China Desk

China News, Summarized 14 Aug 2026, 17:03 UTC 34 stories 17/19 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

Why Chinese platforms make ordinary speech hard to post

# China Digital Times 16h

A translated Chinese essay argues that blunt keyword filtering, platform liability and risk avoidance force users into coded language, revealing how censorship is experienced below the level of formal bans.

More
  • The essay distinguishes published legal prohibitions from the much larger set of terms platforms suppress because they fear penalties, complaints or uncontrolled discussion.
  • Users reportedly turn to homophones, initials and euphemisms to express ordinary feelings or social observations, making communication less precise even when the topic is not illegal.
  • The author blames platform governance and negligence as much as state rules, a framing that makes the problem easier to criticize while leaving the political structure largely implicit.
  • This is an important distinction for engineers: the visible moderation system is not simply a legal blocklist but a feedback loop of liability, automation, human review and institutional risk.
  • Overblocking can reduce trust and distort public sentiment because users learn to speak around the system rather than state what they believe.
  • The essay's continued availability also shows the unevenness of Chinese censorship: criticism of platform implementation may survive even when direct challenges to policy are constrained.
Open at China Digital Times Ask Claude ChatGPT

Former premier Zhu Rongji dies at 97

# SCMP 10h

Former Chinese premier Zhu Rongji has died at 97; Chinese retrospectives stress his anti-corruption image and reform legacy, while the deeper record includes the social costs of centralizing China's economy.

Also: Initium

More
  • Zhu led major reforms in taxation, state finance and state-owned enterprise restructuring during a period when Beijing was consolidating control over economic policy.
  • His public image rests on blunt anti-corruption rhetoric and a reputation for administrative toughness, but that image coexists with painful memories of layoffs and industrial restructuring.
  • The reforms helped create a more coherent central fiscal system and supported China's market transition, while also shifting costs onto workers, households and local governments.
  • His career illustrates a recurring Chinese governance pattern: reform is pursued through concentrated authority, with limited room for affected groups to negotiate the terms.
  • The political legacy is therefore double-edged. Zhu is remembered both as an engineer of modernization and as an emblem of the state-led discipline that accompanied it.
  • His death may prompt a controlled public reassessment of the reform era, especially as current policy again emphasizes central coordination and state capacity.
Open at SCMP Ask Claude ChatGPT

Alibaba Cloud pitches a full stack for China's self-driving race

# GeekPark 7h

Alibaba Cloud is positioning itself as a data-to-model-to-compute production line for Chinese autonomous-driving companies, arguing that data engineering now matters more than simply buying GPUs.

More
  • The company says its Ulanqab facilities support 60% of intelligent-driving compute represented at the forum, a promotional claim with no independent denominator.
  • Its proposed pipeline combines multimodal ingestion, automated cleaning, model-assisted labeling, vectorization and retrieval of mixed training samples, with daily processing in the million-clip range.
  • The architecture targets the real bottleneck in end-to-end and vision-language driving systems: turning messy vehicle data into repeatable training examples quickly enough to support daily iteration.
  • Alibaba also pitches Qwen as a possible foundation model and Wanxiang as a world-model generator for creating rare driving scenarios and evaluating responses.
  • The strategic shift is from selling isolated cloud capacity to owning the production loop, which could make data tooling and workflow integration a stronger moat than stored data alone.
  • The unresolved issue is customer independence: automakers may gain speed but become more dependent on one vendor's cloud, models and evaluation stack.
Open at GeekPark Ask Claude ChatGPT

China and Brazil plan an all-day disaster-warning satellite

# SCMP 3h

China and Brazil unveiled a satellite project to monitor South America around the clock, turning long-running space cooperation into a regional public-infrastructure and diplomatic asset.

