China Desk

China News, Summarized 14 Aug 2026, 20:38 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 18h

New rules from China's finance and tax authorities tax offshore trusts at transfer, income and distribution stages, closing a widely used deferral route and reopening debate over capital versus labor tax rates.

More
  • The rules were issued on July 24 by the Ministry of Finance and the State Taxation Administration, according to the report, and target offshore trust structures used by Chinese tax residents.
  • The change matters because it reaches multiple moments in a trust's life rather than waiting only for money to return to China, making tax deferral harder to engineer through legal structure.
  • A Tsinghua finance professor welcomed the move as a fairness measure but argued that China's top labor-income rate of 45 percent is out of line with the 20 percent ceiling on capital income.
  • The article says industry estimates put domestic offshore-trust assets at about 2.3 trillion yuan, roughly $324 billion; the figure is an estimate rather than an official balance sheet.
  • This is both a revenue measure and a governance signal: Beijing is asserting that wealth held through offshore vehicles remains within the practical reach of the domestic tax system.
  • The immediate effect is concentrated on wealthy families and trust advisers, but the precedent matters for how China treats cross-border wealth, succession planning and capital mobility.
  • The unresolved policy question is whether closing the trust route will be paired with a broader redesign of personal taxation, or remain a targeted enforcement action.
Open at Pekingnology Ask Claude ChatGPT

DeepSeek Harness is an open, plugin-based agent workbench

# GeekPark 16h

DeepSeek Harness passed 50,000 GitHub stars in about half a day, while Chinese hands-on coverage emphasizes a rough local workbench whose radical feature is that every major component can be replaced.

Also: Techmeme

More
  • Harness runs locally through a command-line launch and browser interface, with sessions, logs and data kept on the user's machine rather than in a hosted service.
  • Its plugin model exposes the model, tools, UI, sandbox and control loop, allowing developers to assemble an agent environment without modifying the core source.
  • The rapid star count is a community-attention signal, not proof of production adoption; the review says the interface is still unfriendly to nonprogrammers and the ecosystem is immature.
  • Chinese testing reportedly used it to rebuild a website and query GitHub data, showing a path from prompt to external tool use and artifact delivery rather than chat alone.
  • The architecture resembles an operating system more than a finished application: its strategic value rises with third-party plugins, but so do compatibility, permissions and security risks.
  • DeepSeek is therefore competing for the control plane around agents, not only for model benchmark rankings.
  • The decisive test will be whether teams can maintain reproducible, secure plugin stacks after the initial novelty and star surge fade.
Open at GeekPark Ask Claude ChatGPT

What China's AI talent market is actually paying for

# Ginger River Review 11h

Chinese recruiter and media interviews suggest the market pays less for celebrity credentials than for scarce execution skills spanning data, infrastructure, models, robotics and agents.

More
  • The discussion treats million-yuan offers to new PhDs as real but exceptional, not a normal salary benchmark for China's wider AI workforce.
  • Recruiters say demand is shifting toward people who can connect research to production: data pipelines, compute systems, model post-training, robotics and agent deployment.
  • This is a labor-market correction to the headline salary story. Companies are buying the ability to make a system work repeatedly, not merely a paper, affiliation or fashionable model keyword.
  • The Chinese discussion also shows how AI hiring is spreading beyond specialist labs into ordinary businesses that need people who can translate models into workflows.
  • For executives, the durable shortage is likely hybrid talent that understands both the model and the surrounding organization, while star-researcher premiums remain concentrated at the frontier.
Open at Ginger River Review Ask Claude ChatGPT

Former premier Zhu Rongji, China's economic reformer, dies at 97

# SCMP 13h

Former premier Zhu Rongji died at 97, closing the chapter on a leader remembered in China and abroad as an engineer of market reform and state restructuring.

More
  • Zhu served as premier during a period of major economic transformation and is credited with helping drive reforms that reshaped China's state sector and integration with the global economy.
  • His reputation rests on a willingness to confront institutional resistance, fiscal problems and inefficient state enterprises even when the political costs were high.
  • The supplied coverage frames him as a chief engineer of reform, an image that captures both his technocratic style and the era's belief that administrative redesign could unlock growth.
  • Zhu's legacy is now read against a different policy environment in which the state again emphasizes strategic control, industrial policy and political discipline alongside markets.
  • His death matters because the generation that managed China's reform-era transition is disappearing, while its choices continue to shape debates over what economic openness should mean.
  • The response to his death will also show how much room remains for public discussion of reform-era officials and the tradeoffs they represented.
Open at SCMP Ask Claude ChatGPT

Chinese memory maker CXMT briefly overtakes Tencent in market value

# Solidot 10h

Chinese coverage says CXMT reached about 3.58 trillion yuan, roughly $500 billion, overtaking Tencent in market value and showing how strongly investors are betting on domestic memory and AI hardware.

