China Desk — 28 Aug 2026, 11:13 UTC

MiniMax says lower inference cost is its route to more intelligence

# Zhidx 2h

MiniMax reported first-half revenue up 283 percent and an August annualized run rate above $800M, arguing that cheaper inference expands both commercial usage and model training.

Also: New Wisdom

First-half revenue reached $116.6M, while gross profit rose 464.8 percent; the company says business revenue now accounts for 80 percent of sales.
Token consumption in July was 20 times January's level, and the company has more than 2 million enterprise and developer users, according to management.
The strategy treats inference as both a product cost and a training input: cheaper rollout, evaluation, and synthetic-data generation allows more post-training compute within a fixed budget.
MiniMax says it is pursuing large models with sparse experts and attention mechanisms that reduce long-context cost rather than simply shrinking the model.
Its planned M3.1 targets one-third the initial M3 inference cost, while M3 Pro is expected to reach roughly 3T parameters with a more efficient attention design.
The thesis matters beyond MiniMax: as agents generate many calls per task, unit intelligence cost may become a more important competitive metric than model size or raw benchmark scores.

MiniMax says agents are turning inference economics into a growth engine New Wisdom

Open at Zhidx Ask Claude ChatGPT

A terrible Chinese animated film becomes a $6M internet carnival

# China Digital Times 10h

The ultra-low-budget animated film The Cow Comes went from fewer than 300 tickets to more than 40M yuan, roughly $5.6M, as audiences paid to experience its famously awkward production together.

The film was made over five years by a mother-and-son team and initially earned about 7,000 yuan, roughly $1,000, before clips of its poor animation and strange dialogue spread online.
Audiences now laugh at screenings, recite lines, answer the screen, and share videos, while cinemas draw improvised posters that become part of the joke.
The phenomenon is not simply irony: buying a ticket turns online ridicule into a communal performance and lets viewers participate in a temporary counterculture without needing a conventional hit.
The source reports backlash from state-affiliated outlets and rumors of local intervention, making the film a case study in how official cultural gatekeeping reacts when a meme escapes intended taste hierarchies.
Its projected box office above 100M yuan, nearly $14M, would be remarkable for a film whose appeal is precisely that it looks like an industrial failure.
The episode reveals a consumer mood that is playful, self-aware, and willing to repurpose a commercial product against its own prestige narrative.

Open at China Digital Times Ask Claude ChatGPT

China-Nepal border disaster tests Beijing's rescue and infrastructure story

# Initium 32m

Chinese coverage centers on the deadly China-Nepal floods and Beijing's rescue response, while the roundup also records the disaster's wider regional toll and border implications.

Also: SCMP

The source reports at least 3 deaths and 558 missing on the Chinese side, with far higher casualties in Nepal; the figures vary across reports and should be treated as developing.
Chinese Premier Li Qiang visited the disaster area, while Xi Jinping ordered rescue efforts and Beijing allocated 120M yuan, roughly $17M, for relief in Tibet.
The suspected mechanism is an upstream ice-and-rock collapse that created a temporary blockage before a flood wave, exposing the limits of monitoring in the Himalayas.
Gyirong is a major China-Nepal trade gateway and part of a proposed northward route that Nepal sees as a way to reduce dependence on India.
The event therefore tests whether infrastructure diplomacy can include cross-border early-warning systems and disaster governance, not just roads and ports.

Open at Initium Ask Claude ChatGPT

Xi-Trump summit plans point to another trade truce extension

# SCMP 2h

Chinese reporting says Xi Jinping may bring a business delegation to Washington for a September summit with Donald Trump, where both sides may extend their trade-war truce.

The report cites four people familiar with negotiations and says the delegation is still under discussion, so the trip and agreement are not final.
A business delegation would signal that Beijing and Washington are trying to pair leader-level diplomacy with corporate and commercial reassurance.
The likely truce extension would preserve a managed pause rather than resolve the structural disputes over technology, market access, and industrial policy.
The comparison with Trump's May China visit, when US executives such as Tim Cook and Elon Musk traveled with him, underscores the role of business elites as diplomatic instruments.
For technology companies, the immediate implication is continued uncertainty rather than normalization: export controls, investment review, and supply-chain decisions remain hostage to the next political bargain.

Open at SCMP Ask Claude ChatGPT

China's top legislature passes five laws and removes senior officials

# Jiemian 37m

China's National People's Congress Standing Committee passed five laws and removed senior civilian and military officials, including two Central Military Commission members.

