China Desk — 14 Aug 2026, 13:25 UTC

Why Chinese users feel they cannot speak openly online

# China Digital Times 12h

A translated Chinese essay argues that blunt platform keyword filtering, not only explicit state bans, makes ordinary users resort to coded language and unable to discuss social reality plainly.

The essay says laws define broad categories of prohibited content but do not publish a complete directory of ordinary words that must be blocked.
It attributes overblocking to platforms' liability: because companies bear responsibility for harmful content, automated systems and conservative moderation become safer than nuanced review.
The result is a feedback loop in which users use homophones, initials and coded terms, making context harder for both people and filters to interpret.
The piece is notable because it locates part of censorship's lived experience in platform governance and risk management, where companies have discretion even inside political constraints.
That distinction does not remove state responsibility; it shows how national rules become everyday speech restrictions through private enforcement systems.
The essay reportedly remained online, illustrating that criticism of blunt moderation can sometimes circulate when framed as platform negligence rather than direct political opposition.
For technology leaders, the lesson is that compliance architecture can reshape public language even without a transparent blocked-word list.

Open at China Digital Times Ask Claude ChatGPT

Former premier Zhu Rongji dies at 97

# SCMP 6h

The death of Zhu Rongji is being framed in China as the loss of a reform-era economic architect whose centralizing reforms also left lasting social costs.

Also: Initium, The Wire China

Zhu served as premier during a period of state-sector restructuring, fiscal recentralization and China's deeper integration with the global economy.
Chinese commentary remembers him both as an anti-corruption reformer and as the force behind reforms that produced harsh memories for industrial workers and many families.
The split legacy matters because it captures a recurring Chinese governance tradeoff: stronger central fiscal control and faster modernization paired with concentrated adjustment costs.
His public reputation was unusually personal for a senior official, built around blunt language, discipline and a willingness to confront corruption.
The supplied reporting describes his political inheritance rather than a new policy decision, but his death provides a window into how the reform era is being reassessed.
Any official memorial treatment and the balance between praise and criticism will indicate how today's leadership wants to position that reform legacy.

Podcast revisits Zhu Rongji's economic legacy Initium
Remembering Zhu Rongji, the great premier The Wire China

Open at SCMP Ask Claude ChatGPT

Chinese team extends cold-atom entanglement record to 420 kilometers

# SCMP 2h

Chinese scientists demonstrated cold-atom entanglement across 420 kilometers, more than four times the prior distance and beyond the range where direct transmission faces fundamental limits.

The result was reported in a peer-reviewed Physical Review Letters paper dated August 11, according to the supplied text.
The experiment links cold atoms across an intercity-scale distance, providing a test bed for quantum-network protocols rather than a deployable network itself.
Crossing the direct-transmission limit shifts the engineering problem toward repeaters, synchronization, memory and error management across long optical paths.
The practical significance depends on whether the demonstrated state can be generated, preserved and measured with useful fidelity and throughput; those system metrics are not supplied.
The result is strategically notable for China's quantum-network program because it connects laboratory-scale atomic systems to infrastructure-scale distances.

Open at SCMP Ask Claude ChatGPT

Changxin's market value overtakes Tencent in a memory-chip rally

# Solidot 3h

Changxin Technology reportedly became China's most valuable listed company as investors favored domestic DRAM exposure over Tencent's rising AI costs.

The supplied report puts Changxin's market value at about 3.58 trillion yuan (roughly $504B) and Tencent's at about 4.0 trillion Hong Kong dollars (around $509B), so the comparison depends on currency and market-close conventions.
Changxin is described as the world's fourth-largest DRAM supplier, placing a domestic memory company at the center of China's AI-hardware investment narrative.
The valuation move is market sentiment, not proof that Changxin has surpassed Tencent in revenue, profit or strategic importance.
The jump reflects two investor beliefs: AI demand will keep pulling memory higher, and semiconductor localization deserves a premium even before mature operating metrics are visible.
The key risk is that a sharp listing rally can make market capitalization move far faster than manufacturing yield, customer diversity and sustainable cash flow.

