China Desk

China News, Summarized 14 Aug 2026 34 stories archived day

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 closes an offshore-trust tax loophole

# Pekingnology

New Chinese rules tax offshore trusts at transfer, income and distribution stages, tightening enforcement as fiscal pressure makes wealthy taxpayers more valuable.

More
  • The Ministry of Finance and tax authority rules target offshore trusts used by Chinese tax residents to defer or reduce personal income tax; the supplied account says domestic offshore-trust assets are estimated at 2.3 trillion yuan (roughly $324 billion).
  • The policy is designed to tax the lifecycle of a structure rather than wait for a final distribution, reducing the room for trusts to function as indefinite deferral vehicles.
  • A Tsinghua finance professor argues the move improves fairness but also notes that labor income can face a top rate of 45 percent while capital income tops out at 20 percent.
  • That framing reveals the political tradeoff: Beijing wants more compliance from wealthy households without necessarily making capital income tax rates match labor income.
  • The measure matters beyond the wealthy because it signals a broader shift toward taxing cross-border assets and using information and reporting rules to strengthen revenue collection.
Open at Pekingnology Ask Claude ChatGPT

GLM-5.3 targets coding and security through post-training

# Leiphone

Chinese lab Zhipu says GLM-5.3 keeps the same base model but sharply improves coding and vulnerability discovery through post-training, including a reported 2,436 flaws across 269 projects.

More
  • Zhipu attributes the gains to post-training methods and frameworks including IndexShare, the SAO reinforcement-learning algorithm and its Slime training system, rather than a new base architecture.
  • The company reports a Terminal-Bench 3.0 score rising from 4.6 to 28.3 and a DeepSWE score rising from 46.2 to 66.9, while claiming a lead on several agent and workplace evaluations.
  • Its internal coding evaluation places the model in realistic development environments with code, documents, compute and storage, testing sustained tool use rather than isolated code completion.
  • Zhipu says joint security work found 2,436 vulnerabilities in 269 open-source and commercial projects, including 1,097 medium- or high-severity flaws; these are vendor-reported results without disclosed methodology in the item.
  • The safety design uses layered filtering, an intent-aware reasoning monitor and model-level alignment, a particularly important choice for downloadable weights where hosted-service controls disappear.
  • The release suggests Chinese labs are shifting competition from parameter count to post-training, tool use and secure agent behavior.
  • The open-model claims should be checked against reproducible benchmarks and real repository fixes before treating them as evidence of a durable lead.
Open at Leiphone Ask Claude ChatGPT

Former Chinese premier Zhu Rongji dies at 97

# SCMP

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

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

Apple and Alibaba train a China-specific AI model

# Solidot

Apple is reportedly working with Alibaba on a model tailored for China, while Chinese coverage stresses regulatory approval and local competition as prerequisites for Apple Intelligence on domestic iPhones.

More
  • The model would support Apple’s generative-AI features in China, where Apple cannot simply deploy the same service stack used elsewhere.
  • Working with Alibaba gives Apple access to a domestic model and cloud ecosystem while allowing the iPhone maker to retain control over the product experience and device integration.
  • Chinese regulators have reportedly registered Apple’s generative-AI service, making approval and local data governance part of the technical launch path rather than an afterthought.
  • The partnership shows how China’s AI rules reshape global product architecture: foreign companies may need local training, local partners and locally approved service layers to ship the same category of feature.
  • Apple faces stronger competition from Chinese phone makers that have already marketed AI features, so the project is also a response to a changing domestic hardware market.
  • The report does not specify the model’s size, training data, hosting arrangement, division of responsibility or launch date, and Apple and Alibaba have not provided a full technical account.
  • If launched, the arrangement could become a template for other foreign AI products in China; if delayed, it would highlight the operational cost of regulatory localization.
Open at Solidot Ask Claude ChatGPT

Scientists report an unexpected structure inside nucleons

# Solidot

A China-linked collider result challenges the simple quark-and-gluon picture by indicating a structure that carries baryon number without electric charge or color charge.

