bn), via Leiphone.

Go deeper on this section: ClaudeChatGPT

The river · 34

Xiaomi spends $3.5 million to train an open model in public

# QbitAI

Xiaomi says its MiMo-V2.6 reinforcement-learning run cost $3.5M, produced a leading open model and released the weights, environments and training stack.

Also: Zhidx

More
  • The Pro model has 1.02 trillion total parameters and 42 billion active parameters; Xiaomi reports a 46-point score on Artificial Analysis and a top open-model position.
  • The six-day run used 30 steps, 1,568 prompts per step and 16 trajectories per task, generating more than 750,000 trajectories across coding and other agent environments.
  • DeepSWE performance rose from 58.4 to 72.57 for Pro and from 48.7 to 65.68 for Flash, suggesting transfer to unseen software-engineering tasks rather than only training-set gains.
  • Xiaomi released more than 7,000 reinforcement-learning environments, model weights, a training framework and mini-harnesses, making the release useful to researchers beyond the model itself.
  • The reported API price is about 3 yuan per million input tokens and 6 yuan per million output tokens for Pro, far below leading closed models at similar benchmark scores.
  • The run’s public cost dashboard is part of the product strategy: it turns training expense into evidence of engineering scale and an open-source narrative.
  • The important test is whether the released infrastructure lets outside teams reproduce the gains without Xiaomi’s cluster, data curation and evaluation pipeline.
Open at QbitAI Ask Claude ChatGPT

Alibaba says Qwen is heading toward 10 trillion parameters

# QbitAI

Alibaba says Qwen4 is training and later Qwen models may reach 5 trillion to 10 trillion parameters, while recursive self-improvement is already being used in training, inference and chip design.

More
  • Alibaba reports that Qwen3.8-Max autonomously built training workflows, generated data and designed experiments for 33 effective iterations without human intervention.
  • The company also says the model improved inference throughput 96% on an unfamiliar domestic GPU and reduced a chip implementation’s area 42% through automated design.
  • Qwen3.8-Flash reportedly cuts training cost by nearly 90% through architectural changes and sparse activation, showing that scaling is being paired with efficiency work.
  • The announcement spans language, audio, video, image and world models, reflecting an attempt to build a unified multimodal family rather than a single flagship.
  • Alibaba reports more than 3 billion Qwen downloads and over 300,000 derivative models, making distribution and developer adoption part of the strategy.
  • Parameter targets are forward-looking corporate plans, not delivered capability; communication, data and serving costs may become the limiting factors.
  • The broader signal is that China’s leading cloud companies are treating model, chip and infrastructure co-design as one industrial program.
Open at QbitAI Ask Claude ChatGPT

Alibaba commits to models, chips and cloud as its AI foundation

# QbitAI

Alibaba CEO Eddie Wu says the company will invest long term in AI models, chips and cloud, aiming for a global cloud footprint above 20 gigawatts by 2032.

More
  • The strategy treats machine reasoning as an industrial utility whose supply depends on coordinated models, accelerators, networking, storage and tools.
  • Alibaba says its M890 supernode already serves models above 2 trillion parameters and that the future V900 will offer three times the performance and support clusters scaling to 500,000 cards.
  • The company’s recursive-self-improvement work extends from model training to inference optimization and electronic-design automation, linking software intelligence to hardware development.
  • A 20 gigawatt cloud target signals a capital-intensive infrastructure race in which energy, supply chains and data-center construction may constrain AI more than algorithms.
  • Alibaba is also recruiting 5,000 people in servers, networking, chips, databases and AI, showing the workforce dimension of the buildout.
  • Most figures come from an Alibaba-sponsored presentation, so the target should be treated as strategic intent rather than an assured delivery schedule.
Open at QbitAI Ask Claude ChatGPT

Shanghai brain-computer startup raises $28M for noninvasive ultrasound BCI

# Leiphone

BCI-Sonics raised 200 million yuan (roughly $28M) to combine focused ultrasound, multimodal brain reading and AI decoding in a noninvasive brain-computer platform.

More
  • The company's architecture separates writing, reading and decoding: focused ultrasound targets deeper brain regions, functional ultrasound and EEG provide signals, and AI closes the feedback loop.
  • A key engineering issue is individualized phase correction through the skull, where thickness and curvature distort the acoustic field and can shift the stimulation focus.
  • The startup says it has completed two generations of research systems, is pursuing clinical studies and plans to enter clinical testing in the first quarter of 2027; these are development milestones, not approval.
  • Its three product lines target research, clinical treatment and consumer enhancement, a wide scope that creates both platform leverage and regulatory risk.
  • Chinese policy and capital are increasingly treating BCI as a future industry, but the field still needs reproducible safety, efficacy and long-term neurological evidence.
  • The financing matters because leading Chinese investors are backing a noninvasive route that may be easier to scale than implanted systems, even if its signal quality and control precision are lower.
Open at Leiphone Ask Claude ChatGPT

The Xi-Trump summit is a test of whether mistrust can be managed

# SCMP

Chinese analysis presents the White House meeting as a test of whether tactical detente can become workable trust, with AI governance among the unresolved issues.

