China Desk

China News, Summarized 14 Aug 2026, 22:01 UTC 34 stories 18/19 sources

Today's brief

China's most valuable company now sells memory, and Tencent's own earnings explain the swap

A Hefei DRAM maker passed Tencent

CXMT is now the biggest listed company in China. Seventeen days after its Shanghai debut, the Hefei memory maker closed Friday at roughly 3.58 trillion yuan of market value against Tencent's HK$4.0 trillion — about 3.44 trillion yuan, which is the ranking Chinese outlets are actually computing. Bloomberg put Thursday's crossover at $524bn to $511bn.

The crossover was Tencent falling, not CXMT rising. CXMT's own shares fell 1.2% on the day; the swap happened because Tencent dropped further after disclosing a 176% surge in AI capital spending. Allspring's Gary Tan gave the line of the week: "chips are the new clicks."

Tencent published the bill on page seven. Marketing services revenue hit 43.565bn yuan (~$6.1bn), up 22% against a Chinese internet ad market that grew 4.6% last year on QuestMobile's count — and gross margin on that line slipped from 58% to 57%. Ad serving used to carry near-zero marginal cost. Now every impression burns GPU.

Strip the AI products out and the company looks untouched. Remove Yuanbao, CodeBuddy, WorkBuddy and the rest and non-IFRS operating profit goes from 75.6bn to 86.1bn yuan — 19% growth instead of 12%. That 10.5bn yuan (~$1.5bn) gap is one quarter's cover charge. Depreciation inside fintech and business services rose 44%.

The tell is who Tencent buys from. It signed a $3bn server DRAM agreement with CXMT in June; ByteDance signed a five-year deal worth more than $7bn in July, and server products went from 8.4% of CXMT's revenue in 2024 to 26.5% in 2025 — investors are penalising the buyer of AI hardware while rewarding its supplier, over the same purchase orders. Wednesday's note on Tencent hoarding memory now has a scoreboard.

Go deeper on this section: ClaudeChatGPT

Apple gets a licence, drones get a tariff

Apple trained a model for China only. Reuters reports Apple built a China-specific LLM with Alibaba's support, which would make it the first foreign company to offer a proprietary AI model in the country. The Cyberspace Administration has registered Apple's generative service, clearing Apple Intelligence for mainland iPhones after two years of watching Huawei eat its share with AI handsets.

The registration is the story, not the model. Beijing has spent that time insisting every model filed and approved; the first foreign firm through the gate got there with a domestic partner attached. Buy the licence, share the training.

Washington answered with drones. The White House imposed tariffs on unmanned aircraft and components, including a 100% levy on "particularly sensitive" models. Second day running that the trade move is American and the mainland feed doesn't mention it.

And Beijing quietly shut the offshore trust door. New Ministry of Finance and tax administration rules issued 24 July tax offshore trusts at asset transfer, at income generated inside, and at distribution, against an industry estimate of 2.3 to 2.7 trillion yuan (~$324–380bn) in such structures. Li Daokui, Tsinghua finance professor and dean of its economic-practice centre, welcomed them — then noted capital income tops out at 20% while labour income reaches 45%: "That's clearly unfair" — and proposed capping all personal income tax at 20%. The fairness argument runs both ways, and the second way is a tax cut for the people the first way just caught.

Go deeper on this section: ClaudeChatGPT

GLM-5.3 found a flaw forty years old

Zhipu shipped GLM-5.3 overnight on the same base model as 5.2 — 744bn total, 40bn active, IndexShare architecture — with everything else coming from post-training. Terminal-Bench 3.0 went from 4.6 to 28.3 and DeepSWE v1.1 from 46.2 to 66.9; on Zhipu's own Code Bench it beat Claude Opus 4.8 at the hardest tier, 31.4% to 29.5%, using about 50,000 tokens per task where Opus needs 120,000.

The security numbers are the actual release. CyberGym rose to 84.5% from 5.2's 77.2%, above Anthropic's Mythos 5 and OpenAI's GPT-5.6. Working with Tsinghua, Nankai and Chinese security labs, Zhipu says it found 2,436 vulnerabilities across 269 projects, some up to 40 years old, 1,097 of them medium-to-high severity, now in the national vulnerability database's repair queue. The oldest sits in the design of DNS itself: a few crafted requests amplify server load nearly 80,000-fold.

The other case reads like fiction. A tracking engine flagged an odd temporary file server in Brazil; with the model in the loop, researchers reconstructed an autonomous AI attacker they call Neo — four servers, seven malicious domains, 60,000 mailboxes, over 100 attack scripts and 18,567 phishing emails aimed at accounting and tax firms, all inside two months. Defence AI unpicking offence AI, and the Chinese write-ups know exactly how that sounds.

Go deeper on this section: ClaudeChatGPT

Alibaba sells the shortcut, ByteDance refuses one

Alibaba Cloud brought thirty carmakers into its own data centre. The Ulanqab site carries 60% of China's intelligent-driving R&D compute, and the pitch was that model, chip and cloud used to be three suppliers at three different conferences. What they are selling is not GPU hours but a pipeline: raw clips in from the fleet, automatic cleaning, vectorisation, model-assisted labelling, millions of clips a day, model iteration measured in days.