More
  • The researchers inferred the structure from precise measurements of net baryon number and net electric charge in high-energy nucleus collisions, according to the report.
  • The finding does not mean the standard model has been overturned; it means the simplified picture used to describe nucleon interiors may be incomplete.
  • A baryon-number-carrying, electrically neutral and color-neutral component would force theorists to revisit how composite structure emerges from quarks, antiquarks and gluons.
  • The result is important for China’s research system because it shows domestic and international teams contributing to a foundational physics question through large experimental infrastructure.
  • Independent replication and theoretical reconciliation will determine whether this becomes a durable new description or an anomaly in the measurement or interpretation.
Open at Solidot Ask Claude ChatGPT

Alibaba may sell Lingxi in a $1.5 billion gaming deal

# 36Kr

Chinese reporting says private-equity firm Centurium is nearing a deal to buy Alibaba’s Lingxi gaming unit for more than $1.5 billion, underscoring the retreat from broad internet conglomerates.

Also: Techmeme

More
  • Earlier reports valued the business at 7 to 9 billion yuan (about $1.0 billion to $1.3 billion), so the newer figure would represent a substantial upward revision if confirmed.
  • Lingxi’s main asset is the long-running strategy game Three Kingdoms Strategy Edition, which reportedly has more than 100 million registered users and generates stable cash flow years after launch.
  • The business has not produced a second hit of comparable scale, making it look more like a mature cash-flow asset than a growth platform within Alibaba’s portfolio.
  • The sale fits Alibaba’s broader shift toward AI and cloud while noncore assets such as retail and entertainment businesses face more explicit capital-allocation scrutiny.
  • It also illustrates why game development is hard to buy into: distribution and traffic can be acquired, but creative teams, live operations and a repeatable hit-making system are difficult to integrate.
  • For employees, a sale can leave the product and operating culture intact while changing ownership and investment horizons, a common pattern in China’s gaming industry.
  • The transaction is not confirmed by the parties, so valuation and buyer identity remain subject to change.
Open at 36Kr Ask Claude ChatGPT

BOE veteran’s second chip venture heads for Hong Kong IPO

# 36Kr

Display-industry veteran Wang Dongsheng is taking Eesway Computing public after three years of losses totaling 4.9 billion yuan, roughly $690 million, to fund a RISC-V alternative bet.

More
  • The company raised 9.276 billion yuan, roughly $1.31 billion, across four rounds but ended 2025 with only 188 million yuan, about $26 million, in cash and a reported monthly burn near 96.8 million yuan.
  • Revenue grew from 1.752 billion yuan to 2.431 billion yuan between 2023 and 2025, yet 2025 gross profit of 453 million yuan did not cover 604 million yuan of administrative expense, let alone 1.042 billion yuan of research spending.
  • Its RISC-V market position is small in the overall market and becomes a domestic-leader claim only after foreign suppliers are excluded, illustrating how Chinese semiconductor rankings often depend on the definition of local content.
  • One customer supplied 82.1% of 2023 revenue, falling to 39.3% in the first quarter of 2026; the unidentified customer appears likely to be BOE, but the filing does not name it.
  • Cash recovered in early 2026 mainly through borrowing, asset sales and collections rather than a demonstrated operating turnaround.
  • The company is making a long-horizon architecture bet while its current economics remain fragile, creating a direct test of how far state and strategic capital will fund domestic substitution.
  • The IPO will need to persuade investors that RISC-V control and future scale can overcome concentrated customers, continuing losses and a thin gross-margin base.
Open at 36Kr Ask Claude ChatGPT

China names 75 apps for illegal personal-data practices

# IT Home

A joint Chinese enforcement campaign identified 75 apps and mini-programs for privacy-policy, consent, third-party sharing and account-deletion violations.

