China Desk — 22 Oct 2026, 16:39 UTC

Report: DeepSeek raises 100 billion yuan in China’s largest AI round

# Zhidx 12h

Bloomberg, citing anonymous sources, reports DeepSeek raised 100 billion yuan (roughly $14B), a record Chinese AI round that could fund infrastructure, new models and IPO preparations.

The report says the fundraising target rose from 50 billion yuan to 80 billion yuan and finally 100 billion yuan as DeepSeek spoke with investors; the account relies on sources who requested anonymity.
CATL and Tencent are described as the largest investors in the round, though the report provides no final terms or company confirmation.
The proceeds are reportedly intended to support data-center infrastructure and further open models for Chinese and overseas users.
The reported scale would give DeepSeek far more capacity to compete with US frontier labs while preserving its open-model strategy; the article says it is also preparing for a possible Shanghai IPO.
Whether this changes the competitive balance depends on what the capital buys: sustained model gains, useful infrastructure and international adoption, rather than simply larger spending.

Open at Zhidx Ask Claude ChatGPT

Unitree’s founder says humanoid robots may reach a turning point within two years

# GeekPark 4h

At Unitree’s 10th-anniversary event, founder Wang Xingxing argued that rapidly improving foundation models could accelerate physical AI, while positioning mass-manufactured robot hardware as the company’s strategic edge.

Wang said the expected timeline for a major embodied-AI breakthrough had shortened to roughly one or two years, and suggested a household task demo could be possible as soon as next year; these are his forecasts, not demonstrated capabilities.
He described connecting a third-party model to a robot for tasks such as washing a brush and changing paint without additional training, framing model-to-robot alignment as the remaining challenge.
Wang said Unitree would develop its own models but also work with external model providers, an approach that treats the robot body and manufacturing scale as potential differentiators rather than requiring a leading foundation model.
He argued that production must rise dramatically for robots to become commonplace, citing annual output in the tens of thousands as insufficient for that ambition.
The article also describes a proposed self-improvement loop in which AI searches papers, writes control code, trains in simulation and tests on hardware; its value depends on reliable evaluation and safe transfer to real machines.

Open at GeekPark Ask Claude ChatGPT

Xi invokes the Long March ahead of a key Communist Party meeting

# SCMP 6h

Days before the Party’s fifth plenum, Xi Jinping cast the Long March as a lesson in unity around strong leadership, linking the revolutionary story to China’s present geopolitical challenges.

Xi delivered the speech at a commemoration marking 90 years since the Long March, according to the report.
He said the Communist Party turned defeat into victory by rallying around Mao Zedong’s leadership, making political unity the speech’s central lesson.
The timing, just before the fifth plenum, gives the historical commemoration a present-day political role, not just a ceremonial one.
The framing ties endurance and loyalty to the Party’s current geopolitical environment, a familiar way for official discourse to turn revolutionary history into guidance for present governance.
The article does not describe policy decisions at the plenum; the significance here is the leadership message and its timing.

Open at SCMP Ask Claude ChatGPT

EU-China trade truce gives way to a dispute over who yielded

# SCMP 4h

Chinese coverage describes Brussels and Beijing disputing who blinked in a fragile trade truce: EU officials see collective pressure as effective, while Beijing resists that account and new restrictions are being discussed.

The report says the agreement to avert an escalating trade dispute was reached less than two weeks before the new public dispute over its meaning.
EU officials in Brussels and Berlin reportedly view the outcome as evidence that coordinated pressure can force Chinese concessions and are considering further trade restrictions.
Beijing is described as rejecting the implication that it was strong-armed; the competing narratives matter because they shape the political room for the next negotiation.
The excerpt does not specify the agreement’s terms or the restrictions under consideration, so the immediate economic stakes cannot be assessed from this account alone.

Open at SCMP Ask Claude ChatGPT

China activates a level-four response after a reported 7.0 quake in Tibet

# Jiemian 43m

Authorities activated a national level-four earthquake response after a reported magnitude-7.0 quake in sparsely populated, high-altitude Baqing County, Tibet, and sent a central work team to the area.

The China Earthquake Networks Center measured the quake at magnitude 7.0 and a depth of 10 kilometers, according to the report.
The epicenter is described as being at an average elevation of about 4,800 meters, with a sparse surrounding population; difficult terrain can complicate assessment and response.
The State Council earthquake relief command office and the Ministry of Emergency Management activated the national level-four response and dispatched a team to guide local operations.
The report says central officials directed local authorities to assess damage, conduct rescue and relocation work, and monitor seismic activity.
The item documents the initial government response, not confirmed casualty or damage figures; the scale of impact remained unclear in the provided account.