The session passed laws on medical insurance, farmland protection, agriculture, and national defense mobilization, plus amendments to the Lawyers Law.
It removed the civil affairs minister Lu Zhiyuan and appointed Li Changguan, alongside several changes at the courts, state supervision commission, and legislature.
More consequentially, Zhang Youxia was removed as vice chairman of the Central Military Commission and Liu Zhenli as a commission member.
The announcement provides no explanation for the removals, so the personnel changes are more important than the routine legislative package but remain difficult to interpret.
The pattern reinforces how formal legislative sessions function as a vehicle for announcing elite personnel decisions as well as laws.
For outside observers, the absence of stated reasons is itself part of the signal: personnel changes are public, while the underlying political process remains opaque.

Open at Jiemian Ask Claude ChatGPT

China's robot games expose a split between peak speed and useful work

# TMTPost 47m

At Beijing's second World Humanoid Robot Games, robots broke human speed records while often failing in ordinary tasks, revealing a widening gap between athletic demos and dependable labor.

The event included 51 events, 1,301 matches, and 2,056 robots from 16 countries, with scene-based tasks expanding from six events to 21.
Tiangong robots reportedly ran 100 meters in 8.64 seconds and set records in several races, while other machines stalled, fell, or could not stop safely.
Scene competitions involving hotels, libraries, factories, households, and emergency work were slower but more revealing because they tested perception, manipulation, and long-horizon control.
The dominant failure modes were dynamic balance, motor overheating or insufficient torque, latency, and weak recovery behavior; these are systems problems, not simply model-quality problems.
The event's shift toward autonomous operation and real workplace tasks suggests Chinese robotics institutions are trying to redefine the benchmark from can run to can work.
The commercial test is not a medal but uptime, intervention rate, cycle time, and whether the same policy survives changes in objects, lighting, and layout.

Robot games show why China's next benchmark is useful work TMTPost

Open at TMTPost Ask Claude ChatGPT

Tencent's Hy4 builds a 3D game from one prompt

# AI Zhidx 7h

A hands-on Chinese test of Tencent's Hy4 Preview found stronger long-context planning, self-correction, and tool use, including a playable 3D open-world prototype at low API prices.

Also: InfoQ China, Leiphone, Techmeme

Hy4 has 770B total parameters, 49B active parameters, and a 1M-token context window, up from the prior model's 295B parameters and 256K context.
The tester says it handled an Art Deco website, repaired JavaScript and display errors, and produced a multi-file game with collecting, eating, and building mechanics after about an hour.
The model used visual inspection and task lists to check its own output, suggesting that the agent harness and validation loop mattered as much as the underlying model.
Input pricing was 6 yuan per million tokens, roughly $0.85, with cached input at 0.3 yuan, about $0.04, matching the article's comparison to DeepSeek.
The result is a product demonstration, not a reproducible benchmark; the key unresolved issue is whether long-running delivery remains reliable when requirements, permissions, and external tools are less cooperative.

Tencent open-sources a 770B model built to improve itself InfoQ China
Tencent open-sources Hy4 and targets the first tier of open models Leiphone
Tencent releases Hy4, a 770B open model with a 1M context window Techmeme

Open at Zhidx Ask Claude ChatGPT

Yunzhisheng turns enterprise agents into a $670M business line

# Signal QbitAI 1h

Chinese AI company Yunzhisheng reported 4.78B yuan, roughly $673M, in first-half agent revenue, while token revenue rose about 760 percent from a small base.

Agent applications, platforms, and integrated solutions represented 85.1 percent of the company's 5.62B yuan, roughly $791M, first-half revenue.
The company says it serves more than 470 medical institutions and has deployed agents across healthcare, insurance, transportation, and manufacturing.
Token revenue was nearly 30M yuan, about $4.2M, with second-quarter revenue above 25M yuan and gross margin above 60 percent; the growth rate is striking but the base remains small.
More than 60 percent of revenue came from repeat purchases, according to the filing summary, suggesting agents embedded in operational workflows can generate expansion revenue beyond initial deployment.
The model resembles an enterprise software integrator: a foundation-model layer matters, but the monetizable moat is connecting data, models, and legacy systems to domain processes.
The company remains loss-making, so the important test is whether recurring deployment and token usage narrow losses without requiring proportionally larger services teams.