More
  • The inspection covers mobile apps and mini-programs, with failures including inadequate first-run disclosure, incomplete descriptions of embedded third-party code and missing consent.
  • The listed problems also include collecting data before permission, failing to offer effective deletion or account cancellation, and sharing personal information without required disclosure.
  • The action is part of a coordinated campaign involving cyberspace, industry, police and technical-response authorities, showing that app privacy enforcement is becoming routine multi-agency work.
  • For product teams, the risk is not limited to a privacy statement: permissions, SDK inventories, data recipients and exit mechanisms all become enforcement surfaces.
  • The breadth of the list suggests a compliance regime focused on operational implementation rather than only major data breaches.
  • The strategic implication is that Chinese consumer software must treat privacy controls as a launch and distribution requirement, not a later legal layer.
Open at IT Home Ask Claude ChatGPT

Horizon Starry wins a flagship Chinese EV design slot

# IT Home

Horizon says its Starry chip will power a premium cockpit project, putting a domestically designed cockpit-and-driving platform into another major EV program.

More
  • The single-chip design pools compute across cockpit and driving functions instead of maintaining separate hardware domains, with shared data paths and a unified vehicle-computing base.
  • Horizon claims roughly 50% less in-vehicle hardware space, a per-vehicle hardware saving of 1,500-4,000 yuan (about $210-560), and development time falling from 18 to 8 months; these are vendor claims, not independently verified results.
  • The significance is architectural as much as commercial: Chinese automakers are testing whether centralized, domestically sourced compute can reduce integration work while supporting local multimodal perception and on-device model inference.
  • The project follows an earlier iCar launch partnership, suggesting Starry is moving from a showcase release into a broader vehicle-platform sales process.
  • What matters next is production validation: sustained automotive-grade reliability, software isolation between safety-critical driving functions and consumer AI, and whether the claimed schedule savings survive mass deployment.
Open at IT Home Ask Claude ChatGPT

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

# IT Home

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

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

US imposes tariffs of up to 100% on drones and components

# Techmeme

The White House is targeting drones and components tied to Chinese supply chains, making a previously commercial hardware dependency a formal trade and security-policy issue.

More
  • The tariff package includes a 100% levy on models deemed particularly sensitive, according to the Techmeme item summarizing Financial Times reporting.
  • The stated goal is to reduce US reliance on Chinese drone technology, so the measure affects both finished aircraft and the component ecosystem behind them.
  • For Chinese manufacturers, the immediate issue is market access; for US buyers, the strategic tradeoff is between supply-chain diversification and higher hardware costs.
  • The policy could accelerate alternative manufacturing in sensors, flight controllers, batteries and software, but tariffs alone do not recreate Chinese production scale or integration capability.
  • Because this item is an aggregator summary, the exact tariff schedule, product definitions and implementation date require the underlying government notice.
Open at Techmeme Ask Claude ChatGPT

Qwen 3.8 27B opens under Apache 2.0 for local deployment

# AI QbitAI

Chinese coverage stresses that Qwen 3.8 27B is freely deployable on consumer GPUs, while US discussion focuses on whether local hardware and software stacks can use it effectively.

Also: Leiphone, Hacker News

More
  • The model is a dense 27-billion-parameter multimodal system with a 262,000-token native context, extendable to 1 million tokens, and a reasoning-effort control for resource management.
  • Alibaba released the weights under Apache 2.0, removing a major commercial barrier for local deployment, fine-tuning, and derivative applications.
  • The claimed global Qwen download total exceeds 3 billion with more than 460 released models and 300,000 derivatives, but those are ecosystem metrics rather than direct evidence of production use.
Open at QbitAI Ask Claude ChatGPT

GLM-5.3 claims the open-source coding lead

# AI QbitAI

Chinese coverage highlights GLM-5.3’s coding and vulnerability-finding claims, while US discussion focuses on its narrow CyberGym edge over Anthropic’s restricted model.

Also: Techmeme

More
  • Z.ai reports an 84.5 percent CyberGym score, up from 77.2 percent for GLM-5.2 and slightly above Mythos 5; the figures are vendor claims awaiting broader replication.
  • The model reportedly found 2,404 deduplicated vulnerabilities across 220 projects, including 1,088 medium- or high-risk cases, with some legacy bugs dating back decades.
  • Its coding gains come from post-training in realistic long-horizon environments, where the model must inspect code, use tools, run tests and close a task rather than answer a static question.
  • The model weights are scheduled to open after safety evaluation and hardening, while the most sensitive security functions are restricted to verified users.
  • The release is a notable Chinese attempt to turn open source, cybersecurity and agentic coding into one competitive narrative.
Open at QbitAI Ask Claude ChatGPT

Chinese robotics teams are consuming ten times more compute

# AI QbitAI

A Chinese robotics firm says its compute demand has risen five- to tenfold in two years, while cloud workstations are becoming infrastructure for simulation, data and deployment.