More
  • Brazil's space research agency says the satellite will monitor the whole continent for weather and environmental conditions, according to the report.
  • The project extends cooperation that has lasted nearly four decades and gives Brazil a role in a system whose coverage serves the wider region.
  • For China, the value is not only launch or hardware export: shared environmental data creates a durable institutional relationship with South American agencies.
  • A regional warning system can become strategically influential if countries depend on its data for disaster planning, but the report does not give a launch date, payload specifications or funding split.
  • The project shows a softer form of China's space diplomacy, where practical public goods may build trust more effectively than prestige missions.
Open at SCMP Ask Claude ChatGPT

Qwen3.8-27B opens Alibaba's model weights for commercial use

# AI QbitAI 1h

Chinese coverage presents Qwen3.8-27B as a practical, locally deployable model; US aggregation reduces the story to Alibaba's open release and coding claims.

Also: Jiemian, IT Home, Techmeme

More
  • Alibaba says the 27B model is dense, multimodal and supports 262K native context, with YaRN extending the window to 1M tokens.
  • The model adds a reasoning-effort setting that lets users spend more or less computation per task, a useful control for cost-sensitive production systems.
  • The source says it outperforms Qwen3.7-Plus in coding and office workflows, but supplies no independent evaluation or hardware profile.
  • Apache 2.0 permits commercial deployment, making the release more consequential for enterprises that need model control or cannot send sensitive data to a hosted API.
  • Alibaba claims more than 460 released models, 3 billion downloads and 300,000 derivative models; these figures should be treated as marketing metrics.
  • The release reinforces China's strategy of competing through a large open ecosystem, where downstream fine-tuning and distribution may matter as much as the base model itself.
Open at QbitAI Ask Claude ChatGPT

GLM-5.3 takes aim at coding and open-model security

# AI QbitAI 8h

Chinese hands-on coverage says GLM-5.3 is closer to top proprietary coding models and found thousands of vulnerabilities, while the claims remain dependent on vendor tests and demos.

Also: Leiphone

More
  • The model's reported CyberGym white-box score rises to 84.5% from 77.2% for GLM-5.2, according to the source.
  • Zhipu says its red-team work found 2,404 deduplicated vulnerabilities, including 1,088 medium or high risk, across 220 projects.
  • The source's demonstrations emphasize long autonomous tasks: building a 3D scene, preserving spatial relationships and delivering a tested application with accounts, storage and permissions.
  • Those demos are more revealing than a static coding benchmark because they test planning, tool use, debugging and verification across many steps.
  • The central technical claim is that post-training and agent scaffolding, not a new base model, produced the jump; independent reproduction is still needed.
Open at QbitAI Ask Claude ChatGPT

DeepSeek Harness is a developer-focused agent saddle

# AI QbitAI 1d

Chinese testing finds DeepSeek Harness unusually strong at long-running, inspectable coding tasks, while also exposing rough interfaces and higher integration responsibility.

Also: GeekPark, Techmeme

More
  • The framework makes model, tools, policies, storage and context management independently replaceable, so a team can build a task-specific agent runtime rather than accept one vendor workflow.
  • Its trajectory view exposes event-level execution and token usage, giving developers a more useful debugging surface than a polished chat transcript.
  • Testers report that it asks clarifying questions and keeps iterating on complex tasks, but these results are anecdotal demonstrations rather than controlled comparisons.
  • The local-first design can help companies keep source code and logs in their own environment, though users must provide their own sandboxing, credentials and model governance.
  • DeepSeek's planned price changes make caching and token accounting part of the product architecture, not merely a billing detail.
  • The release is most relevant to engineers willing to assemble infrastructure; it is not yet a finished general-purpose desktop assistant.
Open at QbitAI Ask Claude ChatGPT

China closes an offshore-trust tax loophole

# Policy Pekingnology 14h

New Chinese rules tax offshore trusts at transfer, income and distribution stages; domestic commentary frames them as a fairness measure, while the fiscal backdrop points to stronger revenue collection.