Open at QbitAI Ask Claude ChatGPT

Alibaba's Qoder turns coding agents into a general-purpose workbench

# AI Zhidx 8h

Alibaba redesigned Qoder for both developers and nontechnical users, hiding code execution behind natural language while connecting repositories, cloud services, plugins, and enterprise systems.

The platform separates programming and general modes, with an automatic model router intended to balance quality, speed, and cost for users who do not want to choose models manually.
Qoder says it has more than 6 million users and over 100,000 enterprise users, while its connector, plugin, and skill counts show an attempt to build a broader agent operating layer.
In tests, it built a batch image-processing desktop app, inspected its output with browser screenshots, repaired bugs, and packaged the result for a nondeveloper.
The longer example had Qoder turn a one-line request into an iOS health-management prototype by clarifying features, creating data, installing tools, and iterating on the result.
This is a strategically important expansion of coding agents: programming becomes a general execution capability for design, operations, and product teams.
The risk is that hiding implementation increases the importance of permissions, provenance, testing, and user ability to judge whether the generated artifact is safe to deploy.

Open at Zhidx Ask Claude ChatGPT

GLM 5.3 Flash runs on a Chinese heterogeneous-compute stack

# AI QbitAI 7h

Zhipu launched and open-sourced GLM 5.3 Flash, while SenseTime said a domestic-chip cluster served 62T tokens in testing at Nvidia-class cost and performance.

GLM 5.3 Flash is described as a 320B-total, 18B-active native multimodal model trained on 30T tokens, with a claimed intelligence-index score of 57.
The pre-release model reportedly handled 62T tokens through Chinese chips, and the source says end-to-end service performance improved threefold from the initial hardware baseline.
SenseTime's heterogeneous inference system reportedly delivered 1.25 times the cost performance of Nvidia H-series hardware and 2.5 times the token output of same-cost homogeneous domestic systems.
The platform says it processed 2.42T tokens per day in July and aims to exceed 10T daily by year-end, but these are company claims without external audit.
The architectural lesson is that domestic-chip competitiveness depends on routing different inference phases to different chips, then hiding the differences behind a common model and tool layer.

Open at QbitAI Ask Claude ChatGPT

Alibaba's ABot-Recon rebuilds long 3D scenes from 12 frames

# AI QbitAI 7h

Alibaba's Amap released a streaming 3D-reconstruction model that uses a 12-frame local window to build scenes over thousands of frames, cutting memory growth without long-range anchors.

ABot-Recon predicts local point clouds and relative camera pose, then composes them online into a global trajectory rather than storing the entire sequence in memory.
The source reports 24.45 frames per second on KITTI, 6.71GB peak memory, and a 40.6 percent lower trajectory error than a representative method on a long Oxford sequence.
It uses only monocular RGB video and does not require depth sensors or known camera parameters, lowering deployment requirements for private spaces such as warehouses and malls.
The counterintuitive design is that bounded local context can produce more stable long-run computation than a large memory bank, provided drift correction is strong enough.
The remaining risk is domain transfer: camera motion, scene changes, dynamic objects, and poor texture can still break the local-pose assumptions.
Alibaba has released code and weights, making this one of the more directly testable Chinese physical-AI releases in the batch.

Open at QbitAI Ask Claude ChatGPT

The US and China are turning war propaganda into meme warfare

# Analysis ChinaTalk 1h

An analysis compares Chinese and US military videos that use cute or comic imagery to package real violence as entertainment, showing how platform-native propaganda lowers the emotional barrier to war.

The Chinese example uses AI-generated animals transforming into missiles, drones, submarines, robots, and humanoid systems during exercises directed toward Taiwan.
The US example uses stylized combat footage, action-movie music, and cartoon inserts, suggesting that the tactic is not unique to one political system.
The strategic shift is from formal patriotic messaging toward shareable short-form media designed for algorithmic circulation and emotional speed.
Turning military force into memes can normalize escalation, obscure civilian costs, and make propaganda harder to distinguish from ordinary internet culture.
For technology executives, the lesson is that generative media and platform incentives are becoming part of military signaling, not merely public relations.

Open at ChinaTalk Ask Claude ChatGPT

Beijing's HICOOL turns a startup contest into an innovation network

# Policy Zhidx 1h

HICOOL drew 10,209 projects from 141 countries and regions, while Beijing used the event to connect foreign founders, state-backed services, research programs, and industrial buyers.