More
  • The ranking reversal reflects a sharp capital-market preference for semiconductor exposure over a mature internet platform, according to the report.
  • CXMT is described as the world's fourth-largest DRAM supplier, putting the valuation claim against a real manufacturing position rather than a purely speculative chip design story.
  • DRAM is foundational to phones and AI servers, so the market is pricing memory as a strategic bottleneck in the generative-AI buildout.
  • The figures are market snapshots, not proof that CXMT's profits or technology have overtaken Tencent's business; semiconductor valuations can move far ahead of operating results.
  • The China-specific signal is industrial-policy leverage: a domestic memory champion can attract exceptional capital when investors believe AI demand and supply-chain autonomy will reinforce each other.
  • The key follow-up is whether production yield, capacity and cash generation justify the valuation after the initial listing enthusiasm fades.
Open at Solidot Ask Claude ChatGPT

DJI gets another chance to challenge its Pentagon blacklist

# SCMP 14m

A US appeals court sent DJI's blacklist case back for reconsideration; Chinese coverage stresses a legal opening, while the US action still leaves the designation intact.

More
  • The ruling does not remove DJI from the Pentagon list, but requires a lower court to reconsider part of the company's challenge, extending a dispute that has lasted years.
  • The case illustrates how Chinese technology firms face overlapping US national-security restrictions, with courtroom process becoming one of the few remaining routes to relief.
  • DJI is a Shenzhen-based drone maker whose commercial products sit at the intersection of consumer electronics, surveillance concerns and military-supply-chain policy.
  • Chinese coverage presents the decision as another chance to contest what DJI says is an unjust designation; the immediate US legal result is narrower than that framing suggests.
  • The strategic question is whether a court finding about procedure or evidence can change the broader political consensus in Washington, rather than merely delay enforcement.
Open at SCMP Ask Claude ChatGPT

China and Brazil plan a 24/7 disaster-warning satellite

# SCMP 6h

China and Brazil unveiled a satellite intended to monitor all of South America, with Chinese coverage presenting the project as both climate infrastructure and a vehicle for broader regional cooperation.

More
  • The satellite is designed for continuous meteorological and environmental monitoring across South America, according to Brazil's space research agency.
  • The project builds on nearly four decades of China-Brazil space cooperation, giving Beijing a long-running institutional channel rather than a one-off commercial contract.
  • For Brazil and neighboring countries, the practical value is earlier warning for floods, storms and other hazards; the source does not yet provide launch timing, sensor specifications or financing terms.
  • The strategic value for China is diplomatic and technical: shared space infrastructure creates recurring data relationships and demonstrates an alternative to relying solely on US or European systems.
  • The project will matter more if regional agencies can access and operationalize the data, not merely if the spacecraft reaches orbit.
Open at SCMP Ask Claude ChatGPT

China team extends cold-atom entanglement record to 420 km

# SCMP 9h

Chinese scientists report entangling cold atoms across 420 kilometers, more than four times earlier demonstrations and beyond the range where direct transmission remains straightforward.

More
  • The result was published in Physical Review Letters, according to the report, giving the claim a peer-reviewed basis rather than relying only on a laboratory announcement.
  • A 420-kilometer baseline is relevant because it approaches intercity-scale quantum networking and crosses a threshold where direct transmission encounters fundamental limits.
  • The experiment is a test bed, not a deployed quantum network: the source does not establish usable throughput, fault tolerance or long-term operational stability.
  • The technical importance lies in distributing entanglement between separated atomic systems, a prerequisite for future repeaters and networked quantum sensors.
  • China's broader value in this field is the ability to combine precision experiments with long-distance infrastructure ambitions, although the path from record distance to useful networking remains long.
  • The result matters most if subsequent work demonstrates repeatability, memory lifetime and integration with existing optical or satellite links.
Open at SCMP Ask Claude ChatGPT

RISC-V chipmaker Eeasy heads for Hong Kong after heavy losses

# 36Kr 15h

Eeasy Computing, founded by former BOE chief Wang Dongsheng, is pursuing a Hong Kong IPO after losing about 4.9 billion yuan, roughly $690 million, in three years while betting on RISC-V as an alternative architecture.