More
  • The bottleneck in embodied AI is not only model size: teams must coordinate teleoperation data, simulation, training, hardware testing and multi-user development environments.
  • The company uses a cloud workstation with Isaac Sim and related tools so teams can obtain standardized GPU environments without buying idle local machines.
  • The technical pain is systems integration: CUDA, ROS2, simulators, drivers and data permissions create version and access problems that multiply as teams grow.
  • The cloud approach turns the workstation into a reproducible environment that can scale by task, which is more valuable for robotics than simply renting an isolated GPU.
  • The demand figures and claimed deployment improvements come from the company and cloud provider, but they capture a genuine shift from laboratory experiments to continuous engineering pipelines.
Open at QbitAI Ask Claude ChatGPT

DeepSeek Harness is an agent toolkit, not a finished product

# AI GeekPark

DeepSeek Harness passed 50,000 GitHub stars in 12 hours; Chinese testing stresses its plugin architecture, local execution and unusually low cost, while warning that it is still developer-only roughware.

Also: Techmeme, V2EX

More
  • The system runs locally through a Node command, with the browser acting as a shell over local sessions, logs and data rather than as a hosted agent service.
  • Its deepest architectural choice is that the model, tools, interface, approval rules and main execution loop can all be replaced, making the Harness closer to an operating system than a fixed assistant.
  • The tool can reportedly complete multi-step website and API tasks, but the test also found rough interactions and a poor fit for non-programmers.
  • A plugin ecosystem appeared almost immediately, including design and interface extensions, supporting the thesis that agent value may come from composition rather than one monolithic model.
  • The reported token cost was below three dollars for the test tasks, but that figure is a local experiment and says little about reliability, review time or total engineering cost.
  • The project’s growth rate is a distribution signal, not proof of a stable platform; the interfaces and core plugins are explicitly expected to change rapidly.
  • If DeepSeek maintains compatibility while others build plugins, Harness could become a control layer for interchangeable models and tools; if not, the early ecosystem may fragment.
Open at GeekPark Ask Claude ChatGPT

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

# AI QbitAI

A Chinese humanoid-model company and dexterous-hand maker are defining data standards that include torso, arm and gaze coordination rather than hand motion alone.

More
  • The partners say current manipulation datasets often record only hand degrees of freedom, making it difficult for models to learn stable whole-body coordination.
  • Their proposed standard will align capture formats, labels and evaluation, with a training baseline intended to make results reproducible across hardware.
  • The technical importance is representational: a hand action is conditioned by posture, reach, visual attention and contact forces, so isolated hand trajectories omit the control problem that matters in the real world.
  • The companies plan to validate the loop from data gloves to model training to deployment on direct-drive dexterous hands, targeting tasks such as connectors and threaded fasteners.
  • A shared baseline could reduce repeated integration work across Chinese robotics teams, although the announcement supplies no released dataset or independent performance result.
Open at QbitAI Ask Claude ChatGPT

Why Chinese users feel unable to speak plainly online

# Culture China Digital Times

Chinese coverage argues that blunt platform filters, not only explicit state bans, force users into coded language and make ordinary online speech hard to publish.

More
  • The translated essay says users routinely replace ordinary words with homophones, initials and coded terms after posts or comments are silently blocked by automated moderation.
  • It attributes much of the overblocking to platforms managing legal and political liability, rather than to a publicly enumerated list of forbidden terms.
  • That distinction matters: companies have some discretion in how they meet broad content obligations, so the everyday experience of censorship is shaped by risk-averse product and moderation choices.
  • The result is an online language layer optimized for filter avoidance, which reduces clarity and makes social discussion harder even when the underlying topic is not illegal.
  • The essay itself remains online and frames the problem as platform negligence, a depoliticized argument that may be more publishable than a direct challenge to the censorship system.
Open at China Digital Times Ask Claude ChatGPT

What China’s AI talent market is actually paying for

# Analysis Ginger River Review

Chinese recruitment discussion suggests million-yuan AI salaries are real but rare, with premiums going to people who can connect research, data, infrastructure and production rather than merely hold elite credentials.