More
  • The rules target structures used by Chinese tax residents to defer or avoid personal income tax, according to the analysis.
  • Tsinghua economist David Li argues the immediate impact on ordinary people is limited because few hold offshore trusts, while criticizing the higher top rate on labor income than on capital income.
  • The source estimates offshore trust assets at about 2.3 trillion yuan, roughly $324 billion, but labels the figure an industry estimate rather than an official total.
  • The move fits a broader push to expand tax visibility and reduce the ability of wealthy residents to separate assets from domestic tax obligations.
  • The policy matters beyond wealthy families: it signals that Beijing is willing to use financial reporting and tax rules to capture revenue as local fiscal pressure rises.
  • Implementation will determine the real effect, especially how authorities define residence, beneficial ownership, trust distributions and legacy structures.
Open at Pekingnology Ask Claude ChatGPT

Two Chinese robotics firms plan a whole-body dexterity data standard

# AI QbitAI 10h

Delta Intelligence and Wujin Technology are joining their humanoid-model and dexterous-hand businesses to standardize whole-body manipulation data, targeting a major bottleneck in embodied AI.

More
  • The proposed dataset will capture hand motion together with torso posture, arm trajectories and visual attention, rather than treating the hand as an isolated actuator.
  • The firms plan to publish a shared data standard and reproducible training baseline this year, which could reduce repeated format and evaluation work across robotics projects.
  • The hardware-to-model loop matters because a high-degree-of-freedom hand is useful only when demonstrations, model outputs and edge deployment share compatible representations.
  • The partnership is still an announced plan; no public dataset, benchmark result or robot task performance is provided.
  • If adopted, a common whole-body representation could make fine manipulation more transferable across hardware, but standardization will be difficult where companies see proprietary data as a moat.
Open at QbitAI Ask Claude ChatGPT

What China's AI talent market is really paying for

# Analysis Ginger River Review 8h

A Chinese recruiting discussion says million-yuan AI salaries exist but are not normal, and that companies increasingly pay for scarce data, systems and product skills rather than celebrity credentials alone.

More
  • The source draws on a Chinese business podcast and an AI recruiting firm that says it has served more than 1,500 companies across chips, infrastructure, models, robotics and agents.
  • A reported 3 million yuan salary for a new PhD is framed as exceptional; the market is more differentiated by skill scarcity, company stage and the ability to ship systems.
  • The most valuable profiles combine research with engineering execution, product integration, data pipelines or the ability to turn models into revenue.
  • The market is also spreading beyond model labs: chip firms, robot companies and conventional enterprises are competing for people who can operationalize AI.
  • This helps explain why headline compensation can mislead. A small star segment gets extreme packages while most hiring is shaped by funding, compute access and the maturity of a company's data flywheel.
  • The wider signal is that China's AI labor market is becoming a test of organizational capability: firms pay for people who can make scarce compute and data produce repeatable output.
Open at Ginger River Review Ask Claude ChatGPT

Chinese robot teams say compute demand has risen up to tenfold

# AI QbitAI 9h

A Chinese robotics company says its compute demand has grown five to ten times in two years, illustrating why embodied AI is pushing development infrastructure toward cloud workstations and shared simulation.

More
  • Robot development combines simulation, data collection, model training, physical deployment and repeated testing, so demand is not limited to a single training run.
  • The described cloud workstation packages GPU access with Isaac Sim, robotics software and 3D interaction, reducing the need for each engineer to maintain a separate environment.
  • The infrastructure problem is operational as much as computational: mismatched drivers, simulator versions, access controls and remote collaboration can consume more time than model experiments.
  • The source's promised 80% deployment-efficiency improvement is promotional and lacks a controlled baseline.
  • The strategic implication is that a cloud platform for embodied AI may sell reproducible environments and scheduling as much as it sells raw GPU hours.
Open at QbitAI Ask Claude ChatGPT

Apple and Alibaba train a China-specific model for iPhones

# Policy Solidot 9h

Apple is reportedly working with Alibaba on a model for the China market; Chinese coverage stresses regulatory access and local competition, while Western coverage would likely frame it as Apple's regional compromise.

More
  • Apple's China-specific generative AI plan reflects the need to operate through domestic partners and comply with local model registration and content rules.
  • The partnership could give Apple more control over the user experience than simply embedding a third-party chatbot, while still leaving model and data governance partly outside Apple's global stack.
  • The source says the service is expected to reach mainland iPhones through a future iOS update and that its generative AI service has been registered with China's cyberspace regulator.
  • This is a substantive China-operations story: Apple is adapting its architecture and partnership model to a market where Huawei and other domestic brands are gaining ground through AI features.
  • The unresolved questions are which model components Alibaba supplies, where inference runs, how user data is handled and what features regulators will permit.
Open at Solidot Ask Claude ChatGPT

Chinese team extends cold-atom entanglement to 420 kilometers

# Analysis SCMP 6h

A Chinese-led experiment entangled cold atoms across 420 kilometers, more than four times previous demonstrations and beyond the range where direct transmission faces fundamental limits.