More
  • The article treats the summit as more than ceremony because mistrust has become the default operating condition between the two governments.
  • AI is notable on the agenda, but both sides are described as recognizing risks without yet producing meaningful regulatory cooperation.
  • The downside case is not merely a failed trade announcement: the analysis warns that a summit without durable mechanisms could push the relationship back toward a colder strategic cycle.
  • The useful executive takeaway is that export controls and technology rules sit inside a broader trust problem; a technical agreement can be reversed if the political channel collapses.
  • Chinese coverage gives the meeting a system-level frame, asking whether communication and guardrails can survive beyond the leaders' personal exchange.
Open at SCMP Ask Claude ChatGPT

Rare-earth friction clouds the next US-China trade truce

# SCMP

Chinese and US negotiators are still disputing strategic-mineral exports ahead of a Xi-Trump summit, exposing how fragile even a narrow tariff truce remains.

More
  • US Trade Representative Jamieson Greer accused Beijing of creating uncertainty by not supplying enough strategic minerals, while the report says both sides claim progress on a limited tariff mechanism.
  • The dispute shows that minerals remain a bargaining instrument inside broader trade diplomacy rather than a normal commodity-flow issue.
  • For technology companies, the relevant risk is supply continuity: a political disagreement can reach magnets, batteries, motors, defense systems and other manufacturing inputs.
  • The timing gives China leverage before a leadership meeting, but also creates pressure to demonstrate that negotiated commitments produce tangible shipments.
  • A durable extension would require verifiable export procedures and predictable licensing, not merely a summit statement.
  • The report provides no final agreement, so the immediate signal is unresolved friction rather than a settled policy change.
Open at SCMP Ask Claude ChatGPT

Alibaba Cloud builds an Agentic Cloud stack

# Leiphone

Alibaba Cloud unveiled AgentCore, Agent Sandbox, a new CPFS storage system and context tooling to make long-running agents observable, isolated and economically viable in production.

More
  • AgentCore manages models, tools, skills, permissions, retries, checkpoints and audit trails, treating the agent runtime as enterprise infrastructure rather than an application add-on.
  • Alibaba says Agent Sandbox can create 100,000 environments per minute and wake a deeply sleeping sandbox in under 600 milliseconds, while Agentic OS is designed to reduce token use and raise deployment density.
  • The new CPFS is claimed to deliver hundreds of terabytes per second, 100 million IOPS and a 100-petabyte filesystem, targeting the data-loading bottleneck in multimodal and reinforcement-learning workloads.
  • Tair KVCM coordinates memory pools, KV-cache storage and remote storage; Alibaba reports up to 99% effective cache hits and a 50% reduction in per-token cost.
  • Context Engine treats enterprise data, memory and agent feedback as a governed data asset that must be assembled dynamically at inference time.
  • The stack shows Alibaba trying to own the full control plane for Chinese enterprise agents, from domestic supernodes to sandbox execution and business context.
  • Vendor claims such as 99% task completion and 70% lower total cost require customer-side validation, but the product shape is strategically coherent.
Open at Leiphone Ask Claude ChatGPT

Xi and Trump bring leverage, not trust, to Washington

# SCMP

Xi Jinping and Donald Trump are meeting with tariffs, rare earths, Taiwan and technology controls as bargaining chips, making incremental stability more plausible than a full reset.

More
  • The summit follows negotiations over economic and technology issues but does not erase the structural conflict over security, industrial policy and market access.
  • China’s leverage includes critical minerals, market access and manufacturing scale; the US retains financial, semiconductor and alliance leverage.
  • The practical question is whether both sides can create bounded channels for crisis management and trade while continuing to compete in strategic technologies.
  • A leader-level agreement may calm markets without changing the underlying permissions, export controls or investment-screening regimes companies must operate under.
  • The visit’s business-deliverable problem, also reported separately, suggests political choreography is ahead of commercial trust.
  • Executives should treat any pause as time to reconfigure supply chains and compliance, not as a return to the pre-friction operating environment.
Open at SCMP Ask Claude ChatGPT

Huawei stays in Aito as Seres takes operational control

# TMTPost

Seres will lead Aito's product definition, sales network and service operations, while Huawei continues supplying driving, cockpit and vehicle-cloud technology inside the Harmony Intelligent Mobility ecosystem.