More
  • The report distinguishes headline compensation from the broader market: a 3 million yuan package (about $420,000) for a new PhD exists, but is not normal for most researchers.
  • Recruiters say the scarce talent is increasingly hybrid, combining model research with post-training, data engineering, infrastructure, embodied systems or the ability to ship a product.
  • That preference reflects a maturing market. When many companies can train or fine-tune a model, the bottleneck moves to evaluation, deployment, reliability and integration with a real business.
  • China’s AI hiring market also shows a split between a small star layer commanding extraordinary packages and a much larger engineering workforce whose wages are less spectacular.
  • The podcast framing matters: Chinese business audiences are asking not only who is paid the most, but what kinds of knowledge create organizational leverage.
  • The source is a conversation with a specialist recruiter, not a systematic salary survey, so it is best used to understand employer priorities rather than precise compensation distributions.
  • The long-term implication is that Chinese firms may become less willing to pay purely for academic prestige unless candidates can convert research advantage into repeatable systems and revenue.
Open at Ginger River Review Ask Claude ChatGPT

Chinese 3D compute-in-memory chip firm raises $282 million

# AI Leiphone

NetEase-incubated Qianhe Yibang raised more than 2 billion yuan, roughly $282 million, to commercialize 3D DRAM compute-in-memory chips aimed at the memory wall.

More
  • The company combines stacked DRAM with compute in three dimensions, attempting to reduce the data movement imposed by conventional von Neumann separation of memory and processing.
  • Its pitch is system-level co-design: chip architecture, software stack and cloud-gaming or application-acceleration workloads are developed together rather than treating memory as a standalone component.
  • The financing includes state and industrial investors, including a national fund and a China Mobile-linked fund, suggesting strategic demand and capital support are part of the commercialization plan.
  • The company traces its work to a 2018 exploration of 3D DRAM compute-in-memory and says the new venture was established in 2024.
  • The announcement gives no silicon benchmarks, manufacturing node, yield, customer revenue or comparison with high-bandwidth memory, so the architecture remains an industrial bet rather than a demonstrated AI accelerator.
Open at Leiphone Ask Claude ChatGPT

Two Chinese Fields Medalists trigger pride and debate at home

# Culture China Digital Times

The first two Chinese nationals to win Fields Medals prompted online national pride but also debate over academic support, institutional credit and whether Chinese researchers feel discouraged.

More
  • The report describes Wang Hong’s work in harmonic analysis and geometric measure theory and Deng Yu’s work on partial differential equations and Hilbert’s sixth problem, but the social response is the more distinctive story.
  • Chinese universities publicly claimed symbolic credit for the winners as alumni, while online discussion questioned whether domestic institutions provide enough encouragement and freedom for emerging mathematicians.
  • The episode exposes a tension in China’s science system: national investment and elite university branding can coexist with concern that talent still needs to develop or work abroad.
  • The source is an account of online and institutional reaction, not a survey of Chinese academics, so it should not be generalized to the entire research community.
  • For technology leaders, the signal is about talent policy and prestige: international prizes become evidence of national progress, but also prompts to examine where high-level researchers choose to build careers.
Open at China Digital Times Ask Claude ChatGPT

DJI gets another chance to challenge US military blacklist

# Policy SCMP

A US appeals court sent DJI’s challenge back for reconsideration without removing the Chinese drone maker from the Pentagon blacklist, keeping a major compliance fight alive.

More
  • The court did not clear DJI; it returned the case to a lower court to reconsider one part of the company’s challenge to its alleged military links.
  • For DJI, the designation can affect government procurement, investor perception and counterparties even if commercial drone sales continue outside the US defense market.
  • The case illustrates how Chinese technology firms face several US restriction mechanisms that are distinct from export controls and can be contested through administrative litigation.
  • A remand is procedural rather than a substantive victory, but it creates another opportunity to challenge the evidentiary basis for the listing.
  • The strategic question is whether the litigation produces a narrower designation standard or simply prolongs uncertainty for Chinese hardware companies operating globally.
Open at SCMP Ask Claude ChatGPT

Chinese teams entangle cold atoms across 420 kilometres

# Analysis SCMP

Chinese scientists report entangling cold atoms over a record 420 kilometres, more than four times previous demonstrations and beyond the range where direct transmission is practical.