More
  • The result was published in Physical Review Letters, according to the report, giving it a stronger evidentiary basis than a laboratory announcement alone.
  • The distance is significant because it moves entanglement experiments toward intercity-scale quantum networking rather than room-sized or campus demonstrations.
  • The work does not by itself create a usable quantum internet: practical networks still require repeaters, memory, synchronization, loss management and error handling.
  • China's advantage here is system-level experimental scale, where long baselines and controlled links matter as much as the underlying quantum state preparation.
  • Its importance will rise if the team demonstrates stable rates, fidelity and network protocols rather than only a record distance.
Open at SCMP Ask Claude ChatGPT

ChangXin's market value overtakes Tencent's

# Signal Solidot 7h

ChangXin Technology's roughly 3.58 trillion yuan, or about $504 billion, market value reportedly surpassed Tencent's, showing how China's AI boom is redirecting capital toward domestic memory chips.

More
  • ChangXin is described as China's leading DRAM producer and the world's fourth-largest DRAM supplier, with products used from phones to AI servers.
  • The comparison is a market signal, not proof that ChangXin has greater operating scale or profitability than Tencent; the source gives no earnings comparison.
  • The valuation reflects investor appetite for memory capacity as AI servers increase demand for high-bandwidth and conventional DRAM.
  • Tencent's falling value is attributed in part to concern over rising AI investment, illustrating how capital markets may punish software platforms for spending aggressively before returns are visible.
  • The reported figures are unusually volatile and the source mixes yuan and Hong Kong dollar comparisons, so the ranking should be checked against a consistent closing-price basis.
Open at Solidot Ask Claude ChatGPT

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

# Policy SCMP 6h

China confirmed that its WS-20-powered Y-20B transport can be reconfigured as an aerial refueling tanker, potentially extending PLA air operations beyond the first island chain.

More
  • The aircraft uses pallet-ready fuel tanks and removable refueling equipment, allowing transport and tanker missions without building entirely separate airframes.
  • The WS-20 turbofans indicate a shift toward domestically powered heavy aircraft, reducing dependence on foreign engine supply for a strategic platform.
  • A tanker version could extend fighter and bomber sortie ranges and improve support for operations over the western Pacific, though the source gives no tanker fuel capacity or fleet size.
  • The modular conversion reflects a broader Chinese preference for extracting more missions from a common airframe while industrial capacity catches up.
  • Operational significance will depend on training, escort, basing and the number of aircraft actually converted, not simply on documentary confirmation.
Open at SCMP Ask Claude ChatGPT

China rejects New Zealand's espionage assessment

# Policy SCMP 10h

China's embassy in New Zealand called Wellington's latest security report foreign interference and Cold War thinking, while New Zealand identified China as its only espionage threat operating at scale.

More
  • The dispute shows how security assessments now become diplomatic conflicts, with Beijing challenging not only the conclusions but the legitimacy of the reporting process.
  • New Zealand's framing places China at the center of its threat environment, raising pressure on trade and scientific ties even without a new criminal case.
  • China's response relies on familiar language about politicization and foreign influence, signaling that it will contest allied intelligence narratives as part of the broader relationship.
  • The episode is a useful indicator of how smaller US partners are being drawn into the wider China security framework.
Open at SCMP Ask Claude ChatGPT

Researchers report an unexpected structure inside nucleons

# Signal Solidot 11h

A China-linked collider experiment reports evidence for an internal nucleon structure carrying baryon number without electric or color charge, challenging the simplest quark model.

