China Desk — 15 Oct 2026, 12:07 UTC

Harbin university punishes 20 students for unauthorized VPN use

# China Digital Times 6h

A university disciplined 20 students for using VPNs, prompting debate over academic access to overseas research and allegations that reporting rules were used to settle a personal rivalry.

The university said seven undergraduates and 13 graduate students were disciplined after a two-day enforcement sweep; the account links VPN use to accessing overseas sites for academic research.
Online discussion focused not just on the firewall, but on whether a classmate reported a rival to improve her own graduate-admission prospects; that allegation is reported as rumor, not established fact.
The case exposes a conflict between restrictions on circumvention and the practical need for international research access in Chinese universities.
Commentary cited in the report frames weaponized reporting as a broader threat to trust and open discussion, especially where academic competition is intense.
The university’s action is documented, but the alleged motive for the report and the full circumstances behind the disciplinary decisions remain unclear.
The distinction matters: the immediate enforcement is concrete, while the online dispute about informer culture is a signal of public anxiety rather than proof of a wider coordinated crackdown.

Open at China Digital Times Ask Claude ChatGPT

Luckin workers say AI scheduling is squeezing labor, not easing it

# Initium 1d

Workers at Luckin Coffee say algorithmic scheduling cuts staffing and shifts work onto unpaid overtime, illustrating how retail automation can transfer efficiency gains into tighter labor control.

The report says scheduling software uses sales, orders and hours to allocate staffing, with store managers and employees describing a gap between what the system measures and the work stores actually require.
Workers say rushes, training, cleaning and inventory create labor the system does not account for, leaving staff to absorb the shortfall through unpaid extra work.
The system also appears to narrow managers’ discretion over staffing and problem-solving, while employees are judged against performance metrics that can be difficult to satisfy simultaneously.
This is a labor-governance story as much as a software story: algorithmic management changes who has authority to allocate work and who bears the cost when forecasts fail.
The account is based on reported worker experiences; it does not provide a company response or a quantified estimate of unpaid hours across the chain.
If the pattern is representative, the productivity question is not just whether stores need fewer labor hours, but whether the model is hiding necessary work by shifting it off the schedule.

Open at Initium Ask Claude ChatGPT

Kunlun’s open-source video game model runs in real time on one GPU

# Leiphone 6h

Kunlun says Matrix-Game 4.0 generates playable 480p video at 60 frames per second on one RTX 5090 by letting a conventional game engine own state and rules while a video model renders the world.

The reported configuration reaches 60 FPS at 480p on a single RTX 5090, with under 150 milliseconds from player input to visual response and peak VRAM use of 18.5 GiB.
The key architectural choice is to keep authoritative game state, collision, inventory and progression in code, rather than relying on a generative model’s implicit memory to preserve consistency.
A coding agent builds lightweight game logic; the engine updates the world and supplies geometry and semantic constraints; the video model generates the detailed visuals.
That split addresses a core weakness of interactive video generation: attractive frames do not guarantee persistent object identity or deterministic game rules over long sessions.
Kunlun released the project and game code openly, giving developers a basis to test whether this engine-plus-generator approach generalizes beyond the company’s demo.
The performance is a company-reported result, and the article does not establish whether comparable latency or consistency holds across longer, more complex games.

Open at Leiphone Ask Claude ChatGPT

Beijing and Washington pursue an AI incident framework before leaders meet

# SCMP 11h

Chinese and US teams are discussing an AI incident framework ahead of a Xi-Trump meeting, making risk communication a potential deliverable in a relationship otherwise dominated by strategic competition.

The report says technical teams from both countries are working on a framework for handling AI-related incidents and institutionalizing bilateral dialogue.
The stated aim is to make AI cooperation a possible outcome of the leaders’ meeting on the sidelines of the APEC summit in Shenzhen.
A basic incident channel would be narrower than joint AI development: its immediate value would be reducing the chance that a technical failure or ambiguous event escalates into a political crisis.
The story presents the talks as an active diplomatic process, but does not specify which incidents would qualify, how information would be shared or whether either side has accepted binding obligations.
Any framework’s significance will depend on whether it creates durable points of contact and usable escalation procedures, rather than a summit announcement with no operational follow-through.
The report attributes its account to multiple sources, so the framework remains under negotiation rather than a completed agreement.