The argument underneath is the good one. Stored data depreciates when the architecture switches from end-to-end to vision-language-action; the pipeline doesn't. Meanwhile T-Head's line on superpods — that most on sale are Ethernet server stacks and the only real test is unified addressing with synchronous memory semantics — is a fair shot, and their 64-card single-layer symmetric design is the answer they want judged.

ByteDance is still paying to avoid the shortcut. 36Kr's researchers describe distillation as a cold start that carries a model to 90 so you can work the last 10. Seed banned it at founding and Zhang Yiming reaffirmed that in late July even if it means trailing domestic rivals — Seed-2.1 Pro sits 16th on Code Arena's WebDev board against Kimi K3 at 2. DeepSeek's V4 report describes distilling a dozen domain experts into the generalist; K3 used nine teachers.

The split is by modality, and it explains everything. Seed's video model reached the frontier without distillation because generated video can't substitute for real footage. Language is precisely where borrowing works — which is precisely where ByteDance is behind. Nvidia's Jensen Huang calls learning from other models the foundation of intelligence; Zhang is betting a year of leaderboard humiliation that it's a ceiling.

Go deeper on this section: ClaudeChatGPT

What the filter blocks, and what it lets through

A WeChat essay on censorship is arguing that censorship is incompetent. The post, translated by China Digital Times, makes a careful case: no public directory of blocked terms exists and regulations don't require platforms to restrict ordinary commentary; platforms over-block because automated filters can recognise characters but not context, so they err on the side of caution and keep expanding the lists. Users invent homophones, platforms block those too, and the vocabulary keeps shrinking.

It is still online, which is the point. Six weeks after the Jieyang dog case, the sayable critique is that the machinery is clumsy, not that it is wrong. Initium meanwhile followed the grief across the border, to a Buddhist rite held in Hong Kong for the dog and the invisible limits of cross-border solidarity.

Elsewhere, the filter had a day off. Zhejiang Provincial Museum's Long March display turned the Red Army into a "red car" amid the red-tourism push. And during Wednesday's total eclipse, Xiaomi 17 Ultra photos of the sun came back with ridges and craters — the camera's reconstruction model knew what a bright disc in a dark sky is supposed to look like, and it was wrong.

Go deeper on this section: ClaudeChatGPT

Same model, two verdicts on the weights

Chinese and Western coverage of GLM-5.3 disagree on what happened. Leiphone's three labels — strongest security model, strongest coding model, strongest open-source model — and QbitAI's framing of Tang Jie's promise to Musk moving forward both treat open weights as a settled fact.

They aren't. The weights ship in two weeks, once security reviews wrap up — the first GLM release held back explicitly for safety review. English coverage led there, and on Zhipu adding vulnerability-discovery environments expecting better bug-spotting and getting a model that reasons across whole exploit chains, which it says was not the intended outcome. That admission is the most interesting sentence in the launch, and it's absent from the Chinese write-ups I read.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • DeepSeek's price rise still lands 16:00 UTC Sunday. Harness passed 50,000 GitHub stars in twelve hours, and a third-party comparison circulating among developers claims the same V4-Flash burns roughly three times the tokens in DeepSeek's harness as in a rival open one. Worth testing before Monday's invoice.
  • I under-weighted memory. I read the price rise as load-shedding by electricity. CXMT's re-rating and ByteDance's $7bn DRAM contract say the binding shortage is DRAM at least as much as power.
  • Zhu Rongji: still no funeral grade. SCMP ran a seven-story retrospective; Initium's podcast keeps the frame on centralisation, not liberalisation.
  • Alibaba's promised 27bn-parameter dense Qwen3.8 companion: day three, nothing.
  • New watch — Moonshot. It has reported fake "friend funds" and "reserved allocations" to police, its second such notice since June, after closing an F round at a $35bn valuation and opening a pre-IPO round early at $50bn pre-money. When the counterfeit share market gets busy, the listing is close.

Go deeper on this section: ClaudeChatGPT

The river · 34

China tightens taxation of offshore trusts

# Pekingnology 19h

New Chinese tax rules target offshore trusts at transfers, retained income and distributions; a Tsinghua economist frames them as tax-fairness reform while also criticizing the higher tax rate on labor income.

More
  • The rules were jointly issued by China's finance ministry and tax authority and make offshore structures harder for Chinese tax residents to use for deferral or avoidance.
  • The source says the rules affect wealthy trust holders far more than ordinary people, who generally do not own offshore trusts.
  • Professor David Li argues that taxing labor income at up to 45% while capital income tops out at 20% is inequitable, turning the debate from enforcement into broader tax design.
  • The policy fits Beijing's need to expand revenue collection and tighten oversight of wealth held through cross-border structures amid fiscal pressure.
  • For private wealth and corporate planning, the important shift is that offshore structures are no longer a dependable timing shield even when assets remain outside mainland China.
  • The eventual significance will depend on enforcement guidance, reporting cooperation and how authorities distinguish genuine estate planning from tax avoidance.
Open at Pekingnology Ask Claude ChatGPT

Why Chinese users feel they cannot speak plainly online

# China Digital Times 21h

A translated Chinese post argues that platform keyword filters block ordinary speech and push users toward coded language, revealing how censorship is experienced through private platform risk management.