More
  • The result comes from precision measurements of net baryon number and net electric charge in high-energy nucleus collisions, according to the source.
  • The claimed structure would not fit the naive picture of nucleons as only quarks, antiquarks and gluons, though the interpretation requires confirmation and theoretical work.
  • The significance is foundational: nucleon structure underlies most visible matter, so a new conserved-charge carrier would affect how physicists describe the strong interaction.
  • The source gives no experiment name, uncertainty range or independent commentary, making the headline claim difficult to assess beyond its reported publication in Science.
Open at Solidot Ask Claude ChatGPT

RISC-V chip hopeful Eeasy seeks a Hong Kong listing after heavy losses

# Signal 36Kr 12h

Eeasy Computing is pursuing an IPO after losing about 4.9 billion yuan, or $690 million, in three years, putting China's state-backed push for domestic RISC-V chips under financial pressure.

More
  • The company was founded by former BOE chairman Wang Dongsheng and has raised about 9.28 billion yuan, roughly $1.31 billion, across four rounds while retaining only 188 million yuan in cash at the end of 2025.
  • Revenue grew from 1.75 billion yuan in 2023 to 2.43 billion yuan in 2025, but gross profit was insufficient to cover administrative and research costs.
  • The company reported a monthly cash burn near 96.8 million yuan, or about $13.6 million, and relied on loans and liquidating investments to rebuild cash in early 2026.
  • Its RISC-V share is modest even within the domestic market, illustrating how the label of local substitution can mask a fragmented and intensely competitive field.
  • Customer concentration is another risk: one unnamed customer supplied 82.1% of revenue in 2023, though that fell to 39.3% in the first quarter of 2026.
  • The listing will test whether strategic semiconductor financing can support a long hardware-software ramp before commercial scale and margins arrive.
Open at 36Kr Ask Claude ChatGPT

Meta AI researcher Yu Jiahui leaves to start a company

# Analysis QbitAI 2h

A Chinese researcher who helped build Meta's Muse model family is leaving after about a year to found an undisclosed venture, highlighting the limits of cash-heavy AI talent strategies.

More
  • Yu joined Meta in 2025 and is credited by the source with work across Muse Spark, voice, image and video systems before departing in August 2026.
  • The widely repeated $100 million compensation figure is unverified; the source notes that Meta described such packages as conditional combinations of stock, bonuses and tenure.
  • His career spans computer vision, Google multimodal research, OpenAI perception work and Meta foundation models, making him unusually positioned to turn frontier-lab experience into a startup.
  • The source cites a researcher directory showing hundreds of former Meta researchers, but that count is not an audited measure of attrition or of the lab's health.
  • The strategic signal is that top researchers may view a major lab's compute, salary and product access as a temporary platform rather than a durable destination.
Open at QbitAI Ask Claude ChatGPT

SHEIN's IPO is about capital structure, not cash

# Analysis 36Kr 12h

SHEIN has generated about $6 billion in operating cash over three years and held $14.8 billion in cash, yet its IPO pressure comes from $17.3 billion of redeemable preferred stock and looming investor rights.

More
  • The source estimates total equity funding over eleven years at about $3.84 billion, making recent self-generated cash roughly 1.55 times its historical financing.
  • The unusual constraint is contractual: some preferred investors' return rate rose from 8% to 12%, and redemption rights can revive if an offering fails or is delayed.
  • SHEIN's negative working-capital model lets customer prepayments fund operations, but slowing revenue growth and weaker US sales complicate the idea that cash generation alone solves its financing problem.
  • The company's valuation already fell from a reported $98.2 billion pre-money round to $64 billion, even as orders, customers and manufacturing partners grew.
  • That gap shows the market is repricing platform growth and regulatory exposure, not merely rewarding operational scale.
  • The IPO is therefore a liquidity and governance event: it can convert private claims into a public financing channel and reset the terms around a globally exposed Chinese-founded retailer.
Open at 36Kr Ask Claude ChatGPT

US coastguards train Taiwan amid increased Chinese activity

# Policy SCMP 4h

Chinese coverage reports US coastguard instructors training Taiwan on small-boat operations as Beijing increases activity east of the island, framing a law-enforcement exchange as part of a wider security signal.