More
  • The change is a division-of-labor reset, not a breakup: Aito will build more independent retail and branding while remaining a core member of Huawei's partner platform.
  • Seres owns 10% of Huawei's vehicle-software joint venture, giving it a deeper technical link than a conventional supplier relationship.
  • The strategic tension is brand dependence: Seres wants its own premium identity and economics, but Aito still relies on Huawei for the capabilities that created its initial differentiation.
  • The report says Aito delivered more than 420,000 vehicles in 2025 and represented over 70% of the partner system's deliveries, making the governance change material to both companies.
  • Huawei says it will continue selling autonomous-driving, cockpit and vehicle-control solutions to Seres, so the near-term risk is execution and channel conflict rather than immediate technology withdrawal.
  • If the arrangement works, it could become a template for Chinese automakers to graduate from technology-led co-branding to clearer ownership without losing access to a platform partner.
  • The company's recent stock rise shows investor approval, but it does not prove that service quality or future product access will remain unchanged.
Open at TMTPost Ask Claude ChatGPT

A $2,800 household robot confronts China's valuation fever

# TMTPost

Qiyuan launched two consumer robots at 19,999 yuan, or about $2,800, while its listed parent carries a roughly $85 billion market value despite modest current revenue and losses.

More
  • The company is positioned as the consumer arm of a robotics ecosystem with industrial expertise, but the report says its core household technology is independently developed and only partly shares underlying resources.
  • The products include an 88cm humanoid and a convertible wheeled humanoid quadruped, with shipments scheduled for October 1.
  • The central commercial issue is not whether the robot can perform a staged demo but whether it can operate safely around children, older adults and unpredictable household environments.
  • Industrial robotics experience can help with manufacturing and supply chains, but consumer systems face different requirements for weight, battery life, fall behavior, joint design and liability.
  • The report says the parent had 210 million yuan, or roughly $30 million, in robot-related customer advances, not recognized revenue; that distinction matters when investors price future growth.
  • The comparison with Segway is apt: technical novelty and consumer demand are separate variables, especially at a price near $2,800.
  • This is a useful test of whether China's robotics equity enthusiasm is being converted into repeatable consumer economics.
Open at TMTPost Ask Claude ChatGPT

China's empty rural schools are becoming boutique hotels

# TMTPost

More than 320,000 rural primary schools have reportedly been closed or consolidated over two decades, and hotel groups are repurposing some of the buildings for rural tourism.

More
  • The trend connects demographic decline, urbanization and school consolidation to a new asset-reuse market: intact buildings, courtyards and playgrounds become accommodation for short rural breaks, families and corporate retreats.
  • The reported school count fell from 491,000 in 2001 to 163,000, leaving a large but geographically fragmented stock of public or formerly public property.
  • Hotel operators see lower construction costs and built-in village locations, while local governments see a way to monetize idle assets without large demolition projects.
  • The conversion is not automatically a success: weak transport, unclear land rights, renovation costs, seasonal demand and local employment determine whether a school becomes a viable hotel or another stranded asset.
  • The consumer signal is equally important: younger urban travelers increasingly buy low-density experiences and rural immersion rather than only standardized city hotels.
  • The story makes demographic contraction visible as a physical transformation of public infrastructure, not just as a population statistic.
  • It also shows how policy encouragement for incremental rural redevelopment can intersect with hospitality industry's search for differentiated demand.
Open at TMTPost Ask Claude ChatGPT

Xiaomi open-sources MiMo V2.6 models for science and agents

# IT Home

Xiaomi released open-weight multimodal MiMo V2.6 models; Chinese coverage highlights a Lean proof and a materials-research workflow, while US coverage focuses on the benchmark ranking.

Also: Techmeme

More
  • Xiaomi says the Pro model helped design metal-organic-framework candidates, simulate properties and narrow options for PFAS adsorption, reducing a reported month-long workflow to two or three days.
  • The model also reportedly completed more than 6,000 lines of Lean-verified code for a classical theorem, a stronger technical signal than a fluent demonstration because the proof kernel checks the output.
  • The published API prices are 1 yuan input and 2 yuan output per million tokens for Flash, and 3 yuan and 6 yuan for Pro, roughly $0.14 to $0.85 per million tokens.
  • The US discussion emphasizes that MiMo V2.6 Pro ties a leading proprietary score and tops open-weight models, while Chinese coverage stresses scientific use and low cost.
  • These are vendor or benchmark claims; the practical question is whether the model can reproduce the research results outside curated workflows.
  • Open weights make the event strategically relevant to local deployment, especially where Chinese organizations want control over data and inference.
Open at IT Home Ask Claude ChatGPT

China investigates DeepSeek and Moonshot over alleged data leaks

# Techmeme

US coverage leads with Anthropic’s alleged data-leak claims, while the reported Chinese response is a cyberspace regulator investigation and staff questioning at DeepSeek and Moonshot.