More
  • The experiment, published in Physical Review Letters, is relevant to quantum networking because it tests entanglement distribution over intercity-scale distances rather than within one laboratory.
  • The reported distance crosses a threshold at which direct transmission suffers fundamental limitations, making repeater or satellite-assisted architectures more relevant.
  • The result is a physics demonstration, not a functioning secure quantum network: fidelity, rate, memory lifetime and synchronization determine whether the link can support useful protocols.
  • A longer entanglement baseline could help China connect metropolitan quantum experiments and validate components for distributed sensing or cryptography.
  • The paper’s significance is therefore architectural. It moves the bottleneck from proving short-range quantum behavior toward engineering repeatable long-distance links.
  • The report does not say whether the experiment used deployed fiber, free space or a hybrid path, leaving the deployment relevance difficult to assess.
Open at SCMP Ask Claude ChatGPT

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

# AI GeekPark

Alibaba Cloud is positioning its data center and Qwen models as an integrated training pipeline for Chinese autonomous-driving companies, arguing that data operations now matter as much as chips.

More
  • The company says its Ulanqab facility supports about 60 percent of the computing used in Chinese intelligent-driving research, though the claim is not independently verified.
  • Its proposed pipeline connects vehicle data ingestion, cleaning, vectorization, model-assisted labeling, sample retrieval and training, with a reported capacity of millions of clips per day.
  • The strategic argument is that accelerators are only an entry ticket: the bottleneck is converting messy multimodal road data into high-quality training examples quickly enough for daily model iteration.
  • Alibaba is also positioning Qwen as a possible VLA foundation and Wanxiang as a world-model engine for generating rare scenarios and evaluating driving behavior.
  • This creates platform lock-in across data, model and compute layers, but also gives automakers a way to avoid rebuilding the entire training factory themselves.
  • The report does not provide independent throughput, accuracy or customer-retention data, and the 60 percent figure may reflect Alibaba’s own definition of the market.
  • The important shift is organizational: Chinese autonomy competition is moving from isolated algorithms toward industrialized data and evaluation loops.
Open at GeekPark Ask Claude ChatGPT

SenseTime proposes Tokens Per Watt as an AI infrastructure metric

# Signal Leiphone

SenseTime’s data-center unit argues that AI infrastructure should optimize effective Tokens Per Watt, not just PUE, and reports a roughly 60 percent increase in usable IT capacity through coordinated power and compute scheduling.

More
  • PUE measures facility overhead relative to IT power, but it does not say how much useful model output a watt produces; TPW links electricity, compute utilization, task value and time-of-use pricing.
  • SenseTime describes an eight-level observability layer mapping jobs and GPUs to racks, buildings and incoming power, so operators can see which workloads consume electricity and whether they can move.
  • Real-time inference stays fixed while training, evaluation and batch workloads can shift toward cheaper or less constrained periods, increasing output from existing capacity rather than simply adding megawatts.
  • The company says its Lingang AIDC improved effective IT carrying capacity by about 60 percent under existing power conditions, but the excerpt does not provide an independent audit or baseline.
  • The metric is strategically important if cloud customers begin paying for useful token throughput rather than accelerator-hours, forcing model efficiency, scheduling and energy procurement into one optimization problem.
  • TPW could become a useful commercial language, but it will need standardized definitions of effective tokens, quality, latency and workload comparability before it can support cross-provider benchmarking.
Open at Leiphone Ask Claude ChatGPT

China urged to keep its AI governance body open to the US

# Policy SCMP

Chinese analysts warn that the new World Artificial Intelligence Cooperation Organisation should not become an anti-US bloc, revealing Beijing’s concern about competing technology coalitions.