Open at Jiemian Ask Claude ChatGPT

China plans a program to expand consumer use of smart products

# IT Home 1h

China’s commerce ministry says it will issue a plan to accelerate smart-product adoption, using consumer demand as part of a broader push to expand domestic consumption.

The ministry described the coming plan as part of the 15th Five-Year Plan’s effort to support consumption through coordinated demand- and supply-side measures.
The report links the initiative to smart home devices, connected appliances and newer intelligent terminals, but does not give the forthcoming plan’s specific incentives or requirements.
This is industrial policy through market formation: Beijing is treating consumer uptake as a way to create room for emerging product categories, not merely as a retail promotion.
The policy’s significance will depend on whether it produces interoperable products and sustained demand rather than a temporary subsidy-driven sales bump.

Open at IT Home Ask Claude ChatGPT

The human-data bottleneck in China's humanoid robotics push

# Analysis ChinaTalk 6h

An essay on China's humanoid push identifies first-person human-motion data, especially dexterous hand use, as a bottleneck amid state-backed collection infrastructure and a growing startup field.

The article says embodied intelligence is a national priority in China's 15th Five-Year Plan and reports more than 150 humanoid robotics companies and 40 state-owned robot data-collection centers operating by the end of 2025.
The technical challenge is moving beyond locomotion demos toward manipulation: reproducing the human hand's coordination requires data on detailed human actions, which the essay says cannot be passively harvested like social-media interactions.
Egocentric collection captures action from a first-person perspective, a potentially useful way to record how people interact with objects; the excerpt does not describe specific collection methods or data-quality results.
Citing the Global Times, the article reports 50 billion yuan in embodied-intelligence startup investment deals by September 2025, roughly $7 billion, suggesting a crowded race to build both robots and their training infrastructure.
The supplied excerpt ends before its account of frontline collection work is complete, so it offers little evidence about workers' conditions, consent, or how the data is being used in training.

Open at ChinaTalk Ask Claude ChatGPT

iFlytek says its Spark model was trained entirely on domestic chips

# AI Zhidx 3h

iFlytek unveiled a cloud-and-edge model family trained on Chinese computing hardware, pairing a million-token cloud model with a 4B edge model and claiming major gains in long-context training efficiency.

The company says its cloud model, Spark X2.6 293-A30B, supports a one-million-token context, while the 4B edge model is open source and intended for devices such as cars, robots and PCs.
iFlytek says its Feixing 2 platform cut training time for a 2-trillion-token, 256K-context workload from 4.5 months to 1.2 months; these are company-reported results.
For inference, the company describes a four-layer KV-cache design for domestic heterogeneous hardware and reports cache-hit rates rising from 63% to 95.7%.
Its proposed routing architecture assigns tasks to edge, cloud or joint processing, and the company reports raising routing accuracy from 81.2% to 96.1%. The practical test is whether this works across varied devices and workloads, not just internal benchmarks.

Open at Zhidx Ask Claude ChatGPT

Recursive releases Flint, a 40B coding model designed to run locally

# AI Zhidx 12h

Recursive, co-founded by Chinese researcher Tian Yuandong, released Flint, an open-weight MoE coding model that the company says can generate more than 125 tokens per second on an M4 Pro MacBook Pro.

Flint has 39.8 billion total parameters but activates about 2.5 billion per token, a sparse design intended to reduce local inference cost.
Recursive reports scores of 54.9% on Terminal-Bench 2.1, 89.7% on SWE-Bench Pro v2 and 74.2% on SWE-Bench Verified; those are company-published benchmark results.
The model was built from scratch through tokenizer development, pretraining, supervised fine-tuning and reinforcement learning, according to the report, and is released under Apache 2.0.
The local-first target is codebase analysis, bug fixing, feature work and tests without sending the full workload to a cloud service; the tradeoff is that complex jobs may need more explicit task decomposition.
The story also connects the launch to Tian’s stated interest in recursive self-improvement, but the released model itself is a coding model, not evidence that the broader research goal has been achieved.

