China Desk — 21 Aug 2026, 20:34 UTC

Hong Kong court convicts Alliance for democracy leaders

# Initium 13h

A Hong Kong court convicted the Alliance for True Democracy of subversion-related offenses, extending national-security law into an organization’s historical advocacy and political intent.

The court convicted former leaders Albert Ho and Lee Cheuk-yan and treated them as the organization’s directing mind, attributing their intent to the group itself.
The source says the court rejected claims that the defendants lacked hostility toward the Communist Party or believed supporters would avoid illegal means.
Sentencing is pending, with the reported statutory range reaching up to 10 years for serious subversion; the case therefore remains consequential beyond the verdict.
Family members called the ruling unfair and said the defense evidence was not properly considered, while the defendants’ supporters framed the decision as a top-down judgment.
The case shows how Hong Kong’s security framework evaluates political organizations through intent, leadership attribution and perceived ideological effect, not only acts of violence or operational planning.
Its broader signal is a narrowed space for civil-society groups whose historical mission is treated as potentially incompatible with the current political order.

Open at Initium Ask Claude ChatGPT

Alibaba Cloud’s AI boom is making infrastructure profitable

# Leiphone 14h

Alibaba’s results suggest AI demand is lifting cloud revenue and margins, while internal chips, networking and scheduling improvements could turn scale into a durable advantage.

Alibaba Cloud revenue reached 48.737 billion yuan, roughly $6.86 billion, up 45 percent; the source says AI and compute spending is driving both model workloads and broader cloud adoption.
Capital expenditure reached 190 billion yuan, roughly $26.8 billion, by the second quarter against a three-year 380 billion yuan plan, roughly $53.5 billion.
The company says adjusted EBITA margin reached 11.6 percent and expects AI compute demand to remain tight before 2030, though those expectations are management claims.
A 128-card supernode reportedly raises intra-rack bandwidth to petabyte-per-second scale, while scheduling improvements lift throughput 2.5 to 5.6 percent without more hardware.
The commercial thesis is full-stack optimization: proprietary chips, interconnects, storage and scheduling can make a cloud more profitable even when model APIs look interchangeable.
The risk is that sustained utilization, depreciation and power costs matter more than headline demand; the next tests are cash flow and customer concentration.

Open at Leiphone Ask Claude ChatGPT

Alibaba turns its AI portfolio into a full-stack bet

# GeekPark 10h

Chinese coverage says Alibaba’s cloud and AI revenue surged while Qwen, image, speech, video and music models formed a coordinated portfolio rather than isolated benchmark wins.

Alibaba Cloud’s AI and compute services generated 48.437 billion yuan, roughly $6.82 billion, with adjusted EBITA of 5.628 billion yuan, roughly $793 million, up 133 percent.
The model strategy spans a 2.4-trillion-parameter language model, image, speech, video and music systems, shifting competition from a single model to shared training infrastructure and distribution.
Qwen’s reported 3 billion downloads and more than 300,000 derivatives indicate an ecosystem flywheel, although those figures are company claims and downloads do not equal production use.
The strategic thesis is that open models become industrial inputs: chips, clouds, developers and enterprise applications all reinforce the provider’s position.
The important test is not another leaderboard result but whether Alibaba can convert model breadth into durable external revenue and margins.

Open at GeekPark Ask Claude ChatGPT

Typhoon death toll in southern China rises to 159

# SCMP 12h

Chinese authorities raised the official death toll from Typhoon Maysak’s floods from 39 to 159 after six weeks, with 10 still missing and more than 1.65 million affected.

The update came from briefings in Guangxi and was attributed to state media, making information release and disaster accounting part of the story.
The sharp revision raises questions about how casualty data are verified, centralized and communicated after a major disaster.
The source describes historically rare, persistent rainfall, but does not establish the relative roles of climate change, land use or infrastructure failures.
The social impact is large enough to test local reconstruction capacity, compensation systems and public trust in official reporting.
This is a case where state information management is itself a governance signal, not merely background to a weather event.

Open at SCMP Ask Claude ChatGPT

China uses Jiang Zemin’s centenary to signal continuity

# SCMP 8h

Chinese coverage reads Xi Jinping’s memorial for Jiang Zemin as a deliberate effort to place reform-era history inside the Party’s current modernization narrative.