More
  • The company raised 9.276 billion yuan, roughly $1.31 billion, across four rounds, yet had only 188 million yuan, about $26 million, in cash at the end of 2025, according to the filing analysis.
  • Its 2025 gross profit of 453 million yuan, roughly $64 million, did not cover 604 million yuan of administrative expense or 1.042 billion yuan of research spending.
  • Revenue rose from 1.752 billion yuan in 2023 to 2.431 billion yuan in 2025, but the company still expects another net loss, showing that domestic-chip substitution does not automatically produce viable economics.
  • The business remains concentrated: one customer supplied 82.1 percent of 2023 revenue, falling to 39.3 percent in the first quarter of 2026, and the filing description strongly suggests a connection to BOE.
  • The RISC-V market is fragmented, with the company holding only a small share overall; its stronger ranking appears only after limiting the comparison to Chinese suppliers.
  • The IPO therefore represents a test of China's capital model for strategic semiconductors: national substitution can attract funding long before scale, margins and customer concentration are solved.
  • The key question is whether the company can turn architectural independence into recurring design wins before its cash burn forces another financing cycle.
Open at 36Kr Ask Claude ChatGPT

A Chinese doctor says GPT-5.6 helped prove a 22-year math conjecture

# IT Home 5h

A Beijing doctor reportedly used OpenAI's GPT-5.6-Sol to produce a proof of the Crouzeix conjecture that outside mathematicians reviewed, highlighting a striking but still source-dependent AI research claim.

More
  • The report says mathematicians Alex Townsend, Anne Greenbaum and Michel Crouzeix reviewed the manuscript and found the proof correct; the supplied account does not provide a journal publication or formal peer-review record.
  • The researcher is a Beijing neurosurgery postdoctoral doctor with geology and medical training rather than a conventional mathematics career, and says his mathematical knowledge was largely self-taught.
  • The reported workflow used many parallel agents, adversarial checking and a disconnected environment, with the researcher leaving the system to work autonomously for about 16 hours.
  • The proof was reportedly released with prompts, drafts, Lean 4 code and an axiom audit, which is the right kind of artifact for testing whether the result is reproducible rather than merely persuasive.
  • The broader implication is not that models have replaced mathematicians: human experts still had to inspect the argument, and the claim's durability depends on independent formal verification.
  • If confirmed, the episode would show a more consequential role for models as search systems over proof strategies, especially for researchers who can frame a problem and audit the output without being domain specialists.
Open at IT Home Ask Claude ChatGPT

Qwen3.8-27B targets local deployment on consumer GPUs

# AI QbitAI 4h

Chinese coverage presents Qwen3.8-27B as a commercially usable, locally deployable model, with a 262,000-token context and adjustable reasoning effort.

Also: Leiphone, IT Home, Techmeme, Hacker News

More
  • The model is a 27 billion parameter native multimodal dense system released under Apache 2.0, according to Alibaba's authorized announcement.
  • The reported context window is 262,000 tokens natively and up to 1 million with YaRN; a reasoning-effort control is intended to trade answer quality against inference cost.
  • The claim that quantized weights run smoothly on consumer GPUs is deployment-dependent and should not be read as a universal hardware guarantee.
  • Alibaba says the Qwen family has more than 3 billion global downloads, more than 460 released models and more than 300,000 derivative models; these are company-reported ecosystem figures.
  • The strategic value is distribution: permissive weights plus manageable inference can make Qwen a default base for private deployments and derivative products.
Open at QbitAI Ask Claude ChatGPT

GLM-5.3 claims to deliver stronger coding with fewer tokens

# AI QbitAI 11h

A Chinese hands-on review presents GLM-5.3 as a major coding and security upgrade, with long-running autonomous demos that are more informative than its promotional benchmark labels.

Also: Leiphone

More
  • The review says GLM-5.3 completed a 3D animation task, an interactive watch disassembly and a small game with accounts, persistence, permissions and tests, though the demonstrations were not independently audited.
  • The reported workflow had the model inspect, modify and validate its own code over tens of minutes, testing whether it can maintain a task state rather than answer isolated programming prompts.
  • The company claims much lower token use than a competing model on difficult software tasks; the review does not provide enough experimental detail to establish a general efficiency advantage.
  • Security is a second emphasis: the model reportedly identified more than 2,400 vulnerabilities, including more than 1,000 medium- or high-risk findings, across hundreds of projects.
  • The useful signal is the evaluation target. Chinese labs are increasingly measuring agents inside realistic repositories and tool environments, not only on static question-answer benchmarks.
  • The missing evidence is production reliability: regression rates, hidden tests, failure recovery and performance across unfamiliar codebases.
Open at QbitAI Ask Claude ChatGPT

Why Chinese users feel they cannot speak openly online

# Policy China Digital Times 19h

Chinese coverage argues that blunt platform keyword filters are suppressing ordinary speech, forcing users into coded language while allowing platforms to avoid political responsibility by blaming compliance pressure.