Open at Zhidx Ask Claude ChatGPT

China Telecom open-sources a model for intent and parameter extraction

# AI QbitAI 9h

China Telecom’s Xing4.0-29B-Intent combines intent classification with structured parameter extraction, targeting fast, schema-constrained outputs for customer service and device commands.

The model uses a unified schema-driven parsing format to output structured fields such as action, device and requested setting as JSON.
China Telecom says it averages dozens of output tokens and can respond in hundreds of milliseconds, aiming to avoid verbose free-form generation in high-volume workflows.
The company released model weights, inference images and 2.75 million training examples, and says the model is already used in customer-service and smart-home products.
The company reports better aggregate results than tested general models on four extraction tasks, but the published benchmark setup and independent replication matter for judging the claim.
Unlike a narrow choice model, the system is intended both to select a category and fill in the values downstream software needs.

Open at QbitAI Ask Claude ChatGPT

A data vendor is building a large human-action dataset for robots

# Analysis QbitAI 9h

A report describes Maxinsights as a supplier of first-person human-action data to robotics firms, including a China expansion, highlighting the labor and data infrastructure behind embodied AI.

The report says the company has delivered more than two million hours of data and holds over 1.5 million hours with hand tracking and language annotations; the figures are company-related claims in the article.
Workers wearing head-mounted cameras can capture everyday hand-object interactions at scale, offering a cheaper source of varied demonstrations than collecting every example on robots.
The article names major robotics customers and describes data collection across thousands of environments, but does not independently verify the contracts or dataset quality.
The company reportedly entered China in March, placing local data collection and annotation within a global robotics-training supply chain.
The key technical question is whether human video and sensor data transfer into robot control, not merely whether the dataset is large.

Open at QbitAI Ask Claude ChatGPT

Chinese economist argues public investment must lead demand recovery

# Policy Pekingnology 1d

A Chinese Academy of Social Sciences economist argues that public investment in services, affordable housing and infrastructure maintenance can lift weak domestic demand, challenging Beijing’s caution about borrowing.

Zhang Bin argues that weak demand, slower income growth and falling spending can reinforce one another, and that public investment can help break that cycle.
The case extends beyond new construction to public services, affordable housing and upgrading existing facilities, areas the author says still have room for investment.
He disputes the assumption that additional government borrowing necessarily worsens debt ratios, arguing that growth and repayment capacity also affect the debt burden.
The piece was published by a China News Service-affiliated commentary platform as part of a series on investment; it is an economist’s policy argument, not a newly announced government program.

Open at Pekingnology Ask Claude ChatGPT

A Chinese startup proposes evidence-backed long-term memory for robots

# AI Leiphone 14h

A paper from Chinese startup Beta Infinite presents BetaMem, a robot-memory system that tracks changing world states alongside event logs and source evidence, and reports top results on four memory benchmarks.

The paper treats memory as a robot’s current judgment about the world, not simply a growing transcript of observations.
Its design separates a time-aware state graph, an append-only event log and a capacity-limited archive of selected original images and audio, allowing later claims to be checked against retained evidence.
The team reports state-of-the-art results on M3-Bench-robot, LoCoMo, OpenEQA and EgoLifeQA using a 9B open model; the benchmark claims come from the paper and require independent replication.
This architecture targets a real deployment problem: stale observations should not overwrite reality, but unbounded multimodal history is too expensive to store and retrieve.
Reliable evidence retention and correction matter because a false memory can influence multiple later actions; the excerpt notes that this is still an open engineering and research challenge.

Open at Leiphone Ask Claude ChatGPT

Chinese team reports a 100-channel single-photon detector

# AI SCMP 12h

Researchers in Shanghai say a new interconnect design supports more than 100 photon-detection channels, above the previously reported 32-channel ceiling, potentially widening the hardware options for photonic quantum networks.

The report identifies the advance as an interconnect scheme for a single-photon detector, a component used in photonic quantum technologies.
The stated result is support for more than 100 channels while maintaining detection accuracy; the excerpt gives no measured error rate, fabrication yield or operating conditions.
If independently validated and scalable, higher channel counts could support more parallel signals in quantum-network hardware without multiplying separate detector systems.
The report frames the work as a step toward broadband quantum networks, not as a deployed network or a demonstration of end-to-end quantum communication.

Open at SCMP Ask Claude ChatGPT

AI animation output is surging, but Chinese studios say profits are not

# Culture GeekPark 8h

Chinese AI-animation producers say production costs and labor needs have fallen sharply, but output growth has outpaced revenue, shifting value toward story judgment and intellectual-property selection.