The memorial was held in Beijing on August 17 and attended and addressed by Xi, giving the event more political weight than a routine anniversary.
The framing absorbs Jiang’s era and reform legacy into a continuous Party story rather than presenting it as an alternative political model.
That is useful context for reading elite symbolism in China: historical figures are rehabilitated or bounded according to what they contribute to present legitimacy.
The event also reassures economic and foreign-policy audiences that selected reform language remains part of the official lineage, even as policy control is more centralized.
Its significance will be clearer in subsequent speeches and personnel signals, not from the ceremony alone.

Open at SCMP Ask Claude ChatGPT

China’s reusable rockets could strengthen the PLA’s satellite kill chain

# SCMP 7h

Chinese coverage says recent booster recoveries could make China’s satellite-supported military targeting network more resilient by increasing launch cadence and replacement capacity.

The source identifies recoveries by commercial rocket company LandSpace as a first for China using deployable landing legs after an orbital flight.
Reusable launchers matter militarily because a state could replenish or reconstitute satellite services faster after attack, reducing the value of an adversary’s first strike.
The connection to the PLA concerns the full chain of detection, tracking and engagement, not just launch economics.
Commercial launch firms are therefore strategically relevant even when their immediate customers are civilian; the same industrial base can support national-security resilience.
The military implication remains conditional on turnaround time, refurbishment reliability and actual launch cadence, none of which the source establishes.

Open at SCMP Ask Claude ChatGPT

Southeast Asia is being pulled toward rival AI blocs

# SCMP 5h

Chinese coverage frames Southeast Asia’s AI choice as a contest between a US China-free supply chain and Beijing’s cooperation organization, testing the region’s non-alignment.

The source says Washington is promoting Pax Silica while Beijing wants neighbors to join the World Artificial Intelligence Cooperation Organisation.
This is infrastructure diplomacy, not merely model competition: supply chains, standards, cloud access and political alignment are being packaged together.
For Southeast Asian governments, the practical strategy may be to accept pieces of both systems while preserving room to maneuver rather than formally choosing a bloc.
The story is especially useful because it shows China’s regional AI outreach as an institutional project, not only an export of hardware or applications.
Whether the proposed organizations deliver financing, interoperable standards and usable compute will determine whether the alignment pressure becomes durable.

Open at SCMP Ask Claude ChatGPT

Study says China could meet only 60 percent of wartime drone demand

# SCMP 10h

A study cited by Chinese coverage estimates that China’s industrial system could satisfy only 60 percent of military drone demand in a sudden conflict, exposing supply-chain fragility beneath manufacturing scale.

The model maps assembly plants, component suppliers, inventories, backlogs, budgets and civilian production capacity rather than counting final assembly alone.
The result is a useful warning against treating China’s large commercial drone sector as unlimited wartime capacity.
The estimate is model-based and the source does not identify the researchers, assumptions or validation, so the exact 60 percent figure should not be treated as a forecast.
Its policy value is to redirect attention toward batteries, motors, sensors, electronics and surge logistics, where bottlenecks may be less visible than airframe output.
If the result survives sensitivity testing, stockpiling and dual-use production planning would matter as much as adding military assembly lines.

Open at SCMP Ask Claude ChatGPT

China’s robotics expo shifts from model theater to closed loops

# 36Kr 16h

Chinese reporting from WRC says buyers now care less about parameter counts and more about real work, with world models, fast and slow control loops and deployment data becoming the new industry vocabulary.

The expo reportedly drew more than 300 exhibitors, 3,000 robots and 49 state-owned enterprise groups carrying 12 real engineering scenarios, a stronger demand signal than a purely investor-led show.
Companies are moving beyond single VLA models toward world models, explicit or latent reasoning and closed loops that feed failures back into training.
A recurring engineering pattern is a fast local controller for repetitive motion plus a slower reasoning system for goals, planning and exceptions, reducing latency, cost and heat.
The source describes interfaces between models, navigation, hardware control and enterprise workflows as a central bottleneck, which explains why integration platforms matter as much as foundation models.
The most important commercial change is cultural: Chinese robotics companies are trying to show revenue-producing deployment, not merely intelligence in a controlled demonstration.
Exhibition claims about speed, data volume and success rates remain vendor-reported, so customer retention and intervention rates are the real next tests.

Open at 36Kr Ask Claude ChatGPT

A robot learns new tasks from seconds-long demonstrations

# AI QbitAI 13h

Generalist’s GEN-1.5 reportedly learns unseen tasks from 3 to 12 seconds of physical demonstration, including combining separate demonstrations without fine-tuning.