Open at Solidot Ask Claude ChatGPT

Apple and Alibaba train a China-specific language model

# Solidot 5h

Chinese coverage says Apple is using Alibaba support to train a model for mainland devices, reflecting how China-specific AI regulation is forcing a foreign platform into local model partnerships.

Also: Techmeme

The report says Apple Intelligence is expected to reach mainland iPhones through a future system update after Chinese authorities registered the service.
Apple is described as training a model specifically for the China market rather than simply routing users to a global model, giving local partners a role in data, compliance and deployment.
Alibaba's involvement matters because it turns a market-access requirement into a deeper technology relationship between a global device company and a Chinese cloud and AI provider.
The strategic precedent is significant: foreign companies may need locally governed model stacks to deliver generative features in China, even when their global product architecture is centralized.
The underlying reporting is attributed to sources and the supplied text does not disclose the model's architecture, training data or final approval status.

Reuters: Apple trained a China-specific model with Alibaba Techmeme

Open at Solidot Ask Claude ChatGPT

China says Y-20B can switch from transport to aerial refueling

# SCMP 2h

Chinese state television confirmed that the Y-20B can be converted into an aerial tanker, potentially extending PLA air operations beyond the first island chain.

The Y-20B is an improved heavy transport aircraft powered by domestically developed WS-20 turbofans, according to the report.
Pallet-ready fuel tanks and removable refueling equipment allow the same airframe to shift between cargo transport and tanker missions.
A convertible fleet could increase sortie flexibility and reduce the need to maintain separate transport and tanker inventories.
The operational effect depends on tanker numbers, crew training, escort, basing and the availability of compatible receiving aircraft; the report confirms capability, not deployment scale.
The announcement is strategically important because aerial refueling is an enabling layer for sustained operations farther into the western Pacific.

Open at SCMP Ask Claude ChatGPT

China is urged not to turn AI governance into an alliance split

# SCMP 4h

Chinese coverage says Beijing's proposed World Artificial Intelligence Cooperation Organisation should avoid looking like an anti-US bloc if it wants broad international membership.

Analysts cited by the report say the organization was not designed as a direct alternative to the US-backed Pax Silica partnership.
The warning is about institutional legitimacy: countries may join a China-led forum for practical cooperation but avoid one perceived as a geopolitical counter-alliance.
China's challenge is to make standards and coordination attractive without requiring members to choose between governance ecosystems.
The episode shows AI governance becoming another venue for strategic competition over infrastructure, safety norms and dependency, not only a technical standards exercise.
The supplied excerpt does not describe the body's membership, authority or concrete rules, so its influence remains uncertain.

Open at SCMP Ask Claude ChatGPT

Microsoft is shrinking in China but keeping a strategic foothold

# Solidot 9h

Chinese coverage frames Microsoft's China presence as a retreat from local cloud and AI ambitions but a continued bet on Chinese engineers and companies needing Western technology for overseas operations.

The report says Microsoft closed at least 15 branches and joint ventures over five years as local software preference and export controls narrowed its room to expand.
It says the company remains because Chinese firms such as ByteDance need Western tools to manage international business and because Microsoft wants access to China's engineering talent.
That is a different story from a simple withdrawal: Microsoft is reducing broad platform exposure while preserving narrower, profitable and strategically useful positions.
The company's statement that it operates under the same regulatory environment as other international suppliers highlights the constraint shared by foreign technology firms rather than a special exemption.
The strategic question is whether cloud and AI restrictions make a China engineering and enterprise-support footprint valuable enough to justify the compliance and geopolitical cost.

Open at Solidot Ask Claude ChatGPT

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

# Techmeme 8h

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.

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

GLM-5.3 claims a new open-model lead in coding and security

# AI QbitAI 4h

A second Chinese account emphasizes GLM-5.3's benchmark gains, lower token use and red-team results, while leaving the strongest claims dependent on vendor testing.