More
  • The translated post says users routinely replace ordinary words with homophones, initials and other codes after harmless comments are flagged, muted or blocked.
  • It distinguishes national law from platform practice: laws define prohibited categories, but the large-scale blocking of everyday social commentary is presented as a product of overcautious platform governance.
  • The platform incentive is asymmetrical. Companies can be punished for allowing harmful content, while users bear the cost when automated systems overblock harmless expression.
  • This produces a distinctive form of censorship experienced as technical friction: people learn to write for the filter, not for one another, and meaning becomes less direct and less searchable.
  • The post's choice to blame platform negligence rather than directly challenge the state censorship regime is itself revealing, because it identifies a narrower space in which users can criticize implementation.
  • For technology leaders, the issue is a governance architecture in which risk scoring, moderation operations and political accountability are inseparable.
  • The post reportedly remains online, suggesting that criticism of blunt enforcement can sometimes pass when framed as platform quality and administrative failure rather than a direct challenge to the political system.
Open at China Digital Times Ask Claude ChatGPT

Robot training demand in China has risen up to tenfold

# Signal QbitAI 13h

A Chinese robotics company says its compute demand grew five to ten times in two years, showing why embodied AI is becoming an infrastructure and workflow problem rather than only a model problem.

More
  • Robot development combines teleoperation data, simulation, model training and real-machine validation, each requiring different compute and software environments.
  • The report describes cloud workstations that package GPU resources, Isaac Sim and related toolchains so teams can share reproducible environments instead of configuring each local machine.
  • The five-to-tenfold demand estimate is an executive's rough figure, not a sector-wide dataset, but it captures the shift from single-engineer experiments to continuously running team infrastructure.
  • The strategic bottleneck is environment reproducibility and data movement: driver, simulator and robotics-library mismatches can consume more time than model experimentation.
  • Cloud elasticity may reduce idle hardware and setup friction, but sensitive robot data, latency and physical validation still keep part of the stack close to the lab.
Open at QbitAI Ask Claude ChatGPT

A star Meta researcher leaves after one year to start a company

# Analysis QbitAI 5h

Chinese coverage treats Yu Jiahui's departure from Meta as evidence that huge AI compensation cannot lock in top researchers, while the concrete destination remains undisclosed.

More
  • The report says Yu joined Meta in mid-2025, helped lead the TBD Lab multimodal effort and then left to pursue an unspecified research venture.
  • His year at Meta reportedly produced the Muse Spark, Voice Mode, Muse Image and Muse Video lines, so the move follows a product-building sprint rather than an inactive research period.
  • The widely repeated $100 million compensation figure is not verified for Yu personally; Meta has said such packages could include conditional stock, bonuses and performance terms.
  • The story fits a China-linked talent pattern: researchers with experience across Chinese and US institutions can treat frontier labs as stepping stones toward founding rather than final employers.
  • The cited count of hundreds of former Meta researchers comes from a third-party index and should not be treated as an audited attrition measure.
  • For AI executives, the implication is that retention increasingly depends on research autonomy, ownership and the chance to define a new platform, not only pay or compute.
Open at QbitAI Ask Claude ChatGPT

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

# AI QbitAI 14h

Delta Intelligence and Wuji Technology plan a shared data standard for whole-body dexterous manipulation, targeting a major weakness in humanoid-robot training data.

More
  • Most existing dexterity datasets record hand degrees of freedom while neglecting torso posture, arm trajectories and visual attention, making learned actions physically plausible only in a narrow sense.
  • The proposed standard would unify capture methods, formats, labels and evaluation, while a shared training baseline could make results reproducible across hardware and models.
  • The partnership links a humanoid foundation-model company with a dexterous-hand maker, closing the loop from wearable data capture to model training and edge deployment.
  • Standardization can reduce repeated integration work, but it also creates a strategic question over who controls the benchmark and the resulting data ecosystem.
  • The technical payoff would be strongest in fine assembly and service tasks where hand motion depends on coordinated balance, reach and viewpoint.
Open at QbitAI Ask Claude ChatGPT

China's best music in 2026 so far

# Culture ChinaTalk 10h

A China-focused music roundup spans funk, indie rock and experimental scenes, showing a local culture ecosystem that is more varied than the country's export-pop stereotype.