More
  • The item comes from a US-origin Techmeme headline, so its value is the Western framing of a Chinese regulatory and data-security dispute rather than independently verified investigation details.
  • The alleged conduct involves routing sensitive user data through Claude, making the case about model competition, data sovereignty and cross-border use of foreign AI services.
  • If confirmed, questioning staff would show Chinese regulators treating model-to-model data flows as a national cybersecurity and industrial-policy issue.
  • The story could mark a harder line against Chinese labs using foreign models for training, evaluation or product operation, but the supplied report does not establish the scope or findings of the investigation.
  • The contrast matters: Western coverage emphasizes possible appropriation of Anthropic data, while the Chinese state response is framed through regulatory control and security.
  • No Chinese-origin counterpart is supplied here, so the regulator’s position and any domestic media framing remain unknown.
  • This is a high-value watch item because an investigation could affect model development practices, cloud access and cross-border AI partnerships.
Open at Techmeme Ask Claude ChatGPT

Alibaba's machine-intelligence pitch ties AI to industrial scale

# AI QbitAI

Alibaba's CEO says AI will turn reasoning into a mass commodity, with Qwen, domestic chips and cloud data centers serving as the necessary industrial base.

Also: InfoQ China, TMTPost, Techmeme

More
  • The speech compares today's coding agents with early electric lamps and argues that the larger opportunity lies in new applications enabled by abundant machine reasoning.
  • Alibaba says Qwen is testing recursive self-improvement and plans models in the 5T to 10T-parameter range, but the report supplies no independent validation.
  • The company also targets clusters of up to 500,000 accelerators and more than 20GW of operated data centers by 2032, making power and systems engineering central to its strategy.
  • Chinese coverage is unusually explicit about infrastructure, capacity and national industrial buildout rather than presenting AI mainly as a software product.
  • The useful executive takeaway is that Alibaba expects the scarce resource to shift from model access to reliable, affordable inference at very large scale.
Open at QbitAI Ask Claude ChatGPT

Huawei proposes a modular 3D data-center design for AI

# AI Zhidx

Huawei says its new 3D data-center architecture separates cooling, IT, power and backup layers, using prefabricated modules to support dense AI clusters and faster delivery.

More
  • The vertical design places cooling, IT equipment, power and backup batteries in separate layers, shortening power runs from 17 meters to 5 and raising claimed capacity from 76 megawatts to 168 megawatts in the same footprint.
  • Huawei says mechanical and electrical prefabrication exceeds 90 percent and cuts delivery from about six months to three, turning site construction into assembly of tested power and cooling modules.
  • The architecture creates smaller fault domains because power, cooling and IT systems are separated, while predictive maintenance monitors leaks and coolant health before failures.
  • The proposal addresses a real AI infrastructure constraint: rack power is moving toward 100 to 200 kilowatts and large clusters make flat layouts, long pipes and shared failure modes expensive.
  • The strategic bet is standardization. If 3D pods become repeatable products rather than bespoke projects, Chinese data-center builders could deploy capacity faster and make future accelerator swaps less disruptive.
  • The headline claims are Huawei's own, but the design is a technically substantive response to the spatial, thermal and electrical demands of agent-scale inference.
Open at Zhidx Ask Claude ChatGPT

China's state-owned firms open real quantum-computing problems to startups

# Policy QbitAI

Five central state-owned enterprises launched a quantum innovation contest with more than 50 real industry problems, lab access and a six-month incubation path.

More
  • The tracks cover communications, dual-use technology, energy, finance and exploration, moving quantum work from generic demonstrations toward specific measurement, security and optimization tasks.
  • The organizers say teams can access major experimental facilities and sanitized business data, including quantum-computing cloud resources and field platforms for energy and mineral exploration.
  • A state-owned enterprise-led pipeline links contest selection to incubation, capital, procurement and possible investment or acquisition, reducing the gap between laboratory results and industrial customers.
  • The event also emphasizes domestic quantum programming frameworks, tying application development to China's broader push for controllable software and hardware supply chains.
  • The immediate value is ecosystem formation rather than a new quantum milestone: large industrial buyers are defining demand and offering test environments before the technology is mature.
Open at QbitAI Ask Claude ChatGPT

Qualcomm unveils dual 2nm flagship chips for agentic phones

# AI Zhidx

Qualcomm's new 2nm mobile platforms target on-device agents, shared memory and 30-billion-parameter mixture-of-experts models, with Chinese phones and services among the first demonstrations.