More
  • The essay distinguishes formal legal prohibitions from the much larger set of words platforms suppress to avoid regulatory liability or moderation failure.
  • Its central claim is that overblocking is partly a platform-governance problem: companies choose blunt filters because missing harmful content can be more dangerous to them than silencing benign discussion.
  • Users respond with homophones, initials and coded terms, making everyday communication less direct even when the underlying topic is not clearly prohibited.
  • The framing is politically careful, blaming platform negligence while leaving the wider state censorship system largely implicit; that caution is itself informative about the boundaries of publishable criticism.
  • The post reportedly remains online, showing that criticism of filtering practices can survive when presented as an argument about platform quality rather than a direct challenge to political authority.
  • This is a valuable account of how censorship becomes infrastructure and user habit, not merely a list of banned subjects.
Open at China Digital Times Ask Claude ChatGPT

DeepSeek's Harness turns an agent into a plugin system

# GeekPark 17h

DeepSeek released an MIT-licensed developer preview whose model, tools, UI and control loop are all replaceable plugins, pointing toward an agent ecosystem rather than a sealed assistant.

More
  • Harness runs locally and keeps sessions, logs and data on the user's machine, while the browser serves as a front end to the local process.
  • Its core architectural choice is that even the agent loop and interface are plugins, so developers can replace models, tools, sandbox behavior and approval policies without editing the core.
  • The early repository attracted more than 50,000 stars in roughly half a day, a strong developer-interest signal but not evidence of production readiness.
  • The preview is deliberately unfinished: the framework is aimed at developers who will assemble their own workflows rather than ordinary users seeking a polished productivity app.
  • This approach could become an Android-like substrate for agents if plugin interfaces stabilize and security boundaries remain understandable.
  • The risk is ecosystem fragmentation: maximum composability can make permissioning, reproducibility and support harder than in a vertically controlled product.
Open at GeekPark Ask Claude ChatGPT

Former Chinese premier Zhu Rongji dies at 97

# SCMP 15h

Zhu Rongji, a central engineer of China's reform-era economic transformation, has died at 97; Chinese coverage revisits his managerial pragmatism while the broader lesson is how much the reform state depended on a few trusted operators.

More
  • Zhu is remembered as a co-architect of major economic reforms and as a premier associated with forceful administrative execution.
  • His legacy matters because the reform era combined market opening with strong central coordination rather than a clean withdrawal of the state.
  • The death invites comparison between a period of institutional and international integration and China's current emphasis on strategic control and industrial policy.
  • Zhu's public image was unusually managerial: competence, discipline and willingness to confront entrenched interests were treated as political virtues.
  • The retrospective value is not nostalgia alone; it helps explain why Chinese policy memory still invokes reform-era officials when debating growth, bureaucracy and market confidence.
  • How official commemorations frame Zhu's achievements will show which parts of that reform legacy remain politically useful.
Open at SCMP Ask Claude ChatGPT

Chinese team extends cold-atom entanglement to 420 kilometers

# SCMP 11h

Chinese scientists report entangling cold atoms across 420 km, more than four times earlier demonstrations, a potentially important step toward intercity quantum networks.

More
  • The result was published in Physical Review Letters and is described as exceeding the distance at which direct transmission faces fundamental limits.
  • Cold-atom entanglement is a laboratory resource for studying quantum links; it is not itself a deployable city-scale network.
  • The technical significance is architectural: longer-lived, distributed entanglement could support repeater-based networks instead of relying on direct transmission alone.
  • The result strengthens China's position in experimental quantum communications, where distance records are proxies for control, synchronization and channel engineering.
  • Its practical importance depends on fidelity, rate, memory lifetime and integration with repeaters, none of which are quantified in the supplied summary.
  • Even so, a fourfold distance increase is a meaningful research milestone rather than a routine product announcement.
Open at SCMP Ask Claude ChatGPT

Apple and Alibaba train a China-specific model for Apple Intelligence

# Solidot 14h

Apple is working with Alibaba on a model for China-specific Apple Intelligence features, showing how regulation and local competition are forcing a global platform to split its AI architecture by market.

More
  • The reported model is intended for iPhones and other Apple products sold in China, where Apple cannot simply deploy the same service stack used elsewhere.
  • The arrangement gives Alibaba a role in the model layer while Apple retains control of the device experience and distribution.
  • China's registration of Apple's generative-AI service is described as a prerequisite for launch, making regulatory approval part of the product architecture.
  • The move reflects competitive pressure from Chinese handset makers that already market AI-native features and have local data, service and compliance infrastructure.
  • For multinational technology companies, the strategic implication is that China support increasingly requires a locally governed model and partner ecosystem, not merely a localized user interface.
  • The source does not provide model size, training data, release timing or final commercial terms, so the integration remains a reported plan rather than a shipped product.
Open at Solidot Ask Claude ChatGPT

DJI gets another chance to challenge its Pentagon blacklist

# SCMP 1h

A US appeals court sent DJI's blacklist case back for reconsideration without removing it; Chinese coverage stresses legal persistence, while the US angle is the unresolved military-link claim.