More
  • The selections include a Brooklyn-formed band with Chinese links, a Shenzhen emo group and musicians connected to the post-punk lineage around PK14.
  • The roundup points to a scene built through small labels, pay-what-you-want releases, podcasts and live performance rather than only through major streaming platforms.
  • One band reportedly drew police attention while livestreaming a Shanghai street performance, a small example of how public-space control can shape independent music culture.
  • The value for a technical reader is not a chart ranking but a map of how Chinese musicians form durable subcultures across cities, genres and overseas distribution channels.
Open at ChinaTalk Ask Claude ChatGPT

China is urged to keep its global AI body open to the US

# Policy SCMP 12h

Analysts warn that China's new World Artificial Intelligence Cooperation Organisation could lose members if it is seen as an anti-US bloc, revealing the diplomatic problem behind Beijing's AI governance push.

More
  • The organization is presented as a possible forum for international AI cooperation, but analysts caution against making it look like a natural opponent of the US-led Pax Silica partnership.
  • The warning reflects a broader Chinese strategy: create institutions that expand Beijing's role in rule-setting without forcing countries to choose a rigid geopolitical camp.
  • For governments and companies, the distinction matters because AI governance bodies can influence standards, data norms, safety language and access to technical cooperation.
  • The article does not establish the organization's membership or enforcement power, so its immediate influence remains uncertain.
  • China's challenge is credibility. A body framed as open and cooperative may attract states that want alternatives, while an explicitly anti-US posture could narrow that coalition.
Open at SCMP Ask Claude ChatGPT

China's Y-20B can reportedly become an aerial tanker

# Policy SCMP 10h

China has confirmed that its Y-20B transport can be converted into an aerial refueling tanker, potentially extending PLA air operations deeper into the western Pacific.

More
  • The Y-20B uses domestically developed WS-20 turbofans and reportedly carries pallet-ready fuel tanks and removable refueling equipment, allowing transport and tanker roles from one airframe.
  • A convertible design could make tanker capacity easier to scale and preserve flexibility when aircraft are needed for cargo missions.
  • The capability would address a long-standing operational constraint: fighters and bombers are limited by range and loiter time once they move beyond nearby bases.
  • The report comes from state broadcaster CCTV, so the public description may signal intended strategic messaging as much as current fleet readiness.
  • The real measure will be tanker numbers, crew training, refueling reliability and integration with long-range aircraft, not simply whether the conversion is technically possible.
Open at SCMP Ask Claude ChatGPT

Alibaba Cloud pitches a full data-to-model pipeline for self-driving

# Analysis GeekPark 10h

Alibaba Cloud argues that Chinese autonomous-driving competition has moved from algorithms to an integrated data, model and compute pipeline, with its data center processing up to millions of clips daily.

More
  • The proposed stack links multimodal data ingestion, cleaning, vectorization, automatic labeling, sample retrieval, model training and cloud compute rather than selling isolated GPU capacity.
  • Alibaba says the pipeline can process data at a million-clip scale per day and let models iterate in days; these are company claims, not an independent operational benchmark.
  • The engineering bottleneck is data preparation: expensive GPUs do little if vehicle data remains fragmented across manual transfer, labeling and validation steps.
  • Alibaba also positions its Qwen model as a possible base for driving VLA systems and its Wanxiang model as a generator of rare scenarios for training and evaluation.
  • The strategic contest is therefore over the production line that turns road experience into usable training examples, a capability that may survive model-architecture changes better than a particular dataset.
Open at GeekPark Ask Claude ChatGPT

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

# Policy Solidot 13h

Apple and Alibaba are reportedly preparing a model for China-market Apple Intelligence, showing how local regulation forces even a global device platform to share control of its AI stack.