China and US discuss an AI incident channel ahead of a summit SCMP

Open at SCMP Ask Claude ChatGPT

Microsoft credits GLM-assisted research for 23 Windows vulnerabilities

# InfoQ China 0m

Microsoft credited a Chinese security researcher using Zhipu’s GLM with finding 23 Windows vulnerabilities, including one rated 9.8; his move to Zhipu puts Chinese coverage on both the model and the talent shift.

Also: Zhidx

The report says Microsoft’s acknowledgments name researcher Yuki Chen as using GLM from Z.ai for 23 vulnerability findings, including one with a CVSS severity score of 9.8.
Microsoft’s public credit follows its review and remediation process, making the acknowledgments stronger evidence than a model vendor’s unverified benchmark claim.
The technical value is in accelerating codebase review, call-graph analysis and candidate validation; the item does not show that the model independently discovered and confirmed every flaw.
Chinese coverage presents Chen’s move to Zhipu as a consequential validation of GLM’s security capabilities, while the underlying Microsoft acknowledgments establish the specific contribution.
The report also cites broader GLM vulnerability claims from Zhipu, but those figures are separate company disclosures and should not be conflated with Microsoft’s 23 credited findings.
The strategic test is whether this becomes a repeatable security-research workflow with independent validation, not simply a prominent researcher’s endorsement.

Top security researcher joins Zhipu after GLM-assisted vulnerability findings Zhidx

Open at InfoQ China Ask Claude ChatGPT

ByteDance-linked consortium wins a $32M AI film platform contract

# Analysis Zhidx 21h

A consortium involving Volcano Engine won a three-year, 230 million yuan (about $32 million) contract to build an AI platform for Hengdian’s film industry, including specialized models and a controlled data-sharing system.

The procurement centers on a film-industry model service, production-asset management, creative tools and ecosystem administration, with compute rental included.
The tender calls for at least nine industry datasets and specialized models for tasks such as script work, storyboarding, digital characters, sound and post-production.
It specifies large training and evaluation corpora spanning text, images, video and audio, with a requirement that recent data form a substantial share.
A trusted data-space component is intended to manage permissions, ownership, circulation and compliance for film assets, making rights governance part of the technical platform.
The award shows a local-government-backed effort to build a vertical AI stack around a major production cluster; delivery quality and rights clearance will determine whether it becomes useful infrastructure rather than a procurement showcase.

Open at Zhidx Ask Claude ChatGPT

Chinese interconnect maker Zhixingwei unveils a full-stack platform

# AI Zhidx 27m

Zhixingwei showcased interconnect chips spanning storage, networking, PCIe switching and AI clusters, with more than 800,000 units shipped and reported deployments at major Chinese server vendors.

The company frames effective cluster compute as a product of chip performance, scale and interconnect efficiency, arguing that communication can strand accelerators even when their peak compute is high.
Its portfolio spans HBA and RAID products, network adapters, PCIe switches, InfiniBand products and a newly announced PCIe 6.0 switch and 800G multi-protocol AI NIC.
Zhixingwei claims its PCIe switch uses a self-developed crossbar design and matches Broadcom’s PEX89104 in its own testing; the RAID and bandwidth comparisons in the report are also company-reported.
The firm says cumulative shipments exceed 800,000 chips and names Inspur, Lenovo, H3C and ZTE among customers or solution partners.
Seven partners, including operating-system and cloud companies, signed agreements to work on integration, underscoring that software and platform qualification are part of the competitive challenge.
The meaningful test is sustained deployment at scale with independent performance, reliability and supply evidence; a broad product list alone does not establish parity with incumbent interconnect ecosystems.

Open at Zhidx Ask Claude ChatGPT

China’s September trade data show a record-scale surplus

# Signal Sinocism 13h

China’s exports rose 36.3% year over year in September and imports 31%, according to Sinocism; the newsletter also highlights Li Qiang’s effort to frame China as both a seller and a buyer.