More
  • The DC Circuit ruling does not clear DJI; it gives the Shenzhen drone maker another opportunity to challenge its designation in lower court.
  • The case shows how a company can remain commercially significant while being trapped in a security designation that is difficult to unwind through normal administrative channels.
  • For DJI, the immediate value is procedural rather than operational: the court has reopened a route to contest the Pentagon's reasoning.
  • The dispute illustrates the expanding overlap between US national-security screening and technology-market access for Chinese hardware companies.
  • Its strategic importance depends on whether the lower court addresses the substance of the designation or only corrects the earlier legal process.
Open at SCMP Ask Claude ChatGPT

Chinese doctor says GPT-5.6 helped solve a 22-year math problem

# IT Home 6h

A Beijing doctor reportedly used an OpenAI reasoning model to produce a proof of the Crouzeix conjecture that outside mathematicians reviewed; the unusual China story is also a reminder that verification matters more than the demo.

More
  • The source says mathematicians Alex Townsend, Anne Greenbaum and Michel Crouzeix reviewed the manuscript and considered the proof correct.
  • The researcher is a medical doctor with limited formal mathematics training who used autonomous multi-agent exploration and later released the paper, prompts and Lean code.
  • The claimed workflow matters technically because it treats a frontier model as a long-running search and verification system rather than a chat assistant.
  • The result is presented through a Chinese researcher who crossed disciplines, a pattern that fits China's strong public interest in exceptional educational and technical trajectories.
  • Independent checking of the formalization and the full proof remains the key condition for treating this as a durable mathematical milestone.
Open at IT Home Ask Claude ChatGPT

GLM-5.3 targets coding and open-model security

# AI QbitAI 13h

Chinese coverage portrays GLM-5.3 as a post-training leap that approaches top foreign coding models while finding thousands of software vulnerabilities, though the strongest evidence is vendor-linked.

Also: Leiphone

More
  • The release keeps the base model architecture but adds new optimization, reinforcement-learning and post-training systems aimed at long-running software work.
  • The supplied report says GLM-5.3 reached 84.5% on a white-box code-security benchmark, up from 77.2% for GLM-5.2.
  • Zhipu says red teams found 2,404 vulnerabilities across 220 projects, including 1,088 medium- or high-risk cases; these figures are not independently audited here.
  • The model's safety stack combines outer classification, reasoning-time monitoring and internal alignment, which is especially relevant when weights are released for local use.
  • The strategic claim is not simply that a Chinese model is better at code, but that post-training and security evaluation can narrow the gap without changing the foundation model.
  • The main missing evidence is independent reproduction across realistic repositories and a clear accounting of false positives in the vulnerability count.
Open at QbitAI Ask Claude ChatGPT

Alibaba open-sources Qwen3.8-27B

# AI QbitAI 6h

Chinese coverage highlights Qwen3.8-27B's open commercial license and compact multimodal design; US discussion is more skeptical and fragmented than the release story.

Also: Leiphone, Jiemian, IT Home, Techmeme, Hacker News

More
  • Qwen3.8-27B has 27 billion parameters, native multimodal input and a 262K context window that can be extended to 1M tokens with YaRN, according to the release text.
  • The model adds a reasoning-effort control so developers can trade deliberation depth for resource use, a practical knob for local or cost-sensitive deployments.
  • The source says Alibaba is releasing the weights under Apache 2.0 and claims improved coding and office performance over an earlier Plus model.
  • The release fits Alibaba's broader strategy of using open weights and distribution through public model repositories to build an external developer ecosystem.
  • The US-facing reaction should be read as market attention, not independent confirmation of the claimed performance advantage.
Open at QbitAI Ask Claude ChatGPT

Chinese robotics firms propose a standard for whole-body dexterity data

# AI QbitAI 15h

Delta Intelligence and Wujin Technology plan a shared data standard and training baseline for whole-body manipulation, addressing a major bottleneck in humanoid robotics: fragmented demonstrations that record hands but not posture.

More
  • The proposed dataset would capture hand, arm, torso and visual-focus coordination rather than treating dexterous manipulation as an isolated hand-control problem.
  • The companies plan to connect a lightweight data glove and direct-drive dexterous hand hardware to a foundation-model training and edge-deployment pipeline.
  • A common format, annotation scheme and evaluation baseline could reduce repeated integration work across robot projects.
  • The technical importance is that whole-body coordination is a structured learning problem: a successful grasp can still be physically implausible if balance, reach and gaze are wrong.
  • The standard is only planned for release this year, so its influence will depend on adoption beyond the two companies.
Open at QbitAI Ask Claude ChatGPT

What China's AI talent market actually pays for

# Analysis Ginger River Review 13h

A Chinese business podcast suggests the headline million-yuan AI salaries are rare; companies pay most for scarce combinations of research depth, deployment experience and data or infrastructure judgment.

More
  • The discussion distinguishes spectacular offers from the much larger market, where compensation varies by specialty, company stage and ability to deliver usable systems.
  • Recruiters say demand is shifting beyond foundation-model researchers toward infrastructure, chips, robotics, agents and people who can connect research to products.
  • The Chinese labor market is therefore rewarding compound technical experience: a person who understands models, data pipelines and organizational execution is more valuable than a narrow paper count alone.
  • The episode also suggests AI hiring is spreading beyond technology companies as finance, manufacturing and other sectors build internal applications.
  • This is qualitative interview material, not a systematic salary survey, but it is more informative than isolated compensation headlines because it explains the scarcity behind them.
Open at Ginger River Review Ask Claude ChatGPT

Chinese DRAM maker CXMT briefly surpasses Tencent in market value

# Signal Solidot 12h

CXMT's reported market value overtook Tencent's, a striking signal that Chinese investors are rewarding domestic memory and AI hardware exposure over mature internet platforms.