More
  • The reported arrangement uses Alibaba as a Chinese model partner for Apple devices sold in China, while Apple's service awaits regulatory registration and rollout through an iOS update.
  • The model is not simply a localization choice: it is part of the approval, data and content-control path required to operate generative AI inside China's domestic internet environment.
  • The move gives Apple a way to preserve a coherent product experience while relying on a partner that can navigate local infrastructure and compliance requirements.
  • It also shows the strategic cost of China's market boundary. Apple can keep the hardware and interface, but the intelligence layer may differ materially from the global version.
  • The partnership's success will depend on latency, quality, privacy commitments and whether users perceive the China-specific system as a competitive feature rather than a constrained substitute.
Open at Solidot Ask Claude ChatGPT

Researchers report an unexpected structure inside nucleons

# Analysis Solidot 14h

Chinese and international researchers report that precise collision measurements require a nucleon structure carrying baryon number without electric or color charge, challenging a simple quark model.

More
  • The result was published in Science, according to the report, and comes from measurements of net baryon number and net electric charge in high-energy nuclear collisions.
  • The finding does not discard quantum chromodynamics by itself, but suggests that the intuitive picture of nucleons as only quarks, antiquarks and gluons may be incomplete.
  • A structure carrying baryon number without electric or color charge would force theorists to explain how such degrees of freedom emerge inside ordinary matter.
  • The work matters for fundamental physics rather than immediate engineering, but it is a substantive China-linked research result because Chinese scientists participated in a measurement that tests the composition of visible matter.
Open at Solidot Ask Claude ChatGPT

SHEIN's IPO is about more than raising cash

# Analysis 36Kr 15h

SHEIN generated about $6 billion in operating cash over three years and held $14.8 billion in cash, yet its IPO appears driven by convertible preferred-stock redemption pressure and slowing growth.

More
  • The company reportedly raised about $3.84 billion in equity over eleven years, making recent self-generated cash roughly 1.55 times its historical financing total.
  • The $17.294 billion of convertible redeemable preferred stock on its books carries a clock: investors' cash return rate rose from 8 to 12 percent, and redemption rights can revive if the IPO fails or stalls.
  • That makes listing a liability-management event as much as a funding event, despite the company's strong cash position and negative working-capital cycle.
  • The business is still growing operationally, with orders, active customers and contract manufacturers increasing, but revenue growth slowed sharply and US revenue reportedly fell in 2025 and early 2026.
  • The valuation history shows how private-market exuberance can outrun business growth: the D round's $98.2 billion pre-money valuation fell to $64 billion in the next round even as revenue rose.
  • SHEIN's China-linked significance is organizational rather than domestic-market focused: a China-founded, globally distributed supply-chain platform is approaching public markets under pressure from both geopolitical scrutiny and investor rights.
Open at 36Kr Ask Claude ChatGPT

China rejects New Zealand's espionage assessment as interference

# Policy SCMP 13h

Beijing rejected New Zealand's report naming China as the only country conducting espionage at scale, showing how security accusations are becoming another point of friction in the relationship.

More
  • China's embassy in Wellington called the assessment foreign interference and a Cold War mentality, while New Zealand's intelligence chief presented China as a large-scale espionage concern.
  • The dispute is not only about one report: it reflects competing definitions of legitimate intelligence activity, political influence and national-security evidence.
  • New Zealand's position shows how pressure from partners can narrow the space for a pragmatic China relationship even when trade and economic ties remain important.
  • Beijing's response follows a familiar diplomatic pattern of rejecting the premise and attributing the criticism to external geopolitical alignment rather than debating individual allegations.
  • The practical consequence is a higher compliance and political-risk burden for universities, technology firms and other institutions working across the two countries.
Open at SCMP Ask Claude ChatGPT

Microsoft shrinks its China business but keeps its engineering base

# Policy Solidot 16h

Microsoft has reportedly closed at least 15 China offices and joint ventures, yet remains because local firms need Western tools for overseas operations and Microsoft wants access to Chinese engineers.

More
  • The company faces a double squeeze: Chinese customers increasingly favor domestic software, while US export controls make cloud and AI expansion harder.
  • The remaining business is positioned around serving Chinese companies that operate internationally, rather than assuming that Microsoft's global cloud model can simply be transplanted into China.
  • Microsoft's statement that it operates under the same regulatory environment as international suppliers shows how foreign firms now describe compliance as a permanent operating condition, not a temporary obstacle.
  • The case illustrates a broader decoupling pattern: foreign technology companies may reduce domestic-facing products while preserving engineering, enterprise and cross-border support functions.
  • For technology executives, the relevant question is whether talent access and multinational customer demand can justify the cost of maintaining a politically constrained local footprint.
Open at Solidot Ask Claude ChatGPT

Moonshot's IPO needs a new story beyond model rankings

# Analysis 36Kr 16h

Moonshot AI is again linked to a possible Hong Kong IPO, but Chinese analysis says Kimi's technical lead must become a financing and capacity story before competitors erase it.