The newsletter reports exports up 36.3% in dollar terms and imports up 31%, both far above the cited consensus figures.
It puts the monthly trade surplus at $133.53 billion, the second-largest on record, including a $34.01 billion surplus with the United States.
Li Qiang’s Canton Fair remarks cast China’s trade role as mutually beneficial, stressing that it is a major buyer as well as a major exporter.
The contrast between the strong surplus and the official emphasis on being a buyer is useful context for how Beijing is presenting trade tensions to domestic and foreign audiences.
The supplied text is a newsletter summary rather than the underlying statistical release, so the figures should be checked against official trade data before use in financial analysis.

Open at Sinocism Ask Claude ChatGPT

APUS open-sources an agent that separates control decisions from execution

# AI Zhidx 21h

APUS has released an open-source agent that uses a small decision model to choose actions while a separate harness manages tools and state, targeting faster, more reliable long-running tasks.

Reflex Agent uses a decision model fine-tuned from a 4.5-billion-parameter Qwen3.5 model to select the next action; the harness handles orchestration, tools, state and execution feedback.
APUS says its tests showed 2–3x faster page responses than comparable agents. That is a company-reported result, not an independent benchmark.
For long tasks, the design stores original task information and retrieves relevant passages as needed instead of relying only on compressed conversation summaries. “Unlimited memory” here means external persistence and retrieval, not an expanded model context window; retrieval quality remains a key dependency.
The project is released under the MIT license with desktop, web and SDK versions. Whether the architecture stays reliable across complex real-world workflows remains unproven.

Open at Zhidx Ask Claude ChatGPT

Shanghai lab says its model helped prove a difficult problem in dynamical systems

# AI Leiphone 4h

A Shanghai AI Lab team says its S2 model helped mathematicians prove a three-dimensional case of the Ghys conjecture, with a separate formalization effort used to check the proof.

The report describes a human-model collaboration: the model generated long chains of possible arguments, while mathematicians selected promising directions and corrected the exploration.
The team says an initial model output ran to about 400 pages; later iterations expanded a problem graph from 15 to 389 nodes across seven rounds of work.
A separate team used the SHEAF framework for formal verification, an important check beyond expert review of a natural-language proof.
The model’s reported role was broad search and derivation, not autonomous mathematical discovery: the researchers directed the process and judged which paths were sound.
The preprint and formalization code are reported as publicly available, allowing other mathematicians to inspect the claim and its verification.
The result would matter most if independent experts confirm the proof and the workflow proves useful on further research problems; the headline’s claim of a world-class breakthrough remains subject to that scrutiny.

Open at Leiphone Ask Claude ChatGPT

Huawei’s Kirin 9050 uses dense hybrid bonding for a folded logic layout

# AI Leiphone 9h

Huawei says its Kirin 9050 Pro is its first mass-produced logic-folded chip, using dense vertical connections to shorten critical paths and improve the performance and power budget of its flagship phones.

Huawei describes designing multiple dies as one coordinated system, placing related logic across layers and linking them vertically rather than routing all signals across a flat plane.
The company reports more than 50 million hybrid bonds and over five million effective signal interconnects, with a 1.5-micron bonding pitch.
The reported design goal is to shorten data paths, reduce clock-distribution overhead and create headroom for higher frequency or lower voltage; the article does not provide an equivalent process-node comparison.
Huawei claims the 9050 Pro improves CPU performance by 25%, GPU performance by 40% while using 45% less power, and NPU performance by 140% while using 66% less power versus the specified prior flagship chip.
For on-device models, Huawei says a 30-billion-parameter MoE model can run with under 2.5 GB of additional memory by loading active experts into DRAM and retaining inactive ones in flash.
The engineering milestone is commercialization of a non-planar logic design, but the claims need independent testing and the article gives limited detail on yield, sustained power or manufacturing constraints.

Open at Leiphone Ask Claude ChatGPT

EU hardens its China trade stance after concluding pressure can work

# Policy SCMP 7h

The EU-China hybrid-car deal has not softened Brussels’ approach: the bloc is pursuing dialogue while preparing tougher action on Chinese trade practices, according to SCMP.

The report says European officials believe the recent hybrid-car arrangement shows that threats can produce results, rather than requiring a broader retreat from trade enforcement.
Brussels is pursuing a two-track approach: continue intensive dialogue with Beijing while also strengthening measures to counter practices it considers unfair.
That framing suggests the EU sees negotiation and trade defense as complementary tools, not mutually exclusive choices.
The article does not specify the next enforcement measures or demonstrate that member states agree on every element of the tougher line.
For Chinese firms in Europe, the practical signal is that a negotiated settlement in one sector may not prevent scrutiny in others.