More
  • The source puts CXMT's value near 3.58 trillion yuan, roughly $500 billion, and Tencent's near 4.0 trillion Hong Kong dollars, roughly $560 billion; the two figures use different currencies and should not be compared without conversion.
  • CXMT is described as the world's fourth-largest DRAM supplier, serving devices from phones to advanced AI servers.
  • The reversal reflects investor expectations around memory demand from AI infrastructure, not proof that CXMT has surpassed Tencent as an operating business.
  • A valuation gap based on a single trading day can reverse quickly, especially after a newly listed semiconductor stock's sharp rally.
  • Even with that caveat, the episode shows how China's AI boom is redirecting capital toward the memory and compute supply chain.
Open at Solidot Ask Claude ChatGPT

Chinese robot teams are using ten times more compute

# AI QbitAI 14h

A Chinese robotics company says its compute demand rose five- to tenfold in two years, underscoring that embodied AI needs shared simulation and development infrastructure rather than only larger models.

More
  • The reported workflow combines remote workstations, physics simulation, data generation, model training and real-robot deployment, all of which compete for accelerator time.
  • Cloud workstations reduce the need for every engineer or event participant to maintain a separate CUDA, robotics and simulator environment.
  • The engineering bottleneck is reproducibility: version conflicts, fragmented data and access control can consume more time than model experimentation.
  • The story's deeper point is that robotics infrastructure is becoming a standardizable layer, with environments and scheduling treated as shared services rather than bespoke team projects.
  • The five- to tenfold demand figure comes from the company and is not a sector-wide measurement, but it captures the direction of travel.
Open at QbitAI Ask Claude ChatGPT

Baidu launches standalone Kuku AI office products

# AI QbitAI 13h

Baidu is packaging its document-and-storage agent into dedicated desktop, web, mini-program and enterprise products, with finance as the first specialized office vertical.

Also: Leiphone

More
  • Baidu reports more than 25 million monthly office users and says the broader service has passed 100 million monthly users; these figures come from the company.
  • The platform combines office-specific agents, multimodal editing, private files, long-term memory and cross-device task handoff.
  • The enterprise edition adds shared context, reusable skills and audit controls intended to turn individual AI assistance into organizational workflow infrastructure.
  • The product's differentiator is less a novel model than access to Baidu's large document corpus and the ability to transform stored knowledge into structured work products.
Open at QbitAI Ask Claude ChatGPT

China says its Y-20B can switch from transport to tanker duty

# Policy SCMP 11h

A state documentary says China's Y-20B transport aircraft can be reconfigured as an aerial refueling tanker, potentially extending air operations beyond the first island chain.

More
  • The Y-20B is described as an improved Y-20 variant powered by domestically developed WS-20 turbofans.
  • The proposed configuration uses pallet-ready fuel tanks and removable refueling equipment, allowing one airframe to serve transport and tanker roles.
  • A dual-use fleet could give the air force more flexibility than a dedicated tanker inventory, especially during surges or expeditionary operations.
  • The strategic value is reach: refueling can reduce the distance constraint on fighters and other aircraft operating in the western Pacific.
  • The source confirms a claimed capability through state television, but does not provide numbers for tanker conversion time, fuel offload or fleet size.
  • Whether the system changes the regional balance depends on those operational details and on the availability of escorts, bases and trained crews.
Open at SCMP Ask Claude ChatGPT

US coastguards train Taiwan amid increased Chinese activity

# Policy SCMP 9h

US instructors are training Taiwan's coastguard in small-boat operations as Chinese activity rises east of the island, adding a law-enforcement layer to cross-Strait friction.

More
  • Taiwan's Coast Guard Administration says the two-week program includes small-boat operations and team coordination.
  • The training is notable because it strengthens day-to-day maritime capability without being described as a formal military exercise.
  • Beijing's increased presence and Washington's practical support are therefore being managed through a gray-zone channel where coastguards, rather than navies, carry much of the signaling.
  • For China watchers, the risk is cumulative normalization: repeated operational cooperation can expand deterrence while also raising the chance of a misread encounter.
  • The source does not establish a new US security commitment, so the strategic weight lies in the trend rather than in this single training event.
Open at SCMP Ask Claude ChatGPT

China calls New Zealand's spy report foreign interference

# Policy SCMP 15h

China's embassy rejects a New Zealand security report that names China as the only country conducting espionage at scale, sharpening a bilateral dispute over how security claims are framed.

More
  • New Zealand's intelligence service says China poses a large-scale espionage threat, while Beijing calls the assessment a product of foreign interference and Cold War thinking.
  • The dispute is less about a single allegation than about competing definitions of legitimate security reporting and political interference.
  • New Zealand is balancing intelligence cooperation and economic ties with China against pressure to speak more openly about foreign influence.
  • Beijing's response follows a familiar diplomatic pattern: reject the factual premise, challenge the motives and warn against alliance-driven threat narratives.
  • The episode matters for companies and researchers because national-security language increasingly shapes the operating environment for cross-border technology and academic exchange.
  • The source does not establish whether the report will trigger new restrictions, so the immediate significance is political signaling.
Open at SCMP Ask Claude ChatGPT

Alibaba Cloud pitches a full data-to-model stack for Chinese driving

# Analysis GeekPark 11h

Alibaba Cloud argues that China's autonomous-driving bottleneck is the data-to-model production line, not simply more GPUs, and is packaging cloud, models and simulation as one system.