More
  • The chips combine a 5 GHz custom CPU, larger shared memory, a Transformer accelerator and persistent sensor processing to reduce the cost of moving context between memory and compute.
  • Qualcomm says the platform can run 30-billion-parameter MoE models on-device and supports local context, but model quality, thermal limits and sustained performance remain unverified.
  • The China angle is substantive: Xiaomi, Honor, Nubia and Chinese agent services are part of the launch ecosystem, showing how Qualcomm is using Chinese device makers to seed an agent-phone category.
  • The architectural bet is heterogeneous execution: local models handle sensing and private context while cloud and tool calls handle larger or fresher workloads.
  • This could make the handset an agent runtime rather than a thin client, but only if operating systems and third-party apps expose reliable APIs and permissions.
Open at Zhidx Ask Claude ChatGPT

Hygon enters embedded computing with a 10-watt x86 chip

# AI Zhidx

Chinese chipmaker Hygon launched its Hygon 1000 embedded CPU, combining x86 compatibility, 10-watt power, wide-temperature operation and a promised ten-year lifecycle.

More
  • The four-core, eight-thread chip supports dual 4K displays, DDR4 and DDR5 memory, industrial temperatures from minus 40C to 105C and passive cooling.
  • Hygon is using its data-center C86 software ecosystem to reduce migration costs for factory controllers, robots, edge gateways and machine-vision systems.
  • The product’s strategic value is lifecycle reliability: industrial customers often keep equipment for a decade and cannot accept frequent platform changes.
  • Hygon is expanding from China’s domestic server and accelerator base into the much larger installed base of embedded industrial equipment.
  • Compatibility with existing APIs and operating systems may matter more than peak performance in these deployments, where security, supply continuity and real-time behavior are decisive.
Open at Zhidx Ask Claude ChatGPT

China's AI infrastructure debate moves beyond simply adding GPUs

# Signal QbitAI

Inspur argues that agent workloads require separate solutions for model capability and inference capacity, with new systems designed around memory, communication and heterogeneous load patterns.

More
  • The report cites an IDC forecast that global token consumption could grow at a 4,822.6% compound rate through 2030 while AI-demand fulfillment falls from 79% in 2024 to 71% in 2027.
  • It says the absolute global compute shortfall could reach $380.9 billion by 2030, but these figures come from a vendor-sponsored presentation and should be treated cautiously.
  • The technical distinction is useful: prefill is compute-heavy, decode is memory-bandwidth-heavy, attention stresses growing key-value caches and MoE adds routing overhead.
  • Inspur's SD200 Ultra uses 128 chips, 8TB of memory and 64TB of expansion storage, and claims it can run a 2.8T-parameter model in one system.
  • The architecture uses unified addressing to make remote memory look more like one address space, aiming to reduce explicit message passing and data copies.
  • This is the larger infrastructure trend behind Chinese supernode announcements: effective utilization and communication topology matter as much as accelerator count.
Open at QbitAI Ask Claude ChatGPT

Inspur targets agent scale with a 128-chip AI supernode

# AI Zhidx

Inspur says its SD200 Ultra can run the 2.8-trillion-parameter Kimi K3 on one system, while a heterogeneous compute rack targets the far larger volume of agent inference.

More
  • The SD200 Ultra couples 128 domestic AI chips with 64 TB of memory and 8 TB/s bandwidth, using a 3D interconnect and unified memory addressing.
  • Inspur reports 0.69 microsecond cross-chip latency and a token-generation delay below 5.85 milliseconds for Kimi K3, although the workload and measurement conditions are not fully specified.
  • The HC2000 rack is designed for capacity rather than peak capability, assigning prefill, decode and different model types to different accelerators through a shared software layer.
  • The company says the rack exceeds 300 kilowatts and can hold up to 256 AI cards, underscoring the power and cooling burden of agent-scale inference.
  • The architecture reflects a Chinese infrastructure response to two different bottlenecks: fitting frontier models and producing enough cheap tokens for millions of agents.
  • Whether domestic chips can deliver this performance in production will depend on compiler, runtime and model-porting maturity as much as silicon.
Open at Zhidx Ask Claude ChatGPT

Chinese practitioners say FDE cannot substitute for enterprise data work

# Analysis Leiphone

A Chinese industry investigation argues that many so-called forward-deployment engineers are consultants or demo builders, while weak data foundations and client power dynamics block real AI adoption.