More
  • The two-week program focuses on small-boat handling and team coordination, according to Taiwan's coast guard.
  • The training is not a formal military deployment, but it expands practical interoperability in a domain that can sit below the threshold of overt armed conflict.
  • Its timing matters because maritime encounters around Taiwan are increasingly managed through coastguard and other gray-zone forces rather than only naval units.
  • The source gives no details on the number of instructors, equipment transfer or future exercises, so the long-term operational effect remains uncertain.
Open at SCMP Ask Claude ChatGPT

China is urged not to turn AI governance into a bloc fight

# Policy SCMP 8h

Analysts warn that China's new World AI Cooperation Organisation could repel members if it is seen as an anti-US bloc, revealing the diplomatic constraint on Beijing's AI governance ambitions.

More
  • The organization is presented as a potential forum for international AI cooperation, not simply a technical standards body.
  • Analysts compare it with the US-backed Pax Silica partnership and argue that prospective members may avoid joining an institution defined by opposition to Washington.
  • China therefore faces a familiar governance problem: it wants to shape rules and reduce dependence on US technology without making participation look like geopolitical alignment.
  • The warning suggests that legitimacy will depend on practical benefits and inclusive procedures, not only on China's ability to convene countries.
Open at SCMP Ask Claude ChatGPT

ByteDance bets that avoiding distillation is worth the delay

# Analysis 36Kr 15h

ByteDance is reportedly refusing to distill other models into its core language systems, choosing a slower independent training path even as Chinese rivals use distillation to reach useful coding and agent capability faster.

More
  • Distillation lets a weaker model learn answers or task trajectories from a stronger teacher, often accelerating the cold-start phase of training.
  • Researchers quoted in the analysis say avoiding it can slow coding and agent progress because the model enters real use later and therefore collects less feedback data.
  • ByteDance's video-generation systems are presented as an exception: proprietary real-world data and the different failure modes of video may make external distillation less valuable there.
  • The company has reportedly created a dedicated AI data and safety organization to support sourcing, synthetic data, cleaning and evaluation for its independent training pipeline.
  • The strategic trade-off is between intellectual and data independence on one side and the compounding advantage of early product usage on the other.
  • The policy and supply-chain context matters too: a self-built training stack may be slower, but it reduces reliance on foreign models and contested data sources.
Open at 36Kr Ask Claude ChatGPT

Alibaba's Lingxi Games sale exposes China's maturing game market

# Analysis 36Kr 7h

Chinese analysis treats the reported $1.5 billion-plus Lingxi sale as a test of Alibaba's retreat from empire building, while US coverage focuses on the transaction itself.

Also: Techmeme

More
  • The business is described as a mature cash-flow asset built around a long-running strategy game with more than 100 million registered users, rather than a fast-growing studio.
  • Alibaba reportedly acquired its predecessor for about 1 billion yuan and later moved the unit's reporting line toward the group CFO, a sign of portfolio discipline.
  • The source says the earlier rumored price was 7 to 9 billion yuan, roughly $1.0-1.3 billion, below the newer reported valuation above $1.5 billion.
  • The unit's dependence on one major game illustrates why a stable publisher can still become saleable: private equity may value predictable cash flow while a platform company wants to fund AI and cloud.
  • The episode marks a reversal of the period when Chinese internet firms bought entertainment assets to capture more user time.
Open at 36Kr Ask Claude ChatGPT

Moonshot needs a new IPO story after Kimi K3

# Analysis 36Kr 12h

Kimi K3 gives Moonshot AI a possible frontier-performance narrative for a Hong Kong IPO, but the company's earlier capacity limits show how compute supply can undermine model momentum.