Also: Leiphone

The article reports CyberGym code-review performance rising from 77.2% for GLM-5.2 to 84.5% for GLM-5.3, with the new model positioned as the leading open security model.
It says testing found 2,404 vulnerabilities, including 1,088 medium- or high-risk cases, across 220 projects; these figures differ from another account and should be treated as source-specific claims.
The model's coding demonstrations involve long-running autonomous work such as building a 3D scene and a small application with accounts, databases, permissions and tests.
The technical story is post-training and agent execution: the base model reportedly stays the same while training, evaluation and tool-use loops improve.
The report's enthusiastic tone and comparisons with frontier models make independent reproduction especially important before treating the scores as a durable lead.

GLM-5.3 targets coding and security with post-training Leiphone

Open at QbitAI Ask Claude ChatGPT

DeepSeek Harness is built for self-evolving developer workflows

# AI QbitAI 21h

Chinese hands-on coverage stresses that DeepSeek Harness exposes models, tools, policies, storage and context as editable modules, making it a developer workbench rather than a finished consumer app.

Also: GeekPark, Techmeme, Hacker News

The framework's trajectory view records raw agent events, letting developers inspect where a long task failed and how tokens were spent instead of relying on a summarized chat transcript.
It supports custom agent presets, local project directories and more than 100 built-in plugins, according to the article.
The testing narrative says the same model handled longer tasks more reliably under Harness than under another coding-agent runtime, but this is a single reviewer's experience rather than a benchmark.
The product's local-first design and explicit cost telemetry are unusually relevant to enterprise teams that need auditability and budget control.
The tradeoff is usability: the interface remains rough, the developer preview is not an ordinary desktop app, and multimodal work may require a separate vision model.
Chinese coverage frames the system as an Android-like substrate for agents, while the real strategic question is whether a plugin ecosystem can become a secure standard runtime.

DeepSeek Harness looks like an open agent operating system GeekPark
DeepSeek releases a modular agent framework under the MIT license Techmeme
Hacker News debates DeepSeek Harness and AI-generated plugins Hacker News

Open at QbitAI Ask Claude ChatGPT

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

# AI QbitAI 7h

Delta Intelligence and Wujii Technology plan a shared data standard and training baseline for full-body dexterous manipulation, addressing a major gap in physical-AI datasets.

Most dexterous-hand datasets record hand degrees of freedom while omitting torso posture, arm trajectories and visual attention; the proposed system would capture those coupled signals.
The partnership links a humanoid foundation-model company with a dexterous-hand hardware maker, covering data collection, model training, hardware adaptation and edge deployment.
A common format, annotation policy and evaluation baseline could reduce repeated integration work across robot vendors and make experiments more reproducible.
The target tasks include fine assembly such as inserting connectors and tightening threaded parts, where whole-body posture affects hand control and stability.
The standard is only planned for release within the year, so its adoption and technical quality remain untested.

Open at QbitAI Ask Claude ChatGPT

China tightens taxation of offshore trusts used by wealthy residents

# Policy Pekingnology 11h

New tax rules target offshore trusts at asset transfer, income and distribution stages; Chinese commentary frames them as a fairness measure and reopens debate over labor versus capital taxation.

The Ministry of Finance and State Taxation Administration issued rules on individual income tax for offshore trusts on July 24, according to the supplied report.
The change makes it harder for Chinese tax residents to defer or avoid tax through offshore structures, according to the analysis.
A Tsinghua finance professor says ordinary people are largely unaffected because they do not hold offshore trusts, but argues that capital income taxed at up to 20% is still favored over labor income taxed at up to 45%.
That framing is politically useful: Beijing can present enforcement against wealthy offshore structures as tax fairness while leaving open a broader debate about the design of personal-income taxation.
The rules also signal a more capable tax state, one willing to pursue wealth across legal entities and jurisdictions rather than relying only on domestic reporting.
The article does not provide the full statutory text or examples of enforcement, so the practical burden on different trust structures remains unclear.