The contest selected 200 winning projects from AI, robotics, medicine, quantum information, biotechnology, and other hard-tech fields; 32 of 48 finalists were in the overseas group.
Beijing launched a global startup-services alliance with more than 200 organizations and a support plan aimed at one-person companies, expanding the event beyond prizes and publicity.
The platform says its ecosystem has produced five listed companies, 18 unicorns, and 219 specialized small firms, with winning projects raising more than 719B yuan, roughly $101B, though these figures are official claims.
The planned digital-life science initiative and robot self-evolution academy show how the city is linking entrepreneurship programs to strategic research themes.
The deeper signal is institutional: Beijing is building a funnel that brings projects into local capital, talent, procurement, and policy networks rather than relying only on university spinouts.

Open at Zhidx Ask Claude ChatGPT

WALL-SS tries to make robot world models causal

# AI GeekPark 6h

Chinese startup Zibian released WALL-SS, a next-scale autoregressive world model that predicts action-linked physical states and reports stronger 60-second rollouts than competing systems.

Instead of generating an entire future video through diffusion, WALL-SS organizes the sequence as observation, action, and new observation, making the action part of the state transition.
The model reportedly reached an action-following score of 0.29 versus 0.044 for Cosmos3-Nano and kept object trajectories coherent for 60 seconds.
A multiscale memory system preserves recent detail while compressing older history, addressing the error accumulation that makes long autoregressive rollouts unstable.
In 600 paired real-versus-simulated experiments, the source reports a 0.926 correlation between task success rates and 89 percent accuracy in ranking policies.
The important distinction is between a video that looks plausible and a simulator that correctly predicts the consequence of a robot action.
The remaining question is whether the correlation survives new robots, materials, contact dynamics, and tasks beyond the company's evaluation set.

Open at GeekPark Ask Claude ChatGPT

China's tax crackdown and Unitree's stock slide top a policy roundup

# Analysis Sinocism 1d

A China-focused newsletter connects tax enforcement, a surprise anti-corruption session, Evergrande's founder, and Unitree's stock decline to Beijing's effort to stabilize revenue and manage markets.

The roundup says the central government is using tax crackdowns to help cover local-government shortfalls, linking fiscal stress to more visible enforcement.
It also highlights a surprise plenary session of the Central Commission for Discipline Inspection and the life sentence of Evergrande founder Hui Ka-yan, reinforcing a climate of continued political and financial discipline.
The discussion of Unitree's sharp post-IPO decline asks what the Party wants from the stock market after encouraging strategic technology listings, rather than treating the price move as an isolated market event.
A cabbage scandal in Hebei is included as a reminder that food quality and consumer trust remain politically salient even amid high-tech policy.
The newsletter is commentary, not primary reporting, but its value is synthesis: fiscal extraction, anti-corruption, property cleanup, and technology-market management are being read as parts of one governance posture.

Open at Sinocism Ask Claude ChatGPT

Chinese education-data project advances to the national finals

# Policy QbitAI 3h

A Zhejiang education project won its provincial data-elements contest by linking student records, knowledge graphs, and a planning model across 60 schools and more than 100,000 students.

The project combines academic results, interests, behavior, traits, potential, needs, and resources into student profiles, then uses a model plus structured knowledge graphs for recommendations.
It has reportedly been deployed in 14 provinces and supports 72 education paths, including conventional exams, arts, sports, and overseas routes.
The contest is organized by China's National Data Administration and evaluates governance, data circulation, measurable impact, and replicability rather than model novelty alone.
The reported 60 percent core-feature usage rate is more informative than the award itself, though the source is company-authorized promotional material.
The policy signal is that China is pushing data projects toward sector-specific operating models and official demonstrations, with education framed as a scalable data-governance use case.

Open at QbitAI Ask Claude ChatGPT

China's CPU market enters a five-year contest driven by agents

# Analysis Leiphone 3h

Chinese investors and chip executives see agent workloads reviving the domestic CPU market, because orchestration, databases, tool calls, and scheduling can dominate an agent's non-model compute.

One server vendor estimates that more than 80 percent of a representative agent workload runs on CPUs rather than GPUs, covering tool execution, database queries, and task coordination.
The article says AI-center CPU-to-GPU ratios could move from roughly 1:8 toward 1:4 or even 1:1, though the source notes that workload mixes are still changing.
Global server CPU prices rose nearly 30 percent from late 2025 to mid-2026 in the article's account, while one Chinese model company reportedly quintupled CPU demand.
Domestic Arm vendors with products already in production have the near-term advantage, but RISC-V attracts capital because of licensing and geopolitical autonomy.
The Chinese market is therefore splitting into an execution race for shipped Arm products and a longer ecosystem bet on RISC-V.
For infrastructure planners, the important shift is that agent scale changes the balance between accelerator capacity and the general-purpose control plane around it.