More
  • The source says Kimi K3 ranked near the top of Artificial Analysis while costing less than leading proprietary models, though rankings and prices can change quickly.
  • Moonshot reportedly paused new consumer subscriptions after demand approached cluster capacity, exposing the infrastructure cost of converting model quality into usage.
  • DeepSeek's newer model, with a much lower reported task cost, shows how quickly a valuation narrative based on both frontier performance and efficiency can be challenged.
  • An IPO would provide more than one financing event: it could create a continuing capital channel for compute, hiring and model deployment.
  • The timing matters because model leadership is temporary, and public markets may reward the company that appears closest to the frontier when it lists rather than the one with the best long-term architecture.
  • Moonshot will need to explain durable distribution, capacity and enterprise economics rather than rely on one benchmark window.
Open at 36Kr Ask Claude ChatGPT

The best Chinese music released in 2026 so far

# Culture ChinaTalk 6h

A China-focused music guide highlights funk, indie rock, post-punk and experimental releases, offering a window into a lively scene that rarely reaches mainstream English-language technology coverage.

More
  • The selection spans Shenzhen indie rock, Shanghai street performance, Sichuan recording sessions and China's avant-garde circuit.
  • The guide describes artists negotiating between local scenes and global genres rather than simply imitating Western pop formats.
  • Its account of a band being stopped while livestreaming on a Shanghai street also hints at the practical limits on public performance and online visibility.
  • This is cultural discovery rather than a market report, but it helps explain the creative ecosystems surrounding China's urban youth.
Open at ChinaTalk Ask Claude ChatGPT

Sanrio's China revival shows how local licensing works

# Culture 36Kr 7h

Chinese analysis casts Alibaba's AliFish as the local operator behind Sanrio's China growth, while contrasting its licensing model with Pop Mart's retail-heavy approach.

More
  • Sanrio's China sales reportedly reached 31.44 billion yen, up 83.2%, with Kuromi overtaking Hello Kitty locally; the figures come from company reporting summarized by the source.
  • AliFish reportedly connected Sanrio's characters to more than 200 Chinese brand partners and 30,000 product SKUs in 2024, supplying local channels and commercial adaptation.
  • The case shows why global character businesses often need a domestic operating layer in China rather than simply importing an IP and collecting royalties.
  • Sanrio remains relatively asset-light through licensing, while Pop Mart controls more of its retail network; their collaboration shows that licensing and direct retail can reinforce rather than exclude each other.
  • The growth rate is unlikely to persist indefinitely, and the source notes that Sanrio itself warns its China strategy may not succeed.
Open at 36Kr Ask Claude ChatGPT

China wants Chancay Port on the APEC agenda

# Policy Jiemian 1h

Chinese coverage frames Peru's Chancay Port as a platform for regional integration, not just a bilateral project, revealing how Beijing plans to use its APEC host year to shape connectivity rules.

More
  • The article argues that hosting APEC gives China agenda-setting power and an opportunity to present Chancay as a trade and logistics bridge between Asia and Latin America.
  • The port is therefore being positioned as institutional diplomacy as much as infrastructure: China wants a project it built to become part of a wider regional governance conversation.
  • The framing responds to fragmented trade rules, geopolitical scrutiny of overseas projects and weak trans-Pacific connectivity, according to the article.
  • Whether the initiative matters beyond symbolism will depend on whether other APEC economies treat the port as a shared logistics asset rather than a Chinese strategic outpost.
Open at Jiemian Ask Claude ChatGPT

Xiaomi camera mistakes the sun for the moon

# Signal Solidot 7h

A Xiaomi phone reportedly added lunar-looking ridges and craters to a solar-eclipse photo, a vivid example of computational photography crossing from enhancement into fabricated evidence.

More
  • The reported image suggests the camera's AI reconstruction inferred a familiar moon texture where the captured scene contained the sun, creating details that were not optically recorded.
  • The incident is not proof of the exact pipeline or model used, but it exposes the trust problem of semantic image enhancement in unusual scenes.
  • Similar allegations involving other handset makers show that this is a category-level product risk, not necessarily a Xiaomi-only defect.
  • For engineers, the missing control is provenance: users need a way to distinguish sensor data from generative or template-based reconstruction, especially in scientific or evidentiary images.
Open at Solidot Ask Claude ChatGPT

Museum caption errors embarrass China's revolutionary-history displays

# Culture SCMP 7h

Chinese museums have drawn criticism for basic caption errors in Party-history exhibits, revealing the gap between official historical importance and the quality-control burden of mass cultural promotion.