Open at Pekingnology Ask Claude ChatGPT

What China's AI talent market is really paying for

# Culture Ginger River Review 4h

Chinese labor-market discussion suggests premium AI pay is concentrated in scarce post-training, data, systems and commercialization skills rather than every new PhD.

The source draws on a Chinese business podcast and an AI recruiting firm that says it has served more than 1,500 AI companies across chips, infrastructure, models, robotics and agents.
The headline offer of 3 million yuan for a new PhD is described as real but exceptional, not a normal entry-level market rate.
The deeper shift is from rewarding isolated research prestige toward paying for people who can turn data, training systems and model behavior into repeatable production capability.
Recruiters describe shortages spanning infrastructure, foundation models, embodied AI and agents, suggesting that the labor bottleneck is broadening beyond model architecture.
The discussion also points to AI hiring beyond technology companies, as firms seek people who can translate domain workflows into deployable systems.
Because the article is an edited conversation rather than a survey, it is useful as market texture but not as a representative salary dataset.

Open at Ginger River Review Ask Claude ChatGPT

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

# Analysis GeekPark 3h

Chinese autonomous-driving leaders met at an Alibaba data center where the competitive question was framed as a production line linking data, models and compute, not just a better algorithm.

The report says the site handles 60% of the compute used in China's intelligent-driving research, a company claim that is not independently substantiated.
Alibaba Cloud describes a pipeline from multimodal vehicle data through automated cleaning, vectorization, model labeling and retrieval of training samples, with daily processing in the millions of clips.
The architectural point is that data preparation, model training and hardware scheduling are coupled bottlenecks; buying accelerators does not solve slow labeling and fragmented environments.
Alibaba positions its general model as a possible VLA base and its generative model as a world-model engine for producing rare driving scenarios and evaluating responses.
If the pipeline really supports day-scale iteration, cloud providers can capture a more strategic role in vehicle development than storage or GPU rental alone.
The open question is whether shared cloud infrastructure can satisfy automakers' data-sovereignty, latency and differentiation requirements.

Open at GeekPark Ask Claude ChatGPT

Robot teams say their compute demand has grown up to tenfold

# AI QbitAI 5h

Chinese robotics developers report five- to tenfold compute growth in two years, pushing embodied AI toward cloud workstations that bundle simulation, environments and scheduling.

The engineering load spans teleoperation data, simulation, model training, real-robot deployment and multimodal asset management, making a single local workstation inadequate for a growing team.
The report says inconsistent CUDA, robotics and simulator versions can consume half a day per developer, turning environment setup into a recurring infrastructure tax.
Alibaba Cloud's workstation platform aims to standardize those environments and elastically allocate GPUs instead of forcing teams to buy fixed capacity.
The strategic shift is from renting compute to offering a reproducible robotics development system with access control, 3D interaction and data movement built in.
The five- to tenfold figure comes from a company executive's rough estimate, but the underlying systems problem is credible and general to physical AI.

Open at QbitAI Ask Claude ChatGPT

The best Chinese music of 2026 so far

# Culture ChinaTalk 2h

A China-focused music roundup surveys independent releases across jazz, emo, post-punk and avant-garde scenes, offering a view of cultural production rarely visible in mainstream tech coverage.

The selections span artists from Shenzhen and other Chinese scenes, alongside musicians connected to the country's long-running post-punk and experimental networks.
The source also notes a street livestream in Shanghai being interrupted by police, a small but telling example of the constraints around informal performance and online distribution.
The value is cultural texture rather than a single event: independent musicians continue to build cross-genre communities and international listening paths despite limited mainstream visibility.
The supplied excerpt is only part of the list, so it should not be treated as a comprehensive ranking of Chinese music.

Open at ChinaTalk Ask Claude ChatGPT

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

# Culture Initium 14h

A Hong Kong animal-rights community held a memorial for a dog case in Jieyang, revealing the invisible boundaries that shape cross-border civic solidarity inside greater China.