More
  • The platform claims it can process millions of driving clips per day through ingestion, cleaning, vectorization, automated labeling and retrieval for training.
  • The architectural argument is that data quality and iteration cadence are coupled: a cluster without a fast sample pipeline leaves model teams waiting even when compute is available.
  • Alibaba positions Qwen as a possible VLA base and Wanxiang as a world-model engine for generating and evaluating rare driving scenarios.
  • This is a strategic move to sell an integrated production system rather than isolated storage or compute, binding vehicle companies more tightly to the cloud stack.
  • The reported throughput and model positioning come from Alibaba Cloud's forum presentation, so independent validation remains missing.
  • If the data pipeline genuinely shortens model iteration from weeks to days, it could become a stronger moat than a single model benchmark.
Open at GeekPark Ask Claude ChatGPT

SHEIN's IPO is about preferred-share pressure, not cash

# Signal 36Kr 17h

SHEIN generated about $6 billion in operating cash over three years and held $14.8 billion in cash, yet its IPO is driven partly by $17.3 billion of convertible preferred stock and redemption pressure.

More
  • The company's cash generation exceeded its roughly $3.8 billion of historical equity financing, showing a marketplace model funded heavily by customer prepayments rather than new capital.
  • The preferred shares carry contractual features that can accelerate investor returns and revive redemption rights if the IPO fails or remains inactive, making listing a liability-management event.
  • SHEIN's valuation fell from a reported $98.2 billion pre-money level to $64 billion while revenue and orders continued to grow, illustrating how market multiples can compress even as the business expands.
  • The company's China-rooted supply network remains strategically important, but the IPO story is increasingly about international demand, US exposure and whether growth can reaccelerate.
  • This is a useful case study in how private-company financing terms can force a public listing even when the operating business does not need fresh cash.
Open at 36Kr Ask Claude ChatGPT

A guide to China's best music of 2026 so far

# Culture ChinaTalk 11h

A China-focused music guide highlights funk, indie rock, post-punk and avant-garde releases, offering a view of an urban music scene that is creative but still shaped by public-space controls.

More
  • The selection spans artists from Shenzhen and other Chinese scenes, showing a culture that is not reducible to state-approved pop or imported genres.
  • One band reportedly got into trouble with police while livestreaming a Shanghai street performance, a small example of how public-space regulation can shape independent music.
  • The guide treats Chinese musicians as part of a global listening economy through Bandcamp, podcasts and overseas attention.
  • Its value is cultural texture rather than a hard market claim: informal scenes, regional identities and genre experimentation continue despite constrained venues and distribution.
Open at ChinaTalk Ask Claude ChatGPT

Chinese RISC-V chip startup files for an IPO while burning cash

# Signal 36Kr 17h

Einstar Computing is seeking a Hong Kong listing after losing about $690 million over three years, betting state-backed capital can turn RISC-V into a strategic alternative despite thin market share.

More
  • The company reported revenue growth but annual losses of roughly $260 million, $218 million and $214 million for 2023 through 2025, using 7.1 yuan per dollar for orientation.
  • Its 2025 gross profit was about $64 million against roughly $85 million in administrative costs and $147 million in research spending, so scale has not yet solved the cost structure.
  • The company is described as a major domestic supplier of human-machine-interface chips and a smaller RISC-V controller supplier, with its strongest ranking appearing only when foreign competitors are excluded.
  • A large customer accounted for 82.1% of revenue in 2023, though that dependence had fallen to 39.3% by the first quarter of 2026.
  • The IPO highlights the tension in China's semiconductor strategy: strategic substitution can justify large funding rounds before commercial economics are proven.
  • Its survival depends on converting state and industrial support into diversified demand, not simply winning a national-label market-share ranking.
Open at 36Kr Ask Claude ChatGPT

ByteDance's no-distillation bet makes model iteration slower

# Analysis 36Kr 20h

Chinese reporting says ByteDance's Seed is refusing model distillation to build a more independent training system, accepting slower progress in language models while relying on proprietary data strengths in video.

More
  • Distillation lets a weaker model learn answers or task trajectories from a stronger model, providing a fast cold start before further training.
  • Researchers quoted in the article say abandoning it can delay coding and agent progress because a model must first become useful enough to attract real user data and feedback.
  • The report contrasts language models with ByteDance's video-generation work, where high-quality real video data is harder to replace with outputs from another model.
  • ByteDance has reportedly created a dedicated AI data and safety organization to handle sourcing, synthetic data, cleaning and evaluation, making data independence an organizational commitment.
  • The strategic tradeoff is sovereignty versus velocity: an independent pipeline may produce a more distinctive model, but competitors using distilled capabilities can reach deployment and feedback loops sooner.
  • The claims come from interviews and reporting rather than a public technical disclosure, so it is unclear how absolute the no-distillation rule is across all model teams.
Open at 36Kr Ask Claude ChatGPT

Museum caption errors embarrass China's revolutionary-history campaign

# Culture SCMP 12h

Errors in Communist Party history displays, including a caption turning the Red Army into a red car, expose the quality-control burden behind China's push for revolutionary tourism.