More
  • The company reportedly holds more than 10 billion yuan, roughly $1.4 billion, in cash and has previously said it was not under immediate pressure to list; the attraction is a recurring public-market funding channel.
  • Kimi K3 is described as near the top of Artificial Analysis rankings at lower cost than leading US models, giving Moonshot a temporary performance-and-efficiency narrative.
  • DeepSeek's much lower per-task cost threatens that narrative, illustrating how quickly a model advantage can be repriced when the underlying technology stack is broadly shared.
  • Kimi reportedly had to pause new consumer subscriptions when demand approached cluster capacity, exposing compute availability as a constraint on converting model quality into revenue.
  • The IPO window is therefore tied to product timing: listing while K3 looks frontier-adjacent could secure a premium, but a later filing may face a market that has already moved on.
  • The broader lesson is that China's model startups need capital not only for training but also for serving enough users to generate the data flywheel that sustains quality.
Open at 36Kr Ask Claude ChatGPT

US coastguards train Taiwan counterparts amid Chinese activity

# Policy SCMP 8h

US coastguards began a two-week training program with Taiwan, while Chinese military and maritime activity around the island gives the cooperation a wider escalation context.

More
  • The reported training covers small-boat operations and team coordination, placing the emphasis on coast-guard interoperability rather than an announced military alliance.
  • Taiwan's coast guard describes the activity as partnership-building; the surrounding Chinese presence makes even limited operational cooperation politically significant.
  • The episode shows how pressure around Taiwan is expanding into law-enforcement and maritime domains, where routine training can still alter crisis signaling.
  • The source does not establish whether the training is part of a larger permanent deployment or changes Taiwan's rules of engagement.
Open at SCMP Ask Claude ChatGPT

ByteDance bets that avoiding distillation is worth the delay

# Analysis 36Kr 18h

Chinese researchers debate ByteDance's decision not to distill outside models, framing it as a long-term bet on independent data and training pipelines at the cost of slower short-term progress.

More
  • Distillation lets a weaker model learn answers or trajectories from a stronger one, providing a cold start toward usable capability before a team builds its own data flywheel.
  • Researchers quoted in the article say refusing the shortcut can slow coding and agent progress, and ByteDance's Seed models reportedly lagged leading domestic systems on a web-development benchmark.
  • The company has a counterexample in video generation, where real data and product feedback may be more valuable than synthetic outputs from another model, helping its Seedance line reach the frontier without external distillation.
  • The strategic choice is therefore capability-specific rather than ideological: independence may be valuable where proprietary data is abundant, but expensive where a model needs long chains of successful tool-use trajectories.
  • ByteDance's creation of a dedicated AI data and safety organization suggests that an anti-distillation strategy requires replacing borrowed capability with much stronger sourcing, cleaning, evaluation and feedback infrastructure.
  • The bet only pays if the resulting data system produces differentiated models faster over several generations, not merely if the first model is more original.
Open at 36Kr Ask Claude ChatGPT

China wants Chancay Port on the APEC agenda

# Policy Jiemian 5h

Chinese coverage frames the China-Peru Chancay Port as a candidate for wider APEC connectivity, turning a bilateral infrastructure project into an instrument of regional agenda-setting.

More
  • The article argues that hosting APEC gives China an opportunity to shape discussion around trade corridors, logistics and economic integration rather than merely promote a completed bilateral project.
  • Chancay is presented as a link between China and Latin America, with its proposed role extending beyond port operations into regional economic governance.
  • The move reflects a broader Chinese diplomatic pattern: use a physical infrastructure asset as a concrete entry point for institutional and political influence.
  • The article also acknowledges the constraints, including fragmented trade governance, inadequate trans-Pacific links and greater geopolitical scrutiny of overseas Chinese projects.
  • Whether the port becomes a durable regional hub will depend on cargo economics, local political acceptance and the ability to attract partners beyond China's existing network.
Open at Jiemian Ask Claude ChatGPT

Alibaba's Lingxi Games may sell for more than $1.5 billion

# Analysis 36Kr 10h

Chinese analysis treats the reported Lingxi sale as a test of Alibaba's retreat from content, while US coverage foregrounds Trustar's possible $1.5 billion-plus valuation.