The event connects Hong Kong supporters with an animal-welfare controversy in mainland Guangdong, according to the supplied title and source context.
Animal protection offers a relatively accessible cause, yet cross-border participation still exposes limits on organizing, public attention and the movement of narratives between jurisdictions.
The story matters less for the individual memorial than for how civic identity travels across the Hong Kong-mainland boundary after political space has narrowed.
The supplied excerpt is mostly page code and does not provide enough facts to assess the underlying animal case or any official response.

Open at Initium Ask Claude ChatGPT

ByteDance's no-distillation bet could slow its language-model race

# Analysis 36Kr 11h

Chinese analysis says ByteDance's refusal to use teacher-model distillation buys training independence at the cost of time, data flywheel speed and near-term coding and agent performance.

Distillation can give a weaker model a fast cold start by learning outputs or task trajectories from a stronger model; rejecting it means building more of the capability through original data and post-training.
Researchers cited in the article say ByteDance's language model trails leading Chinese models in coding benchmarks, while its video model has performed better because real video data is harder to substitute with teacher outputs.
The strategic tradeoff is temporal: early deployment creates usage, usage produces feedback, and feedback improves the next model, so a slow start can compound into a durable disadvantage.
ByteDance is reportedly building a dedicated AI data and safety organization to source, clean, synthesize and evaluate training data, suggesting that independence requires an organizational moat, not only a principle.
The decision also carries provenance and differentiation benefits if the company believes models trained without external outputs can reach a distinct frontier.
The supplied analysis relies on interviews and public benchmark references rather than a controlled comparison of identical training budgets.

Open at 36Kr Ask Claude ChatGPT

US coastguards train Taiwan amid rising Chinese activity

# Policy SCMP 1h

US instructors began a two-week Taiwan Coast Guard program focused on small-boat operations and coordination as Chinese activity east of the island increases.

The training covers small-boat operations and team coordination, according to Taiwan's coast guard.
The supplied report frames the program as a practical partnership step rather than a new arms transfer or formal defense pact.
Its significance lies in maritime gray-zone competition: routine coast-guard capacity can raise the operational cost of Chinese presence without crossing directly into military escalation.
The report does not quantify the Chinese activity or describe any incident, so the strategic effect remains prospective.

Open at SCMP Ask Claude ChatGPT

Xiaomi phone camera reportedly invents lunar detail on the Sun

# Analysis Solidot 4h

A Xiaomi 17 Ultra image reportedly adds crater-like detail to the Sun, reviving the central trust problem of computational photography: enhancement can become fabrication.

The image appeared during a total solar eclipse and reportedly showed ridges and craters that belong to lunar imagery rather than the photographed solar surface.
The supplied report attributes the behavior to AI-based detail reconstruction and notes a similar controversy involving Samsung.
This is not merely an image-quality bug: if a camera model inserts learned structure, users cannot assume that an apparently sharp detail was optically captured.
The case exposes a product-design choice between aesthetically pleasing reconstruction and provenance-preserving imaging, especially in rare scientific or documentary scenes.
The report does not establish the exact Xiaomi pipeline, whether the behavior is reproducible, or whether the image was altered elsewhere.

Open at Solidot Ask Claude ChatGPT

Moonshot AI needs a new IPO story beyond model rankings

# Analysis 36Kr 9h

Chinese analysis says Moonshot AI's IPO window depends on turning Kimi K3's frontier performance into capital and capacity before cheap rivals erase its lead.

The article places Kimi K3 near the top of an Artificial Analysis ranking while emphasizing lower usage cost, presenting that combination as the company's valuation hook.
DeepSeek's newer model is described as having lower single-task cost, threatening Kimi's performance-plus-efficiency narrative even if Kimi remains stronger on aggregate scores.
The article says Kimi temporarily paused new consumer subscriptions after demand approached cluster capacity, exposing compute supply as a constraint on monetization.
The strategic issue is not only model quality: a company that cannot serve demand cannot generate usage data, subscription revenue or the market visibility needed to defend a premium valuation.
Moonshot's prospective listing would also create a public financing channel, as recent placements by other Chinese model companies demonstrate.
The analysis is highly dependent on third-party rankings and market reports, so the durability of Kimi K3's lead remains uncertain.