More
  • One Zhejiang caption reportedly turned the Red Army into a red car, a mistake especially damaging in an exhibition tied to the Long March.
  • The errors appear during a push to expand revolutionary or red tourism, where museums are expected to serve both education and local economic development.
  • The episode shows that ideological importance does not automatically produce reliable editorial or translation workflows at the local level.
  • Public criticism is possible because the errors are framed as administrative quality failures rather than as a challenge to the historical narrative itself.
Open at SCMP Ask Claude ChatGPT

Microsoft is shrinking in China while serving firms that face westward

# Policy Solidot 13h

Microsoft has closed at least 15 Chinese branches and joint ventures but is retaining operations to serve firms such as ByteDance that need Western technology for overseas business.

More
  • The company faces two opposing forces: Chinese customers increasingly prefer domestic software, while US export controls constrain cloud and AI offerings.
  • Its remaining value includes access to China's engineering talent and multinational customers that still need globally compatible management and developer tools.
  • The source says Microsoft describes its business as operating under the same regulatory environment as other international suppliers and remains committed to China.
  • This is a substantive China-operations story because the market is becoming less attractive for local cloud growth while remaining strategically useful as a talent and international-business base.
  • The pattern suggests partial retreat rather than exit: reduce exposed consumer and infrastructure bets, preserve relationships and serve cross-border enterprise needs.
Open at Solidot Ask Claude ChatGPT

Tencent moves a senior Hunyuan researcher to WeLM

# AI 36Kr 6h

Tencent has moved post-training researcher Xu Can from Hunyuan to the WeLM team as WeChat accelerates its own models and agent capabilities, signaling tighter competition between Tencent AI organizations.

More
  • Xu is associated with synthetic data, instruction construction, reinforcement-learning feedback and automated evaluation, all directly relevant to improving agent behavior after pretraining.
  • WeChat is testing an assistant that can operate native functions and mini-programs, giving WeLM a product surface where model reliability can be measured through real task completion.
  • The source describes WeLM as a separate research line within Tencent, even as the company consolidates infrastructure, data and foundation-model work under Hunyuan leadership.
  • Recent WeLM systems reportedly use sparse mixture-of-experts designs and additional hidden computation per token, indicating a focus on efficient inference as well as scale.
  • The personnel move is therefore less about a single hire than about where Tencent believes product-specific model ownership should sit: inside the platform with the user loop, not only in a central model lab.
Open at 36Kr Ask Claude ChatGPT

3D compute-memory chip startup raises more than $2.8 billion

# AI Leiphone 12h

Chinese 3D compute-memory chip company Qianhe Yibang says it raised more than 20 billion yuan, roughly $2.8 billion, to commercialize stacked-memory architectures aimed at the memory wall.

More
  • The company combines 3D DRAM stacking with compute-in-memory to reduce the cost of moving data between processors and memory.
  • Its backers include state funds, China Mobile's industry fund and private investors, giving the round both strategic and financial support.
  • The company was incubated by NetEase and says its team has worked on 3D DRAM compute-memory chips since 2018, but the source provides no product specifications, yield data or customer revenue.
  • The financing reflects China's willingness to fund nonstandard architectures that could reduce dependence on conventional memory and interconnect supply chains.
Open at Leiphone Ask Claude ChatGPT

Hong Kong activists hold a cross-border animal-rights vigil

# Culture Initium 18h

An Initium feature follows Hong Kong supporters commemorating an animal in mainland China, using a cross-border vigil to explore the invisible limits on mainland-facing activism.

More
  • The supplied text is mostly site code and does not provide enough reporting to establish the event's participants, location or official response.
  • The headline suggests a cultural and civil-society story about how concern for animals can travel across the Hong Kong-mainland boundary even when political activism faces sharper constraints.
Open at Initium 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 9/10
36Kr 200 9/30
Solidot 200 13/20
Leiphone 200 7/20
GeekPark 200 4/30
IT Home 200 12/60 +40 over cap
SCMP 200 10/50 +5 over cap
Initium 200 5/15
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 ERR 0/0 ConnectionError: ('Connection aborted.', ConnectionResetError(104, 'Connection reset by peer'))
V2EX 200 8/50 +39 over cap