Open at SCMP Ask Claude ChatGPT

Ant Group open-sources a 560B-parameter model built for long workflows

# AI InfoQ China 0m

Ant Group’s Ling-3.1-flash is a 560B-total, 25B-active MoE with a 1M-token context window; its mixed-precision deployment recipe is aimed at making long enterprise workflows cheaper to run.

Also: Zhidx

The model is released with downloadable weights and is positioned for office work, healthcare, software engineering and terminal-based tasks.
Its reported architecture activates about 25 billion of 560 billion total parameters per token and supports up to a 1-million-token context window.
Ant says routed experts use MXFP4 quantization while dense layers use block-FP8, reducing weight storage from about 534 GB to 292 GB; in its Hopper test setup, the deployment fell from eight GPUs to four.
The company reports 1,673 Elo on GDPVal-AA office tasks and a score of 65.35 on HealthBench Professional, alongside strong results on software and terminal benchmarks.
Artificial Analysis is cited for roughly 210 output tokens per second and blended pricing near $0.19 per million tokens, which could make repeated tool-driven workflows economical.
These are vendor- and benchmark-reported results; the business case depends on how well the model performs on a team’s own tasks and whether the claimed speed and cost hold at production load.

Ant’s Ling-3.1-flash is tested on office and coding-agent tasks Zhidx

Open at InfoQ China Ask Claude ChatGPT

JD employee sued over alleged false claims about Douyin

# Policy Leiphone 3h

Douyin’s e-commerce unit has sued a JD employee over social-media claims that it manipulated coverage against JD; the dispute also revives scrutiny of aggressive platform pricing tactics.

The report says Beijing’s Haidian District People’s Court has accepted the case, which concerns claims posted before the June shopping festival.
The article says the viral allegations were unsupported and that the image used as evidence linked to a report that had not appeared on the cited platform’s trending list.
The report places the dispute against allegations that JD has pressured merchants to offer the lowest prices and used unusually low listings to influence cross-platform pricing.
The case shows how platform competition can turn into a legal fight over commercial disparagement, especially when claims spread through coordinated-looking social posts.
The court’s acceptance is not a ruling on the merits; the competing companies’ claims and the underlying pricing allegations require separate verification.

Open at Leiphone Ask Claude ChatGPT

China opens a long-delayed bridge to North Korea to revive trade links

# Policy SCMP 2h

Li Qiang described the New Yalu River Bridge as a landmark of China-North Korea friendship, with Beijing presenting the long-delayed link as a way to expand trade and people-to-people exchange.

The bridge opening was accompanied by a recorded address from China’s premier, who framed the project as infrastructure for closer economic ties and more movement between the countries.
The timing and official emphasis point to Beijing’s interest in keeping practical links with Pyongyang visible despite years of delay.
A bridge can improve the physical capacity for cross-border commerce, but the report gives no traffic, customs or trade-volume data showing how quickly it will be used.
The event is also diplomatic signaling: state media’s friendship language treats the connection as a bilateral relationship marker, not merely a transport project.
The scale of its economic effect will depend on North Korean border policy and actual trade activity, neither of which is established in the supplied report.

Open at SCMP Ask Claude ChatGPT

Young Chinese viewers increasingly question harmful filial expectations

# Culture SCMP 10h

A report on online reactions to family reality television describes younger Chinese viewers rejecting coercive versions of filial duty and arguing that family relationships need boundaries.

The article uses public reactions to television stories about parent-child relationships to show how viewers debate responsibility, emotional harm and what children owe parents.
Its central contrast is between traditional filial expectations and a newer emphasis on mutual respect and healthier boundaries.
Reality television gives these disputes a shared public stage, where viewers examine family behavior and openly judge power dynamics that were once treated as private matters.
The item offers a cultural reading rather than polling or evidence that attitudes have shifted uniformly across generations.

Open at SCMP Ask Claude ChatGPT

Chinese AI infrastructure engineers are seeing a sharp pay premium

# Culture Leiphone 4h

Recruiters and executives say pay for AI infrastructure specialists has risen as model companies focus on GPU utilization and inference costs, making systems engineering a scarce source of measurable savings.