More
  • The central critique is architectural: an agent cannot produce reliable business results when source data is inaccurate, fragmented or full of organization-specific terminology that has not been mapped.
  • Practitioners describe FDE teams that lack a product, cannot meet customers directly or simply feed requirements into a model, turning deployment into a sales accessory rather than an accountable engineering function.
  • The article highlights enterprise semantic layers as a hard problem because local terms and business definitions require human alignment; a general model cannot infer them safely from sparse examples.
  • Chinese buyers often remain the stronger party in the vendor relationship, making it difficult for an FDE team to obtain the system access and authority needed to redesign workflows.
  • The useful distinction is between consulting, implementation and genuine deployment engineering: the latter must own data access, runtime integration, evaluation and business outcomes.
  • This is a valuable corrective to the current FDE hype because it describes the organizational prerequisites that cloud vendors' agent demos tend to hide.
Open at Leiphone Ask Claude ChatGPT

China’s premier highlights advanced manufacturing and AI in Shanghai

# Policy Sinocism

Premier Li Qiang toured Shanghai AI and future-industry facilities, calling for closer integration of manufacturing and digital technology and cheaper upgrades for smaller firms.

More
  • The visit links industrial policy to practical deployment: robotics, AI factories and pilot bases are presented as ways to raise manufacturing competitiveness rather than as standalone research showcases.
  • Li's emphasis on low-cost modernization for small and medium firms suggests Beijing wants AI adoption to diffuse through the industrial base, not remain concentrated in frontier labs.
  • The inspected future-industry community spans AI, advanced nuclear energy, life sciences, space computing, brain interfaces, quantum computing and silicon photonics.
  • This is a policy signal about state-backed ecosystem building and Shanghai's role as an innovation hub, not evidence that every showcased technology is commercially mature.
  • The combination of a national leader, municipal incubation platforms and industrial funds shows how China uses state visits to connect research prestige with capital allocation and factory deployment.
Open at Sinocism Ask Claude ChatGPT

DeepSeek bets on domestic chips while China’s AI sector expands abroad

# AI Leiphone

Chinese coverage bundles several AI developments, but the substantive China signal is DeepSeek’s reported plan to train larger models on domestic chips and distribute Kimi through overseas cloud platforms.

More
  • DeepSeek reportedly wants Huawei or other domestic accelerators to train a 2 trillion-parameter model and eventually an 8 trillion-parameter model, making hardware substitution a strategic priority.
  • The source says Huawei could begin delivering training chips in the fourth quarter, but gives no volume, benchmark or confirmed supplier agreement.
  • Separately, Kimi K3’s availability through Amazon Bedrock would make a Chinese open model directly reachable by global enterprise developers and introduce a revenue-sharing route with foreign clouds.
  • The combination points to two complementary strategies: reduce dependence on foreign compute at home while using global distribution to monetize Chinese model capability abroad.
  • The report also mentions internal security measures at DeepSeek after earlier leaks, showing how secrecy and domestic hardware policy are becoming linked in the model race.
  • Several other items in the source are unrelated headlines, so the claims should be evaluated separately rather than treated as one corporate announcement.
Open at Leiphone Ask Claude ChatGPT

Qwen Office adds a hardware doorway to enterprise context

# AI Leiphone

Qwen Office launched an AI recording card and an enterprise-context layer that turns meetings, chats and documents into agent-ready inputs and actions.

More
  • The QwenNote A2 uses six microphones, claims eight-meter pickup and 98 percent transcription accuracy, and can operate through an embedded cellular connection without a phone.
  • Its product distinction is that audio is not merely transcribed: the agent can analyze the conversation, update systems, send messages and create follow-up work.
  • Enterprise Context addresses the opposite problem by compressing huge internal data stores into the small, task-specific context an agent can actually use.
  • The combination shows Alibaba building both ends of the context pipeline: a physical sensor for offline conversations and a data layer for online organizational memory.
  • The privacy design reportedly keeps text summaries rather than raw audio, but customers still need clear retention, consent, access and cross-tenant isolation guarantees.
Open at Leiphone Ask Claude ChatGPT

Alibaba says unified multimodal models are three years away

# AI QbitAI

Alibaba’s Qixiang conference laid out a portfolio spanning language, image, video, audio and world models, with the company predicting a native unified model within three years.