The model treats recent actions as context, allowing a robot to imitate a demonstrated task and transfer a simulated demonstration to a real machine with zero gradient updates.
Across ten tasks, the company reports 59 percent average success from one physical prompt and 83 percent after five minutes of data plus ten update steps.
That gap matters: the headline capability is not reliable autonomy yet, but the emergence of in-context physical learning from large-scale pretraining.
The architecture is framed as physical prompt engineering, where actions rather than words specify the task and the model fills in transitions and recovery behavior.
The source says these are short, relatively atomic tasks and that in-context learning remains less stable than fine-tuning, limiting immediate production value.
If the capability scales to long-horizon tasks, it could reduce per-site robot training costs far more than another incremental improvement in motion control.

Open at QbitAI Ask Claude ChatGPT

A Chinese robot completes 15 minutes of household work

# AI QbitAI 17h

Moqi’s wheel-based robot completed a 15-minute household sequence without human commands, using a policy model with persistent goals, multi-level memory and progress feedback.

The task combined clearing a table, checking and restocking a refrigerator, moving wet laundry to a dryer and folding it, making it a long-horizon test rather than an isolated manipulation demo.
Moqi’s MoRA architecture puts goal persistence and execution memory inside the fast policy layer, while a slower system handles intent, global planning and unusual failures.
The design addresses a weakness of common agent stacks: repeatedly translating a high-level goal into short text commands can lose context and force the robot to pause for replanning.
The robot’s value depends on maintaining state through navigation, object changes and recovery, so the architecture is more interesting than the 15-minute spectacle itself.
The source gives no success-rate distribution, intervention count or repeatability data, and a trade-show run cannot establish household reliability.
If the fast policy can safely decide when to escalate, this could reduce the cost and latency of using a large model for every physical action.

Open at QbitAI Ask Claude ChatGPT

China gets a 4.2-ton autonomous industrial truck

# AI QbitAI 8h

Ninebot and Yutong launched a production autonomous light truck for factories and logistics parks, shifting Chinese driverless delivery toward heavier, repeatable industrial routes.

The Z20 carries 4.2 tonnes, offers 19.32 cubic meters of cargo space and can handle six to eight standard pallets, according to the company.
Ninebot supplies the L4 autonomy stack, perception, planning, control and fleet dispatch, while Yutong supplies the chassis, manufacturing and quality systems.
The initial target is private industrial infrastructure rather than public roads, where fixed routes, high utilization and measurable labor savings make autonomy easier to justify.
The product includes different cargo and loading configurations, showing that unattended logistics depends on material handling and workflow integration as much as driving.
The company says it has delivered more than 25,000 autonomous vehicles and logged 170 million real-world L4 kilometers, but these figures are not independently audited.

Open at QbitAI Ask Claude ChatGPT

A Chinese robot improves by replaying the same household demo

# AI QbitAI 14h

Qianxun’s unchanged living-room demonstration reportedly reached 85 percent garbage-picking success after a month, illustrating that iteration speed and data closure matter more than novelty at the robotics expo.

The task combines navigation, perception, long-term goals, object handling and recovery, so repeated performance is more informative than a new one-step trick.
The company’s data stack combines web video, first-person, teleoperation and real-machine data; it claims collection cost fell to one-tenth of traditional teleoperation and usable data rose from 30 to 95 percent.
Its central concept is time density: how quickly a failure is identified, reproduced, trained, evaluated and redeployed to the machine.
The source says the system uses an agent for task decomposition, navigation, a base model for action and post-task reports for feedback, making the demo a systems test rather than a single-model test.
The 85 percent figure is company-reported and applies to a selected task, so it should not be generalized to household autonomy.

Open at QbitAI Ask Claude ChatGPT

Minglue and Hikrobot put agents into commercial robots

# AI QbitAI 10h

Minglue and Hikrobot are combining software agents with mobile robots for cleaning, logistics and patrols, betting that organizational integration matters as much as individual robot intelligence.

The demonstrations chain perception, task planning and execution across restaurant cleaning, flexible warehouse picking and patrol anomaly detection.
Minglue’s Octo platform is intended to connect robots, digital agents and legacy IT systems, turning separate machines into a shared organizational control layer.
The partnership pairs Minglue’s VLM and VLA work with Hikrobot’s mobile hardware, motion control, sensor fusion and manufacturing experience.
The strategic claim is that commercial deployment requires shared memory, workflows and escalation paths, not just a stronger action model.
The deployment and training claims come from promotional coverage, so the real test is sustained operation outside a trade-show environment.