One producer says a standard 80-minute animation cost about 100,000 yuan (roughly $14,000) a year ago and now costs about 40,000 yuan (roughly $5,600), while one person can produce several titles a month.
The same producer says monthly output climbed from dozens of titles to hundreds while domestic monthly revenue stayed around 20 million yuan (roughly $2.8M), suggesting capacity gains are pushing down prices rather than expanding returns.
Interviewees say AI compute has become a much larger share of costs as labor falls, so cheaper production does not automatically mean better margins.
The industry is moving toward solo creators and small teams, while platforms and publishers try to bring production, existing story rights and distribution into one pipeline.
The reported pattern is a labor-market shift: production execution becomes less scarce, while editorial taste, audience insight and rights selection may become more valuable.

Open at GeekPark Ask Claude ChatGPT

Beijing protests Taiwan’s ban on officials attending Sun Yat-sen events

# Policy SCMP 15h

Beijing condemned Taipei’s decision to bar Taiwanese officials from mainland events marking Sun Yat-sen’s 160th birth anniversary, turning a shared historical figure into a new point of cross-strait friction.

The report says Taiwan barred its officials from attending mainland events in Beijing scheduled for November 12.
Beijing publicly criticized the decision, presenting Sun Yat-sen as a revolutionary figure revered on both sides of the Strait.
The dispute shows how historical memory remains a tool in cross-strait politics: each side can invoke the same figure while contesting who has the right to represent that legacy.
The excerpt does not provide Taipei’s explanation for the restriction or the full wording of Beijing’s response.

Open at SCMP Ask Claude ChatGPT

Tesla’s profit falls as it touts demand and FSD growth in China

# Analysis QbitAI 9h

Tesla reported sharp operating-profit pressure despite improved vehicle margins, while emphasizing paid FSD growth and claiming its driver-assistance lead over Chinese automakers.

The report says Q3 deliveries were 486,500 vehicles, with Model 3 and Model Y accounting for more than 98% of the total.
Operating profit fell 56% year over year as operating expenses rose, including increased AI and research spending; the report says automotive gross margin improved.
Tesla said global paid FSD users reached 1.74 million and that monthly subscriptions exceeded half of paid users.
Musk claimed other Chinese automakers lack Tesla’s FSD lead, but the article does not provide independent comparative performance data or a timeline for full China availability.
The Shanghai plant’s September wholesale volume and export growth are cited as evidence of improved demand, linking Tesla’s global financial story to its China manufacturing base.

Open at QbitAI Ask Claude ChatGPT

Qwen team describes GPU optimizations for Gated Delta Network attention

# AI InfoQ China 0m

An Alibaba Qwen engineering presentation details FlashQLA, a GPU kernel library for Gated Delta Network linear attention that tackles recurrent-state dependencies, memory traffic and low parallelism.

The presentation explains why linear attention can reduce sequence-length cost relative to quadratic softmax attention, but its recurrent state update creates a sequential dependency that limits GPU parallelism.
FlashQLA uses the WY representation to process tokens in chunks, typically 64 at a time, enabling more of the work to use matrix-multiplication hardware.
The chunked procedure has independent projection and output stages, while the state-update stage still depends on the preceding chunk; this dependency is the central systems bottleneck described.
The article says the implementation combines sequence-dimension blocking, warp specialization and reuse of sliding-window properties to reduce redundant memory traffic and special-function-unit pressure.
The hybrid-model context matters: linear attention compresses history into a bounded state, so the presentation describes retaining occasional softmax-attention layers for long-range recall.

Open at InfoQ China Ask Claude ChatGPT

Ant launches a decision model for high-volume enterprise workflows

# AI InfoQ China 0m

Ant Digital says its 4B and 27B LingDT models choose directly among structured options, cutting token use and latency in selected business tasks while adding multi-select and ordered decisions.

Also: Leiphone

The model targets classification, intent recognition, tool selection and risk judgments where the answer space is constrained, rather than open-ended text generation.
Ant says its 27B model ranked first in a 23-benchmark composite against other decision models; the reported score weights single-choice results at 90% and multi-choice at 10%.
The model supports multi-label output and dependency ordering represented as a directed acyclic graph, which could help route compound requests to several actions in sequence.
Ant reports task-specific reductions in token use and latency, including a customer-service risk-identification accuracy increase from 76% to 89%; these are company-run tests, not independent production evaluations.
The architecture does not itself execute financial or account actions: the article says it must work with business rules, authorization systems and execution tools.