The article cites individual compensation examples ranging from roughly 300,000 yuan (about $42,000) to several million yuan annually, but these are anecdotes rather than a representative salary survey.
The work includes kernel optimization, inference scheduling, expert parallelism and separating prompt processing from token generation to improve accelerator utilization.
One interviewee estimates typical GPU utilization at 25% to 27% and says stronger teams can raise it to around 40%; these are industry estimates, not audited sector-wide measurements.
The shift follows the realization that reducing compute waste can translate directly into lower operating costs as inference scales.
The reported compensation premium is likely concentrated among experienced specialists with production-scale systems skills, rather than a broad-based rise for all infrastructure roles.

Open at Leiphone Ask Claude ChatGPT

Hunan police warn rural men about costly cross-border matchmaking

# Culture SCMP 3h

Police in Hunan warned rural men that for-profit cross-border matchmaking can bring financial and legal risks, revealing official concern about a marriage market shaped by gender imbalance and brokerage.

The warning says marriages arranged through illegal agencies may lack genuine emotional foundations and can lead to financial loss or legal trouble.
China prohibits for-profit international matchmaking, according to the report, so the notice combines consumer protection with enforcement of a regulated marriage market.
The message singles out rural men, making visible the vulnerability of people seeking partners through costly intermediaries.
The article does not provide data on fraud prevalence or identify a particular case, so the notice is a signal of official concern rather than a measure of the problem’s scale.

Open at SCMP Ask Claude ChatGPT

China’s GPU challenge is shifting from chip supply to system efficiency

# Analysis InfoQ China 0m

Chinese AI infrastructure is moving beyond simply acquiring accelerators: cluster communication, scheduling, software adaptation and fault diagnosis now determine how much compute becomes usable capacity.

The interview argues that model compatibility is only a starting point; kernels, collective communication, memory behavior and operations tooling must be retuned for each accelerator stack.
At thousands of GPUs, small delays that prevent communication and computation from overlapping can multiply across training iterations and leave substantial hardware idle.
Mixed fleets from different vendors and generations add another layer of work because drivers, libraries, telemetry and error handling do not align automatically.
The source presents this as a shift from a supply-constrained phase to a systems-engineering phase, where utilization and reliability matter more than theoretical chip specifications.
The piece is based on an infrastructure company’s experience and does not offer independent utilization benchmarks for China’s clusters.

Open at InfoQ China Ask Claude ChatGPT

Major Chinese apps block Doubao’s phone-level automated controls

# Analysis 36Kr 11h

A Chinese news roundup says 151 apps, including WeChat, Taobao, Meituan and DeepSeek, rejected Doubao’s GUI automation permissions, exposing a key obstacle to agents that operate across apps.

The roundup says the apps’ refusal blocks automated clicking and other GUI actions, while leaving ordinary manual use available.
It reports that the refusals span social, shopping and workplace apps, with Tencent-linked apps making up 65 and Alibaba-linked apps 23.
The dispute is over control of the phone interface: an agent needs permission to observe screens and act across services, while app providers may resist ceding that access.
This puts a practical limit on the promise of a system-level phone agent: model capability cannot compensate for platform owners withholding the interfaces or permissions needed to execute tasks.
The item reports app declarations under a published protocol, but does not establish whether the parties will negotiate an alternative access model.

Open at 36Kr Ask Claude ChatGPT

A Chinese robotics team tackles the gap between walking and working

# AI GeekPark 1h

YuanCe Future says its whole-body control stack lets a humanoid move its body while manipulating objects, targeting a practical weakness in demos where robots walk to a task and then freeze to use their arms.

The system separates a fast motion-control model, S0, from a higher-level task model, S1, assigning balance and execution separately from planning.
The company reports reducing hand-endpoint tracking error from about 10 centimeters in its comparison baseline to around 2 centimeters.
Its demos include lifting a fallen bicycle, pressing a rice cooker button and moving a step platform before reaching a high object.
The architectural issue is coupling: stable biped motion and precise end-effector control compete, while flat-ground assumptions make transitions between locomotion and manipulation difficult.
These are company demonstrations and comparisons; reliable work across varied bodies, terrain and unscripted settings remains unproven.