More
  • The analysts say the organization is not intended as an alternative to the US-led Pax Silica partnership, but worry that outsiders may interpret it that way.
  • The advice reflects a practical diplomatic goal: attract countries that want access to Chinese technology and standards without forcing them to choose between Washington and Beijing.
  • For China, an international AI body is useful only if it gains membership and legitimacy beyond existing political partners; an explicitly rival identity would narrow that market.
  • The episode shows AI governance becoming an arena for coalition-building, standards influence and dependency management rather than only safety rules.
Open at SCMP Ask Claude ChatGPT

China calls New Zealand's spy report foreign interference

# Policy SCMP

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

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

Moonshot AI’s IPO needs a new story beyond model rankings

# Analysis 36Kr

Moonshot AI is reportedly preparing for a Hong Kong IPO as Kimi K3 gives it a performance window, but scarce compute and fast model catch-up make timing more important than ever.

More
  • Earlier reports put potential fundraising near $3 billion, but Moonshot denied one filing rumor and the supplied article does not confirm a timetable.
  • Kimi K3 reportedly ranked near the top of Artificial Analysis while costing less than leading Western models, giving investors a clearer performance narrative than a generic low-price pitch.
  • DeepSeek’s newer model reportedly costs less than one-tenth as much per task, showing how quickly a Chinese rival can attack a temporary technical and commercial advantage.
  • Kimi temporarily paused new consumer subscriptions after demand approached cluster capacity, exposing compute availability as a constraint on turning benchmark strength into usage and revenue.
  • Recent fundraising by listed Chinese model companies demonstrates that an IPO can become a continuing financing channel, not merely a one-time cash event.
  • The market is likely to reward a company seen as near the frontier at the moment of listing, but the same mechanism creates pressure to list before the lead disappears.
  • The unresolved question is whether Moonshot can convert high-end model quality into durable distribution, enterprise contracts and sufficient inference supply.
Open at 36Kr Ask Claude ChatGPT

AI opens a new chapter for ancient Chinese literature

# Culture SCMP

AI is helping researchers preserve manuscripts, add punctuation, compare editions and recover ancient Chinese texts, turning a specialist bottleneck into a computational humanities workflow.

More
  • Traditional scholarship requires years of manual transcription, interpretation and cross-referencing across fragmented records.
  • The new workflow can combine image restoration, character recognition, language modeling and reference matching, reducing the cost of locating and comparing variants.
  • The Chinese cultural context matters because the corpus spans thousands of years and includes fragile, incomplete or difficult-to-read materials that are expensive to process at human speed.
  • AI does not remove the need for specialists: uncertain readings, historical context and editorial judgment remain central to producing a trustworthy text.
  • The supplied excerpt does not identify a specific model, archive or accuracy rate, so the item supports a direction of travel rather than a measured breakthrough.
Open at SCMP Ask Claude ChatGPT

Beijing worries about Washington chaos before Xi’s planned visit

# Policy SCMP

Chinese officials are reportedly unsettled by disorder in the Trump administration ahead of a planned Xi visit, showing that summit risk now includes US internal coordination as much as bilateral substance.

More
  • The report says the planned visit would be Xi’s first state visit to Washington since 2015 and that Chinese officials previously became frustrated by disorganized US planning.
  • The concern is operational but strategically meaningful: Beijing needs confidence that agreements, protocol and policy signals from Washington reflect a coordinated position rather than intra-administration competition.
  • The supplied text relies on unnamed sources and does not identify which plans are disrupted or whether the visit will change, so it describes anxiety rather than a confirmed diplomatic setback.
  • The episode illustrates how China reads US political process as part of the bilateral balance; unpredictability can reduce the value of summit diplomacy even when both sides want contact.
Open at SCMP Ask Claude ChatGPT

SHEIN’s IPO is about a redemption clock, not fresh cash

# Analysis 36Kr

SHEIN’s strong cash generation masks a $17.3 billion preferred-share redemption overhang, making an IPO a deadline-driven capital-structure event rather than a funding necessity.