More
  • The mistakes are especially embarrassing because museums and local governments give political-history exhibitions a high degree of symbolic importance.
  • The source links the errors to a broader campaign promoting red tourism, where historical presentation is also a form of civic education.
  • The episode shows how centralized narrative priorities can coexist with decentralized execution and weak editorial controls.
  • Public criticism of a caption is not necessarily criticism of the underlying political narrative, but it does reveal how easily credibility can be damaged by basic errors.
  • For organizations managing sensitive historical content, the operational problem is less ideological consistency than review, provenance and accountability.
Open at SCMP Ask Claude ChatGPT

Sanrio's China licensing machine is powering its rebound

# Culture 36Kr 12h

Sanrio's Chinese growth is being driven by Alibaba's licensing platform, showing how local distribution and character localization can turn foreign intellectual property into a Chinese consumer ecosystem.

More
  • Alibaba Fish has exclusive mainland rights to 26 Sanrio characters for five years, and the article says more than 200 brands signed licensing deals in 2024.
  • Sanrio's China sales reached 31.44 billion yen, roughly $210 million, up 83.2% in the cited fiscal year; Kuromi, not Hello Kitty, led Chinese sales.
  • The operating model combines mass licensing with a rapidly expanding store footprint, while reducing reliance on a single character.
  • This is a useful contrast with Pop Mart: Sanrio remains relatively asset-light and licensing-led, while Pop Mart relies more heavily on direct retail and the blind-box mechanism.
  • The Chinese market is not just an export destination here; it is an operating and merchandising partner that reshapes which characters matter.
  • The next test is whether the exceptional growth rate survives a larger base and whether Sanrio can add narrative depth to otherwise open-ended character brands.
Open at 36Kr Ask Claude ChatGPT

China is warned not to make its AI body an anti-US bloc

# Policy SCMP 13h

Analysts urge China to present its new global AI organization as inclusive rather than an answer to US alliances, revealing the diplomatic constraint behind Beijing's technology-governance push.

More
  • The World Artificial Intelligence Cooperation Organisation is being compared with the US-backed Pax Silica partnership, but analysts say it was not designed as a direct alternative.
  • The warning is about coalition design: potential members may avoid a body that appears to reproduce great-power bloc politics under a different label.
  • China is trying to turn AI governance into a domain of international institution-building, but its credibility will depend on whether the organization can attract states beyond existing political partners.
  • The episode shows that standards and governance forums are now part of technology competition, not merely technical coordination.
  • The source describes analyst concerns rather than a formal change in the organization's mandate, so the practical test is membership and agenda-setting.
Open at SCMP Ask Claude ChatGPT

China and Brazil plan a South America disaster-warning satellite

# Policy SCMP 8h

China and Brazil unveiled a satellite intended to monitor all of South America around the clock, turning a bilateral space project into regional infrastructure and diplomatic leverage.

More
  • Brazil's space research institute says the meteorological and environmental satellite will provide broad coverage of South America after nearly four decades of China-Brazil space cooperation.
  • The project would extend Chinese technical involvement into a public-good service: weather and environmental monitoring rather than a purely commercial launch.
  • Regional value will depend on data access, operational reliability and whether South American governments treat the system as shared infrastructure.
  • The initiative shows how China uses long-running scientific partnerships to build influence through practical infrastructure rather than only through trade or lending.
  • The source provides no launch date or final technical specification, so the capability remains a planned project rather than an operating asset.
Open at SCMP Ask Claude ChatGPT

Tencent moves a senior post-training researcher to WeChat AI

# AI 36Kr 11h

A senior Hunyuan researcher has moved to Tencent's WeLM team as WeChat accelerates its own models and agents, showing that Tencent's AI organization remains internally plural.

More
  • The researcher is known for synthetic data, instruction construction, reinforcement-learning feedback and automated evaluation, all directly relevant to post-training and agent reliability.
  • The move comes as WeChat tests an assistant that can operate native functions and call mini-programs, giving the model team a concrete product surface.
  • Tencent has been consolidating several AI resources under Hunyuan, but WeChat retains a relatively independent model effort through WeLM.
  • That structure matters: the internal competition is not simply between Tencent and outside labs, but between a centralized foundation-model organization and a massive product group with its own data and deployment needs.
  • The transfer suggests WeChat is treating model quality and tool execution as strategic product infrastructure rather than a thin interface layer.
Open at 36Kr Ask Claude ChatGPT

Researchers report an unexpected structure inside nucleons

# Signal Solidot 16h

A China-linked collider study reports evidence for an internal nucleon structure carrying baryon number but neither electric charge nor color charge, challenging the simplest quark-and-gluon picture.

More
  • The result comes from precision measurements of net baryon number and net electric charge in high-energy nucleus collisions and was published in Science.
  • The proposed structure is not an additional ordinary quark: the reported quantum-number combination would require a more complex description of how nucleons are organized.
  • The finding is a fundamental-physics result, not a near-term technology milestone, but it matters because nucleons account for most of the visible mass in the universe.
  • The key question is whether independent experiments and theoretical work reproduce the anomaly and identify the underlying degrees of freedom.
Open at Solidot Ask Claude ChatGPT

SenseTime proposes measuring data centers by tokens per watt

# Signal Leiphone 8h

SenseTime's computing-center unit is promoting tokens per watt as a complement to PUE, arguing that AI infrastructure should optimize useful output and electricity timing rather than power efficiency alone.