Open at Leiphone Ask Claude ChatGPT

China's 6G policy is converging with Qualcomm's AI-network vision

# Policy Zhidx 1d

China says it will accelerate 6G research while Qualcomm presents networks that combine connectivity, sensing, and compute, making AI-native infrastructure a shared direction.

Qualcomm's vision treats 6G as more than a faster data pipe: networks should sense objects, classify motion, allocate resources, and decide where computation occurs.
The company demonstrated full-duplex sensing of vehicles and drones, large MIMO prototypes, and a terminal-to-edge-to-data-center compute continuum.
Chinese officials say the next five-year period will prioritize 6G core technology, trials, standards, and ecosystem development, with commercial deployment targeted around 2029 in the company's account.
The China dimension is substantive because Qualcomm expects Chinese companies to participate heavily in standards while China is also building domestic alternatives and defining its own industrial priorities.
The hard systems problem is coordination across radio, sensing, edge compute, devices, and AI workloads; the marketing vision is much easier than achieving useful latency, energy, and reliability.

Open at Zhidx Ask Claude ChatGPT

SenseTime says domestic compute now supports production video generation

# AI QbitAI 7h

SenseTime and HiDream say they moved a video-generation service onto heterogeneous Chinese chips, using multi-card parallelism and an abstraction layer to preserve quality and scale.

The teams used distillation, sequence parallelism, tensor parallelism, and classifier-free-guidance parallelism to report a 93 percent speedup in multi-card video generation.
They also addressed visible differences between single-card and multi-card outputs, including facial features and poor final-frame details, showing that migration is an application-quality problem as well as a kernel problem.
A unified hardware abstraction layer reportedly supports more than ten heterogeneous domestic chips without changing the application workflow.
The arrangement uses forward-deployed engineers to tune models, operators, communication, and tools against business metrics rather than treating hardware compatibility as a one-time port.
The strategic signal is that Chinese AI infrastructure vendors are selling migration labor and orchestration as much as silicon; real production traffic is the test of whether domestic compute is genuinely competitive.

Open at QbitAI Ask Claude ChatGPT

A single video teaches a robot to open and pass through doors

# AI Leiphone 3h

Researchers from Nua Robotics and two Chinese universities used one RGB video per door to build digital twins and train a robot, reaching 96.57 percent success on five doors and 80.95 percent zero-shot transfer.

The system reconstructs door geometry, handles, hinges, collisions, and scale into a simulated DoorTwin, then searches for executable approach, grasp, rotation, push, and traversal trajectories.
Training randomizes pose, friction, damping, resistance, camera parameters, and depth noise, shifting physical trial-and-error into simulation.
The reported real-robot evaluation covered 175 trials, with 169 successes, while the zero-shot test used three structurally similar doors not seen during training.
Visual processing and policy inference ran locally on the robot, which matters for latency and deployment independence.
The result is not a general door-opening solution: pulling doors, different handles, broader geometries, and cross-robot transfer remain open problems.
Its commercial value will depend on whether digital assets and failure-replay pipelines reduce the time and human labor needed to adapt a robot to each new facility.

Open at Leiphone Ask Claude ChatGPT

A personalized cancer vaccine raises a $475,000 pricing question

# Analysis QbitAI 4h

A Chinese analysis examines the potential price of Moderna and Merck's AI-assisted melanoma vaccine, stressing that the $475,000 figure is an analyst forecast, not an approved price.

The vaccine reached a positive Phase 3 interim result with pembrolizumab, but it is not yet approved and the manufacturers have announced no price.
William Blair's estimated wholesale price of $475,000 is roughly 3.37M yuan, and adding as much as $221,000 for pembrolizumab could put the drug cost near $696,000, about 4.94M yuan.
The treatment is individualized: tumor sequencing, neoantigen selection, mRNA design, manufacturing, and quality release are repeated for each patient, making batch economics difficult.
The article compares the estimate with the original $475,000 price of Kymriah, another individualized therapy, while noting that the analogy is not proof of final pricing.
The social question is central in China as elsewhere: AI can shorten discovery and personalize treatment, but personalization may also create a therapy affordable only through wealthy patients or strong insurance.
The source is an analysis rather than clinical or pricing guidance, so efficacy, manufacturing cost, reimbursement, and approval remain unresolved.