Ant’s LingDT decision model targets structured enterprise tasks Leiphone

Open at InfoQ China Ask Claude ChatGPT

A Chinese smart-home supplier says appliance AI must be cheap and reliable

# Analysis GeekPark 3h

BroadLink’s founder argues that low-cost specialist models, deployed through existing connected-appliance hardware, are a more practical route to AI home devices than putting a general model behind every interaction.

The company says its specialist-model system is already deployed with business customers at an AI cost of about 1 yuan per device per year, though the report does not independently verify the figure.
The founder argues that household-device interactions need deterministic actions, low latency and low cost; an open-ended answer to a command such as feeling cold is less useful than a correct appliance control action.
He says the existing IoT base and over-the-air update paths can let manufacturers add functions without redesigning hardware, which avoids repeating product development and certification.
The interview also says appliance connectivity has not translated into high user activation, suggesting that reliable utility, not simply adding a model, is the adoption bottleneck.

Open at GeekPark Ask Claude ChatGPT

Survey finds young Chinese adults saving for health and job shocks

# Signal 36Kr 3h

A survey of 908 Chinese adults born in the 2000s found that respondents worry about illness and job loss, save an average of 25,400 yuan a year, and often use insurance to hedge risks.

In the survey, 63.8% named sudden illness in the family as a major financial worry and 62% cited unemployment or a prolonged loss of income.
Respondents reported saving an average of 25,400 yuan (roughly $3,600) over the prior year; the report says women in the sample saved more than men.
About 64.9% said they had bought commercial insurance through online channels, reflecting how digital platforms are becoming part of household financial planning.
The findings suggest younger consumers value liquidity and a buffer against income disruption, rather than treating saving only as a route to major purchases.
This is a small, self-selected survey rather than a representative national measure, so it is best read as a mood signal, not a population estimate.

Open at 36Kr Ask Claude ChatGPT

US-led Quad and possible regional bloc talks put China back in focus

# Policy SCMP 1h

Chinese coverage reports a planned Quad leaders’ summit alongside possible talks among the US, Japan, Australia and the Philippines, framing both as part of a regional strategy to counter China’s influence.

The Philippine ambassador signaled possible meetings of the informal four-country defense grouping ahead of an ASEAN gathering in Manila.
Washington confirmed the Quad’s first leaders’ summit in more than two years would take place alongside the regional talks.
The report frames the groupings as elements of a US-led effort to counter China’s regional influence, making the China dimension central rather than incidental.
The excerpt does not provide the summit agenda or indicate that the proposed meetings have been finalized.

Open at SCMP Ask Claude ChatGPT

Beijing condemns Florida warrant for former Cuban leader Raul Castro

# Policy SCMP 17h

China accused the United States of using its courts to pressure Cuba after Florida issued an arrest warrant for Raul Castro, signaling Beijing’s public support for Havana.

The foreign ministry response came after a question from the state-run Global Times at a regular briefing in Beijing.
Spokeswoman Mao Ning referred to Castro as a comrade and a revered leader, language that places the response in a political rather than purely legal register.
China characterized the warrant as pressure on Cuba; the excerpt does not give the US allegations or China’s further diplomatic actions.
The episode fits Beijing’s broader rhetorical posture of opposing what it describes as US interference in other states’ affairs.

Open at SCMP Ask Claude ChatGPT

Former US official discusses realistic goals for China-US AI risk talks

# Policy The Wire China 7h

A former Biden-era State Department official who led the first US-China AI talks discusses what dialogue between the two powers can realistically achieve on technology risk.

The interview is presented as a reflection on the first bilateral talks about AI and their limits.
Its China dimension is direct: the question is how rival governments can discuss AI risks despite strategic competition.
The source summary provides no specific proposals, outcomes or technical risk categories, so the value here is the diplomatic perspective rather than a reported policy change.

Open at The Wire China Ask Claude ChatGPT

Huawei executive says China is moving from driver assistance toward L3

# Policy Leiphone 14h

Huawei’s automotive unit says L3 admission testing is underway in China and a national standard is due in 2027, where responsibility shifts from the human driver toward the automated system.