More
  • TPW is intended to connect electricity use, accelerator utilization, model efficiency and business output in one metric, extending beyond PUE's facility-level view.
  • The proposed control loop maps jobs to physical power use, then shifts delay-tolerant training and batch work toward cheaper or less constrained periods.
  • SenseTime says its approach raised effective IT carrying capacity by about 60% within existing power limits; this is a company claim from sponsored coverage.
  • The conceptual shift matters because AI infrastructure is increasingly sold by completed work or tokens rather than by server capacity alone.
  • The missing piece is an industry-wide definition of a useful token and comparable measurement rules, without which TPW can become another vendor-specific metric.
Open at Leiphone Ask Claude ChatGPT

Moonshot AI needs a new IPO story for Kimi K3

# Analysis 36Kr 17h

Moonshot AI's rumored Hong Kong listing would turn Kimi K3's temporary frontier performance into capital, but compute shortages and rapid model competition make the window fragile.

More
  • Kimi K3 reportedly ranked near the top of a third-party model leaderboard while maintaining lower task costs, giving Moonshot a stronger listing narrative than a generic domestic-model story.
  • The company previously said it had more than 10 billion yuan, roughly $1.4 billion, in cash and did not need a near-term listing, but public markets offer a continuing financing channel as well as capital.
  • A surge in K3 demand reportedly pushed the company's cluster toward capacity limits and temporarily suspended new consumer subscriptions, exposing compute as a constraint on monetization.
  • DeepSeek's much lower reported task cost threatens the combination of frontier quality and efficiency that K3 is using to distinguish itself.
  • The strategic issue is timing: an IPO can monetize a lead before competitors copy it, but a lead in foundation models decays quickly.
Open at 36Kr Ask Claude ChatGPT

Alibaba may sell its Lingxi game business for more than $1.5 billion

# Culture 36Kr 12h

A reported private-equity sale of Alibaba's Lingxi game unit would mark another retreat from Chinese internet conglomerates' old ecosystem-expansion strategy toward cash generation and AI.

More
  • The reported buyer is PAG, with a possible valuation above $1.5 billion; neither side had confirmed the deal in the supplied text.
  • Lingxi's main asset is a long-running strategy game with more than 100 million global registered users, while newer titles have not reached comparable scale.
  • The article frames the unit as a mature cash-flow business whose value depends heavily on one durable product rather than a broad pipeline of hits.
  • Alibaba's shift in reporting lines and wider noncore-asset restructuring suggest gaming is being judged against group priorities such as AI and cloud.
  • The strategic lesson is that traffic and capital did not automatically reproduce the development culture built by Tencent and NetEase; content businesses remain organizationally specialized.
Open at 36Kr Ask Claude ChatGPT

Insta360 wants users to forget the camera after pressing record

# AI GeekPark 11h

Insta360's new X6 shifts the product from capture to autonomous local editing, using on-device analysis, head tracking and 3D reconstruction to close the content-production loop.

More
  • The camera's AI Director analyzes a day's footage locally and produces a video without first sending clips to a phone or cloud service.
  • Head movement becomes an intent signal: a tracking module tells the system where the user looked and therefore which parts of a 360-degree scene may matter.
  • A separate 3D time-capsule feature uses Gaussian splatting to reconstruct static scenes so viewers can move through them rather than only rotate their viewpoint.
  • The product illustrates a broader edge-AI pattern in which physical gestures become prompts and the device performs a long-running workflow while unattended.
  • The reconstruction feature is still early because dynamic scenes are difficult, but the local editing pipeline appears closer to a repeatable consumer behavior than a novelty demo.
Open at GeekPark Ask Claude ChatGPT

China wants Chancay Port on the wider APEC agenda

# Policy Jiemian 6h

Chinese coverage frames the China-Peru port as a vehicle for regional connectivity and agenda-setting during China's APEC host year, not merely a bilateral infrastructure deal.

More
  • The article argues that Chancay Port could move from a China-Peru project into a topic for wider Asia-Pacific cooperation when Shenzhen hosts the 2026 APEC leaders' meeting.
  • Its strategic logic is diplomatic as much as commercial: the host country can use agenda control to promote connectivity and economic integration.
  • The port is presented as a test of China's response to fragmented trade governance, geopolitical scrutiny and uncertainty around overseas infrastructure projects.
  • The framing reveals how Chinese policy commentary links physical logistics assets to institution-building and regional economic governance.
  • The article is advocacy-leaning analysis rather than evidence that APEC members have accepted the proposal.
Open at Jiemian Ask Claude ChatGPT
Source ledger
ChinAI 403 0/0 HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk 200 1/20
Pekingnology 200 1/20
Ginger River Review 200 1/20
Sinocism 200 0/20
QbitAI 200 8/10
36Kr 200 9/30
Solidot 200 11/20
Leiphone 200 8/20
GeekPark 200 4/30
IT Home 200 12/60 +40 over cap
SCMP 200 10/50 +6 over cap
Initium 200 4/15
Jiemian 200 8/30 +14 over cap
China Digital Times 200 1/15
The Wire China 200 0/50
Techmeme 200 1/15
Hacker News 200 2/20
V2EX 200 8/50 +39 over cap