Open at QbitAI Ask Claude ChatGPT

Chinese robotics scientist says factories need reliability before hype

# Analysis Leiphone 4h

A Chinese robotics leader says humanoids have moved into factories but remain limited by reliability, generalization, precision, speed, and the irreducibility of physical interaction.

The team says it has tested humanoids in electronics, clothing, logistics, and automotive assembly, with some applications reaching scaled deployment.
The gap from demo to production includes abnormal conditions, object and lighting changes, submillimeter force control, execution speed, cost, and recovery after model failure.
The proposed response is to combine visual and force-tactile representations, world models, and layered control rather than expecting a vision-language model alone to capture physical causality.
The speaker's L2-to-L4 framing treats current robots as capable of bounded tasks while reserving true multi-scene, nonstructured work for a later autonomy level.
The Chinese industry context is unusually fragmented, with nearly 200 humanoid integrators, making reusable data, tactile sensing, and common evaluation more valuable than another isolated demonstration.
The practical benchmark is whether a robot can keep working through variation without repeated data collection and manual retuning.

Open at Leiphone Ask Claude ChatGPT

A once-unfashionable Chinese photonics startup reaches a $2.8B valuation

# Analysis 36Kr 8h

Guangben Technology says years of rejected fundraising preceded several billion yuan of new capital and a valuation near 20B yuan, roughly $2.8B, after its optical-computing products reached customers.

Also: TMTPost

The company reportedly saw hundreds of investors before getting only a handful of term sheets, then used a 128-by-128 chip, customer proof-of-concept work, and early orders to change its financing story.
It combines optical interconnects, which move data between processors, with optical computing, which performs parts of inference in the optical domain.
The firm says it has shipped more than ten optical-computing systems and secured orders or intent agreements worth several hundred million yuan, roughly $30M to $60M.
Its glass-substrate and 256-by-256 chip work aims to address area, optical loss, and manufacturing cost, but the source offers no independent yield or performance validation.
The financing illustrates how Chinese investors are using Nvidia's photonics push and domestic AI-capex growth as valuation anchors for local alternatives.
The technical challenge is still systems integration: optical boards must fit existing electrical software, memory, packaging, and data-center conventions before theoretical efficiency becomes customer savings.

Chinese photonic-computing startup raises billions of yuan TMTPost

Open at 36Kr Ask Claude ChatGPT

Why vision-based tactile sensors may be a weak robotics moat

# Analysis QbitAI 5h

A Chinese analysis argues that camera-based tactile sensors often repackage mature vision components, leaving weak hardware and algorithmic barriers for industrial robotics.

The critique focuses on a recurring architecture of miniature camera, light source, elastic material, and marker field, with much of the algorithmic stack borrowed from computer vision.
The author argues that dependence on external image sensors, rigid optical cavities, and commodity components limits supply-chain control and makes prototypes easy to copy.
Thickness, focal distance, and rigid packaging make these sensors difficult to place in compact dexterous fingers or extend across curved robot bodies as electronic skin.
The analysis is intentionally adversarial and does not present independent comparative measurements, so its claims about the entire sector should be treated as a thesis rather than established fact.
The useful question for investors and builders is whether a tactile vendor owns calibration, materials, sensing physics, and deployment reliability, or merely assembles an off-the-shelf optical stack.

Open at QbitAI Ask Claude ChatGPT

China expands data-labeling pilots to 32 more cities

# Policy Jiemian 26m

China's National Data Administration will add 32 cities to data-labeling pilots, bringing the total to 39 across manufacturing, healthcare, transport, and embodied intelligence.

The policy treats labeled data as infrastructure for AI training and is extending pilots beyond isolated local experiments into a national city network.
The source says the initial seven cities included Chengdu and Hefei; the second group is intended to cover industrial and physical-AI use cases.
This is a labor and governance story as much as a technical one: Beijing is trying to formalize a data-production industry with local-government participation.
The success criterion will be whether these pilots produce repeatable quality, provenance, and pricing standards rather than simply more labeling capacity.
For companies building models in China, the move signals continued state support for domestic data supply chains and for treating annotation as a strategic industrial capability.