Open at GeekPark Ask Claude ChatGPT

UK AI infrastructure firm Nscale’s growth includes ByteDance-linked business

# Analysis Zhidx 21h

A Chinese report on UK data-center company Nscale says a ByteDance-linked entity was a major customer in 2025, illustrating how Chinese demand can appear inside Western AI infrastructure supply chains.

The article says an unnamed customer accounted for 73% of Nscale’s 2025 revenue and links that customer to ByteDance through a Singapore entity, citing filings and press reporting.
Nscale’s business spans power, land, data centers, GPU clusters and orchestration software, making it a full-stack supplier rather than a simple cloud reseller.
The report describes large financing and capacity commitments, but also substantial losses as the company builds infrastructure ahead of revenue.
The Chinese dimension is ByteDance’s role as a customer, not evidence that Nscale is a Chinese company or that the reported contracts are under Chinese control.
The customer identification is presented as an inference from documents and media reporting, not as an explicit disclosure by Nscale.

Open at Zhidx Ask Claude ChatGPT

As AI makes code cheaper, Chinese engineering leaders say complexity rises

# Analysis InfoQ China 0m

Chinese engineering leaders argue that AI-generated code raises the bar for architecture, context and testing: the bottleneck is shifting from writing software to keeping larger, faster-changing systems correct.

Interviewees say code generation can sharply increase output per developer, but also expands the volume of changes a person must understand and own.
They emphasize project context and engineering harnesses: a model needs business rules, repository history and architecture knowledge, not just a prompt and a codebase.
One leader expects testing to move into the coding loop, especially where agents cannot observe real device behavior or end-to-end outcomes.
The discussion reflects a management tension: conventional productivity metrics may become less informative when a small team can produce far more code.
These are practitioner views collected ahead of a conference, not measured evidence that development teams have already achieved the cited productivity gains.

Open at InfoQ China Ask Claude ChatGPT

China and Togo elevate ties and pledge port and logistics cooperation

# Policy Jiemian 26m

A joint declaration elevates China-Togo relations and backs zero-tariff trade, industrial cooperation and port logistics, positioning Lome as a potential West African trade hub.

The declaration raises the relationship to a new bilateral partnership and includes commitments on trade, investment and development cooperation.
Beijing says its zero-tariff measures for African partners should support Togo’s industrialization, agriculture and local employment.
The parties say they will explore port infrastructure and logistics cooperation to strengthen Lome’s role as a regional gateway.
The statement also includes political alignment on issues important to Beijing, including the one-China principle and opposition to Taiwan independence.
These are diplomatic commitments, not evidence that specific port projects have been financed or built; the declaration’s practical value depends on follow-through.

China-Togo declaration includes port and logistics plans Jiemian

Open at Jiemian Ask Claude ChatGPT

Apple plans a China beta of Apple Intelligence using Alibaba’s Qwen

# Policy Leiphone 2h

Tim Cook said Apple Intelligence has received approval for a late-October China beta using Alibaba’s Qwen model, with translation, writing and photo tools planned but not an AI Siri launch.

The announcement reflects Apple’s need to adapt its AI stack to China’s regulatory and operating environment, including a local model partner.
The first features named are real-time translation, writing tools and photo cleanup; Siri AI is explicitly excluded from this beta.
The report does not describe the exact division of work between Apple’s own models and Qwen or how user data will be processed.
This is a strategically important market adaptation, but the feature set and availability remain subject to the planned beta.

Open at Leiphone Ask Claude ChatGPT

A Chinese drugmaker puts its president in charge of AI adoption

# Analysis 36Kr 8h

Innovent’s president took direct oversight of IT and is pushing AI into drug discovery and operations, reflecting how organizational authority—not just model access—can shape adoption in Chinese pharma.

The company says it is embedding AI staff into business teams, starting with individual use and aiming to integrate it into target discovery, clinical work, manufacturing and administration.
Its current research use focuses on AI-assisted molecule design and early target discovery; it has also entered a drug-discovery collaboration with Schrödinger.
The executive connects early-stage research capability to a broader ambition: originating more differentiated drug targets for global development rather than relying mainly on mature, validated assets.
The report says innovative drugs accounted for nearly 85% of first-half revenue, providing a business reason to seek faster R&D iteration.
The article describes a management-led experiment, not demonstrated productivity or clinical outcomes; the test will be whether AI changes the quality and speed of programs that reach patients.