The executive cited government figures showing L2-or-higher systems in 70% of new vehicles by September; this is an attributed figure in the interview.
He described L3 as a significant liability transition because the system, rather than the driver, carries responsibility in the relevant operating conditions.
The company says it has disclosed nearly 17 billion kilometers of accumulated assisted-driving distance and safety data to build user confidence; the claim is not an independent safety assessment.
Huawei framed its automotive subsidiary as a supplier to automakers rather than a vehicle competitor, an important positioning claim as it expands ecosystem partnerships.

Open at Leiphone Ask Claude ChatGPT

China launches an index intended to benchmark computing prices

# Signal Leiphone 14h

A new computing-price index aims to standardize comparisons across hardware, clusters and services, part of a push to make AI compute a measurable and tradable commodity.

The platform is led by China Economic Information Service, with Jiuzhang Cloud identified as a core contributor to measurement methods and data collection.
The proposed unit, called one degree of compute, is intended to normalize heterogeneous hardware and services into a common measure.
The index plans to use market data to track prices and supply, demand, transactions and fulfillment, according to the organizers.
A credible benchmark would require transparent methodology and broad participation; the launch announcement does not establish that the unit is already widely adopted.

Open at Leiphone Ask Claude ChatGPT

Chinese apps are redesigning workflows for Apple’s foldable iPhone

# Culture GeekPark 16h

Chinese developers have adapted shopping, maps, office and media apps for Apple’s new foldable, using the larger screen for side-by-side workflows rather than simply stretching phone interfaces.

The article tests eight localized apps and highlights dual-pane layouts for product images and details, merchant messages, map routes and office documents.
A key difference from generic split-screen is app-level handoff: a location shared in chat can open directly in the adjacent map pane, if the app supports the integration.
The half-folded posture also enables video controls and comments to sit below the picture, while the phone stands on a desk.
The story is less about the hardware launch than the Chinese app ecosystem’s effort to turn a new form factor into task-specific interfaces.

Open at GeekPark Ask Claude ChatGPT

HarmonyOS reportedly passes 100 million installs, but the count is a proxy

# Signal IT Home 20m

A third-party tracker puts HarmonyOS 5+ above 100 million installs, a milestone for app economics but a manually collected estimate rather than a Huawei systemwide census.

The tracker reports about 102.2 million installations, using figures collected manually from Huawei's AppGallery; the method makes the number useful as a signal, not an audited count of active users.
Huawei executives have framed 100 million users as a threshold where the developer ecosystem could sustain a commercial cycle, rather than merely a headline adoption milestone.
Huawei reported more than 11 million developers and 380,000 available apps and services as of June 2026; a later figure cited in the article puts the available catalog above 450,000.
The article says HarmonyOS reached 10 million installations in July 2025 and crossed 100 million in October 2026, indicating rapid reported growth, but the install metric alone does not show retention or app engagement.

Open at IT Home Ask Claude ChatGPT

Report raises questions about the alleged fall of veteran official Huang Qifan

# Policy Initium 17h

Initium’s headline attributes to the Wall Street Journal a report that Chinese official Huang Qifan has fallen from power; the supplied article text is unavailable, so the claim cannot be assessed here.

The headline presents Huang’s alleged political downfall as a lens on problems under Xi Jinping, but the provided page extract contains no article body or supporting evidence.
The item concerns Chinese elite politics, yet the missing text makes it impossible to distinguish confirmed developments from interpretation or rumor.

Open at Initium Ask Claude ChatGPT

China newsletter highlights Long March messaging, youth unemployment and AI controls

# Signal Sinocism 18h

A China-focused newsletter lists Party history, youth unemployment, technology-industry integration and AI export controls among its leading topics, offering a snapshot of the issues it sees shaping debate.

The newsletter says its issue covers Xi’s Long March anniversary messaging, Premier Li’s Russia visit, youth unemployment and AI export controls.
It notes its normal feed was thinner because services used in its publishing workflow were offline awaiting PHP patches, a reminder of the shared infrastructure behind niche reporting.
This is an editor’s roundup rather than original reporting; its value is the agenda it surfaces, not verification of the individual stories.

Open at Sinocism Ask Claude ChatGPT

China and Djibouti call for open Red Sea shipping lanes

# Policy SCMP 5h

China’s first joint statement with a Red Sea nation on the US-Iran war’s fallout backs efforts to keep international shipping routes safe and open, giving Beijing a regional diplomatic voice on maritime security.