Open at Jiemian Ask Claude ChatGPT

Congo truce may open minerals to US investors, but China remains ahead

# Policy SCMP 3h

A possible Congo peace deal could give US investors access to coltan and other deposits, yet Chinese firms still dominate the mining assets, relationships, and supply chains built over decades.

The report presents the peace plan as a potential opening around the rebel-held Rubaya coltan deposit, not as a completed investment opportunity.
China's advantage is described as an accumulated system of mines, local relationships, processing capacity, and logistics rather than a single acquisition.
A lasting truce would therefore narrow a political access gap but would not quickly erase China's operational and supply-chain lead.
For technology companies dependent on critical minerals, the story illustrates why supply-chain position is built through long-term presence and infrastructure, not only through strategic announcements.

Open at SCMP Ask Claude ChatGPT

Qunhe's AI revenue jumps 177 percent as spatial intelligence commercializes

# Signal 36Kr 9h

Chinese spatial-computing company Qunhe reported first-half adjusted profit of 55.4M yuan, about $7.8M, up 211 percent, while AI application revenue rose 177 percent.

Revenue was 405M yuan, roughly $57M, with an 83 percent overall gross margin, according to the company report.
Its LuxReal platform combines spatial generation with 3D data for video, while SpatialVerse sells physically coherent synthetic data for world-model and robot training.
SpatialVerse booked 6.8M yuan, about $1.0M, in orders in the first half and is working with Chinese physical-AI companies including Agibot, Galaxy General, and Hesai.
The company is also building a self-owned GPU cluster that processed about 12.9 million compute tasks per day, tying infrastructure, models, and applications into a data flywheel.
Its strategy is to be the picks-and-shovels supplier for spatial intelligence: applications create data, data improves models, and better models expand applications.
The open question is whether high gross margins and fast AI revenue growth can persist as 3D generation becomes more competitive and customers demand production-grade physical accuracy.

Open at 36Kr Ask Claude ChatGPT

Chinese inference-chip startup Sunrise raises $280M at a $2.8B valuation

# Signal 36Kr 10h

Sunrise reportedly raised 2B yuan, roughly $282M, lifting its valuation near 20B yuan, about $2.8B, less than two years after being spun out of SenseTime.

The company has already mass-produced two chips and is positioning its S3 as an inference-first design rather than a general training-and-inference GPU.
S3 reportedly uses LPDDR instead of HBM to favor capacity, cost, and supply reliability for long-context inference, accepting lower memory bandwidth.
The investment thesis is being reshaped by agents, which create many serial model calls and make token cost, memory capacity, and predictable supply more important than peak training throughput.
The startup's backers include industrial and state-linked capital, showing how China's AI-chip financing is combining strategic autonomy with commercial demand.
The valuation is still a forward bet: large-scale delivery, software compatibility, and the promised cost curve have not yet been established by public customer data.

Open at 36Kr Ask Claude ChatGPT

Chinese robot makers show revenue growth before profitability

# Analysis TMTPost 55m

First-half filings from Unitree, Dobot, Topstar, and others show fast revenue growth and selective profit, but the industry's shift from industrial robots to embodied systems remains expensive.

Unitree reported about 1.15B yuan in revenue, roughly $162M, up 48.54 percent, and about 274M yuan in attributable profit, roughly $39M, while Dobot grew revenue 106.6 percent but remained loss-making.
Gross margins were comparatively strong at several firms, including about 47.4 percent at Dobot and 46.6 percent at Estun Robotics, but margins alone do not prove that humanoid demand is durable.
Traditional industrial-robot companies are moving toward embodied intelligence because standardized production is giving way to smaller batches, frequent line changes, and more variable tasks.
Their main advantage is accumulated factory access, application knowledge, and operating data; their main risk is spending on general-purpose systems before customers pay for the added flexibility.
The crucial indicators are confirmed orders, deployment uptime, service economics, and how much existing industrial hardware and data can be reused in humanoid products.

Open at TMTPost Ask Claude ChatGPT

Bangladesh confirms a Chinese fighter-jet deal

# Policy SCMP 2h

Bangladesh confirmed plans to buy China's J-10CE fighters and cited their reported performance against Rafale jets, suggesting Chinese weapons are gaining credibility in export markets.