Open at 36Kr Ask Claude ChatGPT

China tech roundup: Hong Kong probes an IPO, ByteDance creates an AI safety unit

# Policy Leiphone 20h

A Chinese news roundup reports a Hong Kong IPO-manipulation investigation, a new ByteDance Seed safety team and Huawei’s next Kirin chip entering tests, combining regulatory and industrial signals.

The roundup says Hong Kong’s securities regulator suspended TeaChaDao shares amid suspected manipulation of IPO demand and restricted assets held in accounts linked to suspected groups.
It reports that ByteDance’s Seed lab is creating a dedicated model-alignment and preparedness department focused on agent safety and advanced capability risks.
Huawei says its next-generation Kirin chip has produced test results in line with expectations, but the roundup gives no technical specifications.
The three items point to distinct pressures: market-integrity enforcement, formalizing safety work inside a Chinese frontier lab and continued domestic chip development.
Because this is a multi-story digest, each claim should be read as a separate report rather than a single connected development.

Open at Leiphone Ask Claude ChatGPT

A Chinese startup wants to make camera moves repeatable with a lightweight robot arm

# Analysis 36Kr 8h

A startup is adapting six-axis robotic-arm mechanics for smaller film crews, using mechanical counterbalancing to reduce motor load while preserving camera movement across three dimensions.

The report says the startup has received seed and angel investment and is targeting creators who need repeatable camera motion without the footprint and specialist crew associated with large film-production arms.
Its Cilin X design mechanically balances the camera and arm, which the company says can reduce the torque needed at a joint from about 100 N·m to under 10 N·m, easing motor and transmission demands.
The proposed 3+3 arrangement assigns three degrees of freedom to camera position and three to orientation; the company says its control approach is intended to reduce singularities that can interrupt motion in conventional six-axis configurations.
The initial workflow still has the user set key positions, with software smoothing the path. Visual subject tracking and a model of users’ visual preferences are described as longer-term ambitions, not established capabilities.

Open at 36Kr Ask Claude ChatGPT

Orbbec’s robot-sensor growth is hard to assess from its disclosures

# Signal TMTPost 1h

Orbbec’s robot-vision business is growing, but the company does not break out revenue by end market; its dependence on Ant and reliance on non-recurring gains complicate a high valuation story.

The report cites a 29% share of global robot 3D-vision revenue in 2025, while noting that Orbbec does not disclose robotics revenue separately from broader product categories.
The company says first-half global robot-vision solution revenue more than doubled year over year, but that figure does not show the business’s absolute size or profitability.
The article says non-recurring items made up a substantial portion of reported profit in 2025 and the first half of 2026, while revenue growth slowed and net profit fell.
Ant is described as a significant commercial partner, creating potential customer-concentration and related-party questions as well as a source of demand.
The story is useful as a capital-markets warning: market share and customer lists do not establish durable margins, production utilization or diversified demand.

Open at TMTPost Ask Claude ChatGPT

Deep Robotics returns to loss as robot-dog sales dominate its IPO story

# Signal TMTPost 1h

Deep Robotics reported a first-half loss after a profitable 2025, with robot dogs still driving sales and only six humanoid units sold, highlighting the gap between robotics valuations and commercial maturity.

The report says revenue reached 259 million yuan (about $36 million) in the first half, while net income turned negative at 8.8 million yuan (about $1.2 million).
Gross margin fell as the company delivered lower-margin products, and R&D spending remained high relative to current revenue.
Industrial quadrupeds account for much of the business, especially power-inspection deployments; humanoid robots remain at an early commercialization stage.
Customer concentration is material, and the article flags a major customer that is developing products of its own, creating a possible future supplier-versus-customer conflict.
The IPO question is whether field-specific robot-dog deployments can fund a transition to broader products without weakening margins or relying on a small set of buyers.

Open at TMTPost Ask Claude ChatGPT

Why Chinese commodity futures can swing far ahead of physical demand

# Signal TMTPost 2h

A sharp fall in lithium futures despite ongoing downstream demand illustrates how Chinese commodity prices trade future supply expectations, delivery constraints and leveraged positioning—not just current consumption.