Open at 36Kr Ask Claude ChatGPT

AI opens a new chapter for ancient Chinese literature

# Culture SCMP 9h

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

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

Embodied-data startup raises millions after two rounds in 40 days

# AI QbitAI 1d

Chinese startup Yuandian Technology is building a physical-AI data platform around MatrixOS, betting that real-world interaction data will become the scarce infrastructure layer for humanoid robots.

The company says high-quality physical-interaction data is far scarcer than language data because it requires synchronized vision, touch, joint trajectories, object mechanics and temporal context.
Its platform aims to connect data collection, processing, model training and deployment, with a production network for collecting real-world examples rather than relying on web-scale scraping.
The report cites an estimate that Chinese embodied-intelligence financing reached 93.5 billion yuan (roughly $13.2B) in the first half of 2026, up fivefold year over year; the figure is not independently sourced in the supplied text.
Yuandian says it has a million-scale data order from a named customer and partnerships with model, robot-body and solutions companies, but customer delivery and dataset quality are not independently verified.
The strategic bet is that neutral, reusable data infrastructure will matter more than another vertically integrated robot company once hardware becomes easier to copy.
The company is very young, so the central risk is whether physical data can be standardized and sold across different embodiments without losing task-specific value.

Open at QbitAI Ask Claude ChatGPT

SHEIN's IPO case is about forced liquidity, not fresh cash

# Signal 36Kr 8h

Chinese analysis argues that SHEIN can fund itself but is pushed toward an IPO by redemption terms attached to $17.294B of preferred stock and a slowing growth curve.

The article says SHEIN generated $6B of operating cash flow across 2023 to 2025 and held $14.831B in cash at the end of March 2026, while its historic equity financing totaled about $3.84B plus 169 million yuan (roughly $24M).
The pressure comes from 17.294 billion dollars of convertible redeemable preferred shares and a return rate that rose from 8% to 12% for certain investors in March 2026.
The business model is unusually cash-efficient: customers pay before fulfillment, net cash-conversion days remain negative and inventory turns are reported at 36 days.
Yet growth has slowed sharply, with the article saying US revenue fell 3.5% in 2025 and 14.3% year over year in the first quarter of 2026.
The strategic lesson is that IPOs can be liability-management events and a route to reset investor redemption rights, not simply a way to finance expansion.
The detailed valuation reconstruction comes from the article's analysis of a prospectus and should not be confused with confirmed market pricing.

Open at 36Kr Ask Claude ChatGPT

Alibaba's Lingxi sale reflects a harsher reality for Chinese games

# Analysis 36Kr 3h

Chinese analysis treats Alibaba's reported $1.5B-plus Lingxi sale as a decade-long shift from internet ecosystem expansion toward monetizing mature cash-flow assets.

Also: Techmeme

The article says Trustar Capital is the leading buyer and that the reported valuation may exceed $1.5B, while noting that neither side confirmed the deal.
Lingxi's flagship strategy game has more than 100 million registered users and has remained a stable revenue generator, but the portfolio has not produced a second comparable hit, according to the article.
The contrast with US transaction coverage is the strategic framing: Chinese reporting focuses on organizational fit, Alibaba's AI and cloud pivot, and the limits of buying content capability.
The sale illustrates why a successful studio can become a divestiture candidate when its growth depends on one aging title and its parent changes capital priorities.
The article compares the case with another Chinese game-company sale, suggesting repeated ownership changes are becoming normal even for strong regional franchises.

Alibaba nears a reported sale of Lingxi Games for $1.5B-plus Techmeme

Open at 36Kr Ask Claude ChatGPT

RISC-V bet leaves Yiswei Computing burning cash before its IPO

# Signal 36Kr 8h

Yiswei Computing is seeking a Hong Kong listing after losing 4.9 billion yuan in three years, betting state-backed capital can turn a low-margin RISC-V business into a strategic platform.