More
  • The company generated $6 billion of operating cash flow across 2023 to 2025 and held $14.8 billion in cash at March 2026, while its historical equity funding totaled about $3.84 billion plus a smaller yuan round.
  • The pressure comes from $17.294 billion of convertible redeemable preferred shares; return rates for later investors reportedly rose from 8% to 12% in March 2026.
  • If the IPO is withdrawn, rejected or not relaunched within the specified period, redemption rights can revive, giving investors leverage even though the operating business is cash generative.
  • SHEIN’s operating engine remains unusually efficient: negative working-capital days, 36-day inventory turnover and consumer prepayments fund growth before suppliers are paid.
  • The risk is deceleration rather than immediate liquidity: US revenue fell 3.5% in 2025 and 14.3% year over year in the first quarter of 2026, while overall growth slowed sharply.
  • The company’s valuation already reset from a $98.2 billion pre-money round to $64 billion in the next financing, showing how a strong cash machine can still lose market confidence.
  • The IPO narrative therefore has to reconcile a profitable cross-border supply chain with slowing growth, US exposure and investor redemption mechanics.
Open at 36Kr Ask Claude ChatGPT

China confirms the Y-20B can become an aerial tanker

# Policy SCMP

Chinese state television confirmed that the Y-20B transport aircraft 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 one airframe to switch between transport and tanker roles.
  • A convertible design can increase fleet flexibility and reduce the number of dedicated tankers needed, though mission conversion time and usable fuel capacity are not provided.
  • Aerial refueling is a range-and-persistence multiplier: it can support larger patrol areas, longer loiter times and operations farther from mainland bases.
  • The capability does not by itself prove sustained operations beyond the first island chain, which would also require escorts, basing, maintenance, communications and protection from opposing forces.
  • The disclosure through state television is itself a signal of confidence and strategic messaging, presenting domestic engine development as part of a broader military-industrial achievement.
  • The key unknowns are the number of aircraft configured, refueling rate, receiver compatibility and operational testing; without them, the report establishes potential more than deployed capacity.
Open at SCMP Ask Claude ChatGPT

Baidu turns its GenFlow agent into a standalone office product

# AI Leiphone

Baidu is separating its document-and-drive agent into a standalone office product after reporting more than 25 million AI-office monthly users, using private documents, memory and enterprise controls as its moat.

More
  • The product combines Baidu’s document corpus, academic content and user cloud files with Office agents that can generate and edit presentations, spreadsheets, documents, web pages and other media.
  • Baidu says GenFlow had more than 100 million monthly active users in April and that AI-office monthly active users exceeded 25 million; these are company-reported figures.
  • The enterprise version emphasizes shared context, task handoff, reusable skills, identity controls and full audit trails, addressing the organizational governance gap between chatbots and business systems.
  • A finance-focused mode packages company data, research reports and spreadsheet skills into workflows for valuation, research and meeting records, showing a verticalized route to monetization.
  • The separation from the storage product is strategically notable: the file repository becomes persistent memory and proprietary context for an agent, rather than merely a place to save generated files.
  • The claims of industry-leading evaluation scores and broad adoption need independent validation, but the architecture reflects where enterprise agents are heading: data access, tools, memory and governance in one surface.
Open at Leiphone Ask Claude ChatGPT

China and Brazil plan a 24-hour disaster-warning satellite

# Policy SCMP

China and Brazil announced a satellite intended to monitor all of South America and provide round-the-clock weather and environmental warnings, deepening a long-running space partnership.

More
  • The project extends nearly four decades of China-Brazil space cooperation from remote sensing toward an operational regional public-good service.
  • A satellite covering the whole of South America would improve access to weather, environmental and disaster data for countries that cannot maintain comparable systems themselves.
  • For China, the partnership demonstrates how space infrastructure can create durable influence through services and shared technical capacity rather than military alignment.
  • The report does not give a launch date, payload specifications, financing structure or data-governance terms, so the project’s readiness remains unclear.
  • Brazil’s participation also matters institutionally: a major regional space agency is helping legitimize Chinese cooperation in a domain where Washington and Beijing compete for standards and partnerships.
  • The real test will be continuity of data delivery and whether South American agencies can integrate the information into local warning systems.
Open at SCMP Ask Claude ChatGPT

ByteDance's no-distillation bet makes model iteration slower

# Analysis 36Kr

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

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