Bangladesh's government acknowledged a long-expected draft proposal for the 4.5-generation fighter, according to the report.
The justification invokes the aircraft's alleged combat performance in the India-Pakistan conflict, turning battlefield perception into an export-sales argument.
The deal would give China a high-profile customer and strengthen its position in the market for relatively advanced but less politically constrained combat aircraft.
The claim about combat performance is politically and technically contested, so the more durable signal is Bangladesh's willingness to use it publicly as part of the procurement rationale.

Open at SCMP Ask Claude ChatGPT

Australia risks losing access to research by cutting Chinese university ties

# Policy SCMP 5h

Australia's decision to end partnerships with two Chinese universities is framed as alignment with Washington's research-security policy, with a cost to Australian scientific access.

The report says the decision followed expanded US restrictions on Chinese research entities, placing Australian universities inside a broader allied screening regime.
One analyst argues that cutting the partnerships will not materially slow China's scientific progress but could leave Australia outside important research networks.
The policy tension is between security alignment and scientific capacity: universities may lose access to talent and collaboration faster than restrictions change Chinese capability.
For technical companies, the case illustrates how research-security rules can reshape ordinary academic and innovation relationships before they become export-control cases.

Open at SCMP Ask Claude ChatGPT

Chinese chemical subsidiary faces prosecution over alleged pollution

# Policy Jiemian 37m

A court case against Jinpu Titanium's subsidiary could trigger China's major-violation delisting rules if the company is found guilty of environmental pollution.

The subsidiary Nanjing Titanium White was indicted for alleged environmental crimes; the case has not yet gone to trial and the financial impact is unknown.
The company says a conviction for corporate criminal liability could expose the listed parent to mandatory delisting under major-violation rules.
This is a sharp example of environmental enforcement reaching the capital market rather than remaining a fine-and-remediation matter.
The case also shows how Chinese listed-company risk can hinge on the legal status of a wholly owned operating subsidiary.

Open at Jiemian Ask Claude ChatGPT

Extreme weather becomes a Taiwan Strait readiness variable

# Policy SCMP 7h

An analysis argues that heat, typhoons, and other extreme weather could damage equipment, interrupt training, and divert forces toward disaster relief around the Taiwan Strait.

The story treats climate stress as an operational variable rather than only an environmental concern, affecting bases, logistics, communications, and maintenance.
Weather emergencies could consume military personnel and infrastructure capacity even without a conflict, complicating planning for rapid mobilization.
The analysis is about regional military readiness rather than a particular Chinese decision, so its value lies in broadening how Taiwan Strait risk is modeled.

Open at SCMP Ask Claude ChatGPT

The former Zibao network keeps producing Chinese cross-border founders

# Culture Leiphone 4h

A Chinese analysis uses Shenzhen company Hesheng Innovation's 2.99B yuan revenue, roughly $421M, and 260M yuan profit to examine how alumni networks and category timing drive export brands.

Hesheng was founded by former Zibao executives who already understood product selection, supply chains, overseas channels, and operating controls, reducing the usual startup coordination cost.
The company entered a gap between premium offline appliance brands and low-cost online white labels in North America, especially fans, heaters, humidifiers, and tower appliances.
Its formula was not radical invention but product-level engineering, fast marketplace execution, and a complete management team that had already worked together.
The broader alumni network produced more than 15 independent companies with annual revenue above 100M yuan, roughly $14M, according to industry estimates cited by the source.
This is a useful Chinese labor-market story: organizational know-how travels through trusted teams and becomes a repeatable founder pipeline.

Open at Leiphone 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	10/10
36Kr	200	8/30
Solidot	200	11/20
Leiphone	200	20/20
GeekPark	200	2/30
IT Home	200	12/60	+46 over cap
SCMP	200	10/50	+14 over cap
Initium	200	6/15	+1 over cap
Jiemian	200	8/30	+7 over cap
China Digital Times	200	1/15
The Wire China	200	0/50
Techmeme	200	3/15
Hacker News	200	0/20
V2EX	200	8/50	+27 over cap
Zhidx	200	7/20
InfoQ China	200	8/20	+12 over cap
New Wisdom	200	8/15	+2 over cap
TMTPost	200	8/18	+10 over cap

The flood that erased Gyirong has a sequel, and Beijing is trying to compute when it arrives

Gyirong is gone and a lake is filling upstream

Xi lands in Washington on 24 September

Four ministries come for "pseudo-innovation"

GLM-5.3-Flash served 62 trillion tokens on Chinese chips

Shawshank sells out a thousand-seat hall

Threads we are pulling