The joint communique followed a meeting in Beijing between foreign ministers Wang Yi and Abdoulkader Houssein Omar.
Both sides emphasized freedom of navigation in the Red Sea and Bab al-Mandeb, a chokepoint important to international shipping.
The phrasing supports safe passage without, in the excerpt, naming a mechanism or assigning responsibility for threats to shipping.
The statement is notable as a bilateral diplomatic position, but the account does not establish a new Chinese security commitment.

Open at SCMP Ask Claude ChatGPT

US tests new ways to command drones over long distances

# Analysis SCMP 5h

US military tests have demonstrated remote control of uncrewed aircraft from a fighter and a carrier, a capability relevant to competition with China and a potential Taiwan conflict.

The report describes a US pilot controlling an autonomous YFQ-44A combat drone from an F-22 during flight.
It also cites a separate US test involving command of uncrewed aircraft from a carrier at sea.
The strategic question is whether long-range command links can coordinate crewed and uncrewed platforms under contested communications conditions.
The article frames the tests against the PLA, but the excerpt provides no direct comparison with Chinese systems or operational readiness data.

Open at SCMP Ask Claude ChatGPT

IM Auto sticks with CATL rather than developing its own batteries

# Analysis Leiphone 4h

IM Auto is choosing CATL cells and packs for an upcoming electric SUV, betting that supplier quality systems and manufacturing scale can outweigh the control offered by in-house battery development.

The article describes CATL’s production process as using more than 7,000 quality-control points and AI-assisted inspection, claims attributed to company representatives.
The upcoming vehicle is described with 93 kWh LFP and 103 kWh ternary battery options, including cold-weather performance and battery-management features.
IM Auto’s choice contrasts with automakers developing battery brands and reflects a strategic tradeoff between owning product definition and relying on a specialist supplier.
The source is a factory and product feature, so its claims about defect rates and performance should be read as supplier-reported rather than independently audited.

Open at Leiphone Ask Claude ChatGPT

Chinese manufacturer proposes another semiconductor-gas acquisition after a costly write-off

# Analysis TMTPost 2h

New Asia Process plans to buy a controlling stake in semiconductor-gas supplier Qiyuan, but its prior acquisition losses, cash constraints and ambitious profit targets raise execution and valuation concerns.

The proposed two-stage purchase would give New Asia Process 67.13% of Qiyuan Gas through cash, shares and a planned financing; final pricing remains undetermined in the report.
Qiyuan supplies industrial and specialty gases and says its lithography gas has been certified by major lithography-equipment suppliers; the article also cites customer qualifications in Chinese semiconductor fabs.
Qiyuan earned 30.7 million yuan (roughly $4.3M) in 2025 but only 10 million yuan (roughly $1.4M) in the first half of 2026, while sellers promised at least 46.35 million yuan (roughly $6.5M) in annual adjusted profit from 2027.
New Asia’s earlier acquisition generated 327 million yuan (roughly $46M) of goodwill that was later fully impaired, and the article notes related unresolved litigation.
The deal illustrates how domestic semiconductor supply-chain ambitions can collide with the financing and diligence risks of listed-company diversification.

Open at TMTPost 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	0/20
Sinocism	200	1/20
QbitAI	200	3/10
36Kr	200	2/30
Solidot	ERR	0/0		ConnectionError: HTTPSConnectionPool(host='www.solidot.org', port=443): Max retries exceeded with url: /index.rss (Caused by NewConnectionError("HTTPSConnection(host='www.solidot.org', port=443): Fail
Leiphone	200	17/20
GeekPark	200	8/30
IT Home	200	12/60	+41 over cap
SCMP	200	10/50	+4 over cap
Initium	200	5/15
Jiemian	200	8/30	+20 over cap
China Digital Times	200	0/15
The Wire China	200	1/50
Techmeme	200	1/15
Hacker News	200	0/20
V2EX	200	8/50	+35 over cap
Zhidx	200	9/20
InfoQ China	200	8/20	+12 over cap
New Wisdom	200	8/15	+2 over cap
TMTPost	200	8/17	+9 over cap

China is closing ranks: Xi around one core, industry around DeepSeek

Xi turns the Long March into a loyalty lesson

DeepSeek's investors are industrial China

Narrow models, hard kernels, robot faith

China's helium price crashed behind a wall

Threads we are pulling