Also: InfoQ China

More
  • The company says Qwen 3.8, Image 3.1, audio and music models, video systems and a world-model preview are being developed as a coordinated stack rather than separate products.
  • A film project reportedly rebuilt a historical disaster scene in five days, showing the intended path from model capability to a professional production workflow.
  • Alibaba’s argument is that real tasks are inherently multimodal, so separate models must eventually share context, editing state and agent control.
  • The proposed five-to-ten-trillion-parameter scale for later Qwen generations is a capacity and systems challenge as much as a modeling ambition.
  • The practical bottleneck is stable delivery across long workflows: preserving characters, facts, assets and user intent matters more than a single impressive generation.
Open at QbitAI Ask Claude ChatGPT

Cainiao finds that 90% AI-generated code barely speeds delivery

# AI InfoQ China

Cainiao says AI-generated code rose above 90% of commits while delivery cycles improved only about 10%, prompting an agent-led cloud-sandbox pipeline for end-to-end work.

More
  • The result separates code production from software delivery: requirements, design, testing, deployment and human handoffs still dominate the critical path.
  • Cainiao's proposed system lets agents run continuously in cloud sandboxes, carry organizational playbooks and move from requirements through staging deployment with humans approving selected gates.
  • This is a practical architecture for legacy-heavy enterprises: the agent needs plugins, environment access and repeatable workflows, not just a coding model.
  • The reported contribution rate is a line-count metric and should not be read as proof of quality; the more meaningful result is the small change in delivery-cycle time despite high automation.
  • The company is effectively treating engineering throughput as a queueing and coordination problem, with human attention as the bottleneck between automated stages.
Open at InfoQ China Ask Claude ChatGPT

Taiwan coastguard fires an anti-ship missile in a rare drill

# Policy SCMP

Taiwan's coastguard fired a supersonic anti-ship missile during an exercise, signaling deeper operational integration with the navy and complicating PLA planning.

More
  • The unusual part is institutional: a civilian law-enforcement agency is practicing a naval strike role, blurring the boundary between maritime policing and wartime defense.
  • Analysts cited in the report say missile-equipped coastguard vessels would create additional targets and defensive problems for the People's Liberation Army in a conflict.
  • The move fits Taiwan's broader effort to make more of its maritime fleet useful in a gray-zone or invasion scenario without presenting every platform as a conventional warship.
  • The report does not establish a new permanent deployment or a change in rules of engagement, so the exercise is a signal rather than proof of a doctrinal shift.
  • For US planners and suppliers, the relevant trend is distributed defense: civilian hulls, reserve forces and dual-use systems could enlarge the battlespace before a formal conflict begins.
Open at SCMP Ask Claude ChatGPT

Huawei adds a HarmonyOS coding model and full development agent

# AI InfoQ China

Huawei Cloud says its HarmonyOS coding model cuts generated-code errors by more than 80% and connects requirements, ArkTS coding, compilation and testing.

More
  • The model was trained on more than 100,000 high-quality HarmonyOS examples plus APIs, development rules, component assets and layout templates.
  • Huawei reports an 80% plus reduction in errors per thousand generated lines, a 78% plus increase in one-pass compilation and a 20% reduction in token use in internal tests.
  • The associated agent can invoke the DevEco command-line tools and a device simulator, turning generated code into a compile-and-preview loop rather than a text-only assistant.
  • The product is also available through Huawei's model marketplace, giving enterprises a route to call the specialized model outside the main development environment.
  • This is ecosystem defense through tooling: as HarmonyOS tries to attract developers, a specialized model can lower migration and debugging costs even if the language ecosystem remains smaller than Android's.
  • The reported metrics are vendor tests, so real value will depend on performance on large legacy projects and cross-platform migration.
Open at InfoQ China Ask Claude ChatGPT

China's vehicle complaints rise as software quality deteriorates

# Signal 36Kr

Chinese automotive data show nearly 180,000 valid complaints in the first eight months, with quality complaints up 78.86% and first failures arriving after only 7.4 months on average.

More
  • The main trouble spots are intelligent cabins, driving assistance and battery systems, including incorrect NOA responses, inaccurate range estimates and charging failures.
  • Traditional engines and transmissions improved, while software freezes, slow systems and neglected updates affected both electric and combustion vehicles.
  • The report says pure-electric vehicles delivered only 71.46% of advertised summer range on average, with winter performance about 11 percentage points lower.
  • The pattern suggests that rapid development cycles are consuming the verification margin: hardware, software policy and manufacturing integration are each becoming failure surfaces.
  • China has approved a national standard for automotive software quality and defect management, signaling that regulators are moving from post-sale complaints toward pre-release software governance.
  • For automakers, the competitive metric is shifting from component cost and feature count toward the gap between promised and experienced behavior.
Open at 36Kr Ask Claude ChatGPT

TokenRhythm links model routing to agent self-improvement

# AI QbitAI

Chinese AI infrastructure company TokenRhythm says model routing can preserve task quality while cutting cost, then feed agent execution traces back into model improvement.