The company reported revenue of 1.752 billion, 2.025 billion and 2.431 billion yuan (roughly $247M, $285M and $342M) from 2023 to 2025, but losses totaled 4.9 billion yuan (about $690M).
Its 2025 gross profit of 453 million yuan (roughly $64M) did not cover 604 million yuan of administrative expense, before 1.042 billion yuan of research spending.
The article says four financing rounds raised 9.276 billion yuan (about $1.31B), while cash at the end of 2025 was only 188 million yuan (roughly $26M).
Yiswei is framed as a national substitution bet on RISC-V against x86 and Arm, but its global market share remains small and the category is highly fragmented.
The company remains dependent on a major customer that supplied 82.1% of revenue in 2023, though that share fell to 39.3% in the first quarter of 2026.
The IPO question is therefore whether strategic importance and state capital can outrun weak unit economics, customer concentration and continuing cash burn.

Open at 36Kr Ask Claude ChatGPT

China rejects New Zealand's latest espionage assessment

# Policy SCMP 6h

Beijing's embassy in Wellington called New Zealand's threat report foreign interference and Cold War thinking after the report identified China as the only country conducting espionage at scale.

New Zealand's security service said China conducts espionage at scale, according to the supplied report.
China's response rejects the legitimacy of the assessment rather than addressing its specific allegations, framing the dispute as politicized intelligence and alliance pressure.
The exchange shows how security reporting is becoming a standing feature of China's relations with advanced democracies, with each side treating the other's framing as evidence of strategic bias.
The report does not supply the underlying cases or evidence behind New Zealand's assessment, so the item is about official positioning rather than adjudicated espionage claims.

Open at SCMP Ask Claude ChatGPT

China rebukes Philippine defense chief over detention dispute

# Policy Jiemian 11m

Chinese coverage frames the dispute as lawful treatment of foreigners and counters Manila's accusation with a personal attack on the Philippine defense chief.

A Chinese foreign ministry spokesperson said foreigners in China must obey Chinese law and that immigration authorities act lawfully while protecting legal rights.
The response rejected the claim that China targeted more than 100 Filipinos as coercion, saying enforcement was not directed at a specific nationality.
The spokesperson also accused the Philippine defense chief of having personally detained Chinese workers in the Philippines, turning a bilateral legal dispute into a sovereignty and credibility contest.
The rhetoric shows Beijing combining a formal legal justification with public condemnation of a foreign official, a common pattern in contentious diplomatic exchanges.

Open at Jiemian Ask Claude ChatGPT

Tencent moves senior post-training researcher into WeChat AI

# AI 36Kr 2h

Tencent reportedly moved senior researcher Xu Can from Hunyuan to WeLM as WeChat expands its own models and agent capabilities, signaling a sharper split between platform-specific AI systems.

Xu Can previously led a post-training direction in Hunyuan and has experience in synthetic data, instruction construction, reinforcement-learning feedback and automated evaluation.
WeChat is testing an assistant that can operate native features and mini-programs, while hiring data points to continued investment in inference optimization and agents.
Tencent's model organization has been repeatedly consolidated under Hunyuan, but WeChat retains a comparatively independent research track for product-native AI.
The supplied report says newer WeLM models use sparse mixture-of-experts designs, including variants with roughly 80 billion and 130 billion total parameters but far fewer active parameters per inference.
The move suggests Tencent sees post-training and tool-use data as portable strategic assets that can be redeployed between a general model organization and a high-frequency consumer platform.

Open at 36Kr Ask Claude ChatGPT

Sanrio's China licensing machine outgrows Hello Kitty dependence

# Culture 36Kr 3h

Chinese analysis presents Alibaba's licensing platform as a key driver of Sanrio's China expansion, while character diversification and Chinese consumer demand reshape the Japanese IP business.