Also: Techmeme

More
  • The reported buyer is Trustar Capital, and the parties have not confirmed the deal; the Chinese article says the valuation could exceed $1.5 billion.
  • Lingxi's main cash engine is the long-running Three Kingdoms strategy game, which the article describes as stable but no longer a breakout-growth business.
  • The Chinese framing is that a mature hit can become a saleable cash-flow asset when a conglomerate changes priorities, even if the underlying game remains commercially healthy.
  • Alibaba's shift from ecosystem expansion toward AI and cloud gives the sale strategic meaning beyond gaming, especially after years of trying to build a broader entertainment footprint.
  • The case also illustrates a structural feature of China's game industry: development culture and durable live operations are difficult to acquire or replicate through conglomerate capital alone.
Open at 36Kr Ask Claude ChatGPT

Caption errors embarrass Chinese museums promoting Party history

# Culture SCMP 11h

Chinese museums have made basic caption errors in Communist Party-themed exhibits, exposing the tension between centralized historical messaging and uneven execution at local cultural institutions.

More
  • A Zhejiang museum reportedly rendered the Red Army as a red car in a Long March caption, an error especially embarrassing in an exhibition category tied closely to official political memory.
  • The mistakes matter because red tourism turns revolutionary history into a state-supported cultural and economic product, where signage is part of the authority being sold.
  • The episode suggests that political importance does not guarantee institutional quality: local museums still depend on ordinary editing, translation and review workflows.
  • Public criticism can target the museum's competence without directly contesting the historical narrative, offering a relatively safe way to express frustration with bureaucratic management.
Open at SCMP Ask Claude ChatGPT

Tencent moves a senior post-training researcher to WeChat AI

# Analysis 36Kr 9h

Tencent researcher Xu Can has moved from Hunyuan to WeChat's WeLM team as WeChat expands its native assistant and agent work, revealing a continuing internal contest over model ownership.

More
  • Xu's background in synthetic data, instruction construction, reinforcement-learning feedback and evaluation fits WeLM's reported push toward reasoning and tool use.
  • The move comes after Tencent reorganized model, data and infrastructure units under a more centralized Hunyuan leadership while WeChat retained a relatively independent research line.
  • WeChat is testing an assistant that can operate native functions and mini-programs, so the destination team is closer to a product surface with hundreds of millions of potential users than a standalone model lab.
  • The technical context is unusual: WeLM has explored sparse mixture-of-experts models and hidden internal decoding, while other WeChat systems reportedly use Hunyuan models in recommendation and advertising.
  • The personnel move therefore signals integration pressure inside Tencent, but not necessarily the end of parallel model stacks; product-specific data and latency needs can justify keeping them separate.
Open at 36Kr Ask Claude ChatGPT

Sanrio's China licensing engine meets Pop Mart competition

# Culture 36Kr 11h

Chinese analysis shows how Alibaba's licensing operation helped Sanrio scale in China, while the market increasingly compares Hello Kitty's authorization model with Pop Mart's direct retail machine.

More
  • Sanrio's China sales reportedly reached 31.44 billion yen, roughly $222 million, in the latest fiscal year, up 83.2 percent, with Kuromi ahead of Hello Kitty locally.
  • Alibaba's IP platform signed exclusive mainland rights for 26 characters and helped more than 200 brands launch tens of thousands of licensed products, according to the article.
  • The Chinese market is not merely a royalty stream: Sanrio has expanded physical stores, while Alibaba supplies localization, distribution and a large network of consumer brands.
  • Pop Mart is the useful comparison because it owns more of the retail relationship and turns scarcity and blind-box mechanics into a direct customer habit.
  • The strategic challenge is that rapid China growth creates a high base and may expose both companies to the limits of character-led consumption without deeper stories, games or film.
Open at 36Kr Ask Claude ChatGPT

Baidu says Kuku AI has more than 25 million office users

# AI QbitAI 12h

Baidu's Kuku AI launch combines a claimed 25 million office MAU with a standalone agent product for documents, finance and enterprise workflows.

Also: Leiphone

More
  • The product is built around Baidu's document and cloud-drive corpus, with the company saying its broader service has more than 100 million monthly users.
  • Its enterprise design combines public professional content, private company knowledge, shared context and auditable skills, aiming to make an agent produce business artifacts rather than just answers.
  • The reported user and evaluation rankings are company or sponsor claims, so the stronger signal is distribution through an existing document repository and office workflow.
  • Baidu is betting that specialist skills and persistent organizational memory can create switching costs that a generic chatbot cannot.
Open at QbitAI 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 3/15
Hacker News 200 3/20
V2EX 200 8/40 +29 over cap