More
  • Its OpenSquilla system reportedly retained 99.96% of a fixed flagship-model baseline while reducing cost 88.9% in a specific evaluation configuration.
  • The architectural idea is to route each step to a model selected for capability, latency and price rather than committing one model to an entire long task.
  • The company’s feedback loop uses task evaluation to identify weak models or routing policies, then feeds successful and failed trajectories into later training.
  • Its NeoHorse models reportedly improved benchmark averages after post-training on Qwen-derived models, but the figures come from company materials and need independent replication.
  • The strategic implication is that the agent runtime becomes a data-collection and model-selection layer, potentially giving infrastructure vendors influence over future model quality.
Open at QbitAI Ask Claude ChatGPT

China’s model market is concentrated in four low-cost leaders

# Signal Leiphone

A weekly Chinese market tracker says four lightweight models account for about 66% of observed domestic token usage, with price cuts rapidly reshaping demand.

Also: 36Kr

More
  • DeepSeek V4 Flash, DeepSeek V4.1 Flash, Tencent Hy4 preview and GLM 5.3 Flash each exceeded 10 trillion weekly tokens.
  • DeepSeek V4.1 Flash usage rose 483.62% after a price cut, showing how elastic inference demand can be when models are interchangeable enough.
  • The top three models captured nearly half of usage and the top five more than three quarters, indicating that distribution and cost matter as much as benchmark leadership.
  • The report’s data comes from OpenRouter, vendor pricing and arena tests, so it is a directional market sample rather than a census of China’s private enterprise deployments.
  • For infrastructure buyers, the market is already bifurcating into high-cost frontier calls and enormous volumes of inexpensive flash inference.
Open at Leiphone Ask Claude ChatGPT

China appoints a new party chief in Tibet amid regional reshuffling

# Policy SCMP

Beijing named Hu Changsheng party chief of Tibet and shifted leaders in Gansu and Qinghai ahead of next year's expected Party congress.

More
  • The moves rotate provincial and autonomous-region party chiefs rather than announce a policy change, making personnel placement the main signal.
  • Hu moves from Gansu to Tibet, while Qinghai and Gansu also receive new leadership, suggesting a coordinated regional reshuffle rather than an isolated Tibet appointment.
  • Party secretaries control the political direction of provinces and autonomous regions, so these changes matter more than ordinary government cabinet turnover.
  • The timing ahead of the 21st Party congress makes loyalty, administrative experience and succession management more relevant than any individual policy portfolio.
Open at SCMP Ask Claude ChatGPT

Qwen plans a personal agent for 300 million users

# AI Leiphone

Alibaba’s Qwen is building a personal agent that can use authorized health, financial and activity data while expanding an ecosystem of more than 1,000 partners.

More
  • The plan moves Qwen from chat toward persistent, user-specific assistance in health, exercise, finance, work and study.
  • Alibaba says financial and health workflows use specialized skills, external data and user-authorized holdings or sensor information rather than generic model knowledge alone.
  • The product raises the central governance question for Chinese consumer AI: who controls the permission, memory and liability boundaries when the assistant acts on sensitive personal data?
  • Qwen says more than 60% of users with complex finance, health or lifestyle needs also turn to a work assistant, a company statistic with no independent methodology.
  • An ecosystem of skills, agents, data providers and hardware could make distribution and context more important than the base model itself.
Open at Leiphone Ask Claude ChatGPT

China hosts a global forum on governing medical AI

# Policy Leiphone

A United Nations-backed health AI meeting held in Hangzhou highlighted China's effort to shape international standards for clinical safety, evaluation and primary care.

More
  • The meeting brought together the World Health Organization, International Telecommunication Union and World Intellectual Property Organization, giving Chinese medical-AI companies an international standards forum rather than a domestic showcase.
  • China's participating companies presented primary-care assistants, health profiles and integrated clinical workflows as examples of practical deployment.
  • The recurring governance themes were common benchmarks, intellectual-property protection, regulatory coordination and cross-border standards, all areas where deployment claims need more than model accuracy.
  • The event's arrival in China signals greater institutional participation in global medical-AI governance, but hosting a forum is not evidence that the proposed systems have passed independent clinical validation.
  • For executives, the relevant distinction is between AI that assists triage and administration and systems making clinical decisions; the latter will require much stronger evidence and liability rules.
Open at Leiphone Ask Claude ChatGPT