Alibaba Fish secured exclusive mainland China licensing rights for 26 Sanrio characters for five years starting in 2023, according to the report.
The platform says more than 200 brands signed licensing deals in 2024, while Sanrio's Chinese store count reached 59 by the end of 2025.
China sales reached 31.44 billion yen (roughly $210M) in Sanrio's latest fiscal year, up 83.2%, with Kuromi overtaking Hello Kitty as the top-selling character in China.
Sanrio's global dependence on Hello Kitty fell from 60.6% of revenue in 2016 to 37.3%, showing how licensing and a multi-character portfolio can reduce single-IP risk.
The comparison with Pop Mart is revealing: Sanrio relies more on lightweight licensing, while Pop Mart owns more of the retail relationship and distribution economics.
The article warns that China's growth rate will become harder to sustain as the base expands and competition for character-driven attention intensifies.

Open at 36Kr Ask Claude ChatGPT

Tencent's ad engine funds an AI spending surge

# Analysis Jiemian 22m

Chinese coverage stresses the tension inside Tencent's results: marketing revenue rose 22% to 43.565 billion yuan, while its margin slipped as AI costs entered the engine room.

Tencent reported second-quarter revenue of 204.785 billion yuan (roughly $28.8B), up 11%, with marketing services up 22% to 43.565 billion yuan (about $6.14B).
Marketing-services gross margin fell from 58% to 57%, even as group gross margin improved from 57% to 58%; the report argues that AI spending is partly embedded in this line.
Domestic games generated 47.3 billion yuan (about $6.66B), while financial technology and enterprise services generated 60.3 billion yuan (about $8.49B).
The strategic story is not simply advertising growth: WeChat Video Channels monetization is expanding while Tencent is funding models, infrastructure and product integration behind the scenes.
For an operator, the important watchpoint is whether AI raises ad relevance and inventory value faster than it raises inference and research costs.

Open at Jiemian Ask Claude ChatGPT

China's canal project reaches its first full transit test

# Policy SCMP 9h

A China-focused weekly roundup highlights the first full transit test of a 134.2-kilometer canal link to Southeast Asia, putting a roughly $10.8B infrastructure project into an operational phase.

The project is described as connecting a large inland region with Southeast Asia through a major canal system.
A full transit test is a milestone in systems integration, but it does not establish commercial throughput, payback or geopolitical viability.
The project matters because Chinese infrastructure planning increasingly links inland development, logistics resilience and regional economic influence in one asset.

Open at SCMP Ask Claude ChatGPT

Baidu's Kuku AI claims 25 million monthly office users

# AI QbitAI 4h

Baidu spun out its document-and-cloud agent Kuku AI into dedicated clients after reporting more than 25 million monthly AI-office users, but the claims come from a sponsored launch article.

The product combines document search, private cloud files, multimodal generation and office agents, with a financial-services mode as its first specialized workflow.
Baidu says the system can route one instruction across presentation, spreadsheet and document agents and maintain tasks across devices through a shared context.
Its enterprise pitch centers on governance, identity controls, audit trails and a mix of public professional content with private corporate knowledge.
The reported 25 million monthly users and first-place evaluation results are company or affiliated-institution claims, not independent adoption or quality measurements.
The strategic pattern is familiar in China: large document stores and cloud accounts become distribution and memory layers for enterprise agents.

Open at QbitAI Ask Claude ChatGPT

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	10/10
36Kr	200	9/30
Solidot	200	13/20
Leiphone	200	5/20
GeekPark	200	4/30
IT Home	200	12/60	+39 over cap
SCMP	200	10/50	+7 over cap
Initium	200	5/15
Jiemian	200	8/30	+16 over cap
China Digital Times	200	1/15
The Wire China	200	1/50
Techmeme	200	6/15
Hacker News	200	2/20
V2EX	200	8/40	+30 over cap

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

A Hefei DRAM maker passed Tencent

Apple gets a licence, drones get a tariff

GLM-5.3 found a flaw forty years old

Alibaba sells the shortcut, ByteDance refuses one

What the filter blocks, and what it lets through

Same model, two verdicts on the weights

Threads we are pulling