The report says the January 2027 lithium contract fell below 100,000 yuan per ton while spot was about 108,500 yuan, a gap of roughly 9,700 yuan.
For lithium, the article points to expectations of mine restarts, imports and higher refining output; futures can price anticipated supply before it reaches buyers.
Its comparison with 2023 soda ash shows the reverse case: near-term delivery scarcity can drive a nearby contract sharply higher even while distant contracts price in future capacity.
A key distinction is between total inventory and exchange-deliverable stock that meets location, quality and warehouse requirements.
The article also notes that changing inventory samples can disrupt time-series comparisons, a reminder to check methodology before treating a reported stock build as a market fact.

Open at TMTPost Ask Claude ChatGPT

Former EU diplomat weighs industrial protection without a trade war

# Policy The Wire China 21h

A former EU diplomat discusses Europe’s readiness to protect industry and Germany’s harder line toward Beijing, reflecting an ongoing European debate about how to compete with China without triggering escalation.

The item focuses on European policy choices, including industrial protection and the degree of unity among EU governments on China.
It offers a China-relevant European perspective, but the supplied text contains no specific policy proposal or new action.
The piece is an interview teaser, so its strategic value is limited without the full argument.

Open at The Wire China Ask Claude ChatGPT

Oppo brings Manus agents into its phone software as sales pressure mounts

# Analysis TMTPost 1h

Oppo plans to integrate Manus’s multi-step agent into its phone assistant, choosing an external agent partner as it tries to make its premium devices more distinctive amid weaker shipments and slower replacement demand.

The article says Manus capabilities will be integrated into AI Mind Pilot and are planned for the Find X10 series running ColorOS 17.
This contrasts with companies trying to build a tightly integrated agent stack themselves; Oppo is using a third-party agent to add task execution to its system layer.
The source cites third-party estimates of falling Oppo shipments and describes a broader move to consolidate overlapping brands and favor higher-margin products.
Phone agents create a new control point between the operating system, apps and cloud services; the integration’s value will depend on what actions apps permit and how user data is handled.
The article does not establish that Manus integration will reverse Oppo’s sales decline or that consumers will pay a premium for agent features.

Open at TMTPost Ask Claude ChatGPT

China’s edge-AI debate turns on cost, interfaces and who controls the agent

# Analysis TMTPost 1h

At a Chinese industry meeting, speakers argued that agents could accelerate edge AI, but also warned that shifting computation onto devices changes permissions, risks and the allocation of responsibility.

A Fudan researcher cited the dispute over Meta’s Muse agent as an example of how interface access and commercial rules can constrain personal agents.
A domestic chipmaker said agent workloads are increasing token demand and pointed to edge inference as a way to lower recurring costs, while acknowledging that complex tasks still need cloud support.
The company described SRAM-based compute-in-memory products and said its M50 chip had secured more than 150 design wins; these are company-reported figures.
Other speakers stressed that edge models need to fit NPU and memory limits and that deployment still faces workflow, reliability and business-model barriers.
The event captures industry positioning rather than an independent forecast; the promised 1- to 2-year inflection depends on model efficiency, device economics and useful applications.

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	0/20
Ginger River Review	200	0/20
Sinocism	200	1/20
QbitAI	200	2/10
36Kr	200	8/30
Solidot	200	11/20
Leiphone	200	14/20
GeekPark	200	2/30
IT Home	200	12/60	+37 over cap
SCMP	200	10/50	+12 over cap
Initium	200	4/15
Jiemian	200	8/30	+11 over cap
China Digital Times	200	1/15
The Wire China	200	1/50
Techmeme	200	3/15
Hacker News	200	0/20
V2EX	200	8/50	+32 over cap
Zhidx	200	10/20
InfoQ China	200	8/20	+12 over cap
New Wisdom	200	8/15	+2 over cap
TMTPost	200	8/19	+11 over cap

China's biggest apps just decided an AI agent has to be invited in

151 apps told ByteDance's phone agent no

Beijing waves through Apple's easy features

Exports surged and Brussels hardened

Zhipu's cyber bet, and idle GPUs

Harbin's informer, Luckin's algorithm

Threads we are pulling