Open at QbitAI Ask Claude ChatGPT

RayNeo’s iO glasses add always-on AI memory

# AI QbitAI 11h

RayNeo launched 34-gram AI glasses with two-day battery life, multimodal memory and proactive prompts, making persistent context the product’s main differentiator.

The glasses use a three-layer waveguide, a 1,800-nit in-eye display and a 37-degree-of-freedom? no, the source says a 33-inch equivalent image; the optical claims are company-provided and should be independently checked.
RayNeo’s memory engine processes audio, location and visual context into summaries, reminders and unsolicited conversational hints rather than waiting for a button press.
The glasses initially support DeepSeek and Qwen models, with more models planned, illustrating how Chinese consumer hardware companies are making model choice a replaceable layer.
The privacy problem is central: a device that continuously records and semantically compresses a user’s day needs clear retention, consent and sharing controls, none of which the source details.
The product is strategically interesting because comfort, battery and low-friction input are being treated as prerequisites for an ambient assistant, not afterthoughts.

Open at QbitAI Ask Claude ChatGPT

DeepSeek Science Space puts AI agents inside research workflows

# AI QbitAI 13h

DeepSeek Science Space is a desktop environment that connects literature, data, models, compute and experiment workflows so researchers can delegate execution while retaining scientific judgment.

The system includes domain agents for biology, medicine, drug discovery and materials, with tasks spanning literature review, data retrieval, simulation, analysis, visualization and draft checking.
Its value is orchestration: an agent can reproduce a paper by connecting public data, code and software environments, rather than merely summarizing the paper.
The workflow keeps evidence and intermediate results available for inspection, an essential design choice for scientific reproducibility and for catching model hallucinations.
The product illustrates a broader Chinese AI strategy of embedding agents into existing institutional processes rather than selling a general chat interface.
The source is product-launch coverage and provides no independent success rate, benchmark or evidence that the generated research is scientifically valid.

Open at QbitAI Ask Claude ChatGPT

China rolls out prefabricated AI compute containers

# Signal QbitAI 18h

Taichu Meta launched containerized compute units that it says can be factory-built and commissioned within 24 hours, targeting distributed AI demand outside conventional data centers.

The product line ranges from a 64-accelerator air-cooled container to a 128-accelerator liquid-cooled unit and a 256-accelerator cluster, with a claimed peak of 307 FP16 petaflops for the largest configuration.
The company says factory prefabrication exceeds 90 percent, deployment efficiency improves 70 percent and the units can operate outdoors at sites such as industrial parks and mines.
The architecture bundles compute, networking, storage and cooling, reflecting a move from centralized hyperscale construction toward modular regional capacity.
Its platform is intended to schedule distributed nodes and expose model and scientific-computing services, but the source gives no customer utilization or independent benchmark.
The strategic constraint is not only chip supply: water, power, site access and deployment time are becoming first-order variables in China’s AI infrastructure buildout.

Open at QbitAI Ask Claude ChatGPT

Chinese clouds are discounting rival AI models at a loss

# Analysis Leiphone 13h

Chinese cloud providers are reportedly selling competing models such as GLM, Kimi and DeepSeek at discounts as low as 30 percent, reviving the margin-destroying resale cycle of 2021.

Salespeople cited in the report say identical third-party models leave price as the main differentiator, with discounts falling from 60 to 80 percent toward unprofitable levels.
Cloud companies are accepting outside models because customers, especially coding teams, often demand a named model rather than the provider’s own system.
The practice makes revenue targets look healthier while weakening gross margin and creating conflict between sales teams and model builders.
The historical parallel is important: Chinese cloud providers previously used third-party resale to inflate project revenue, then had to redesign incentives around proprietary product share and profitability.
This cycle could delay domestic model consolidation by giving weaker models distribution, but it also shows that customer choice remains stronger than cloud-brand loyalty.

Open at Leiphone Ask Claude ChatGPT

Trump’s shortened Korea drills push Seoul toward defense autonomy

# Policy Initium 10h

Chinese-language analysis says Trump’s abrupt reduction of US-South Korea exercises may accelerate Seoul’s push for wartime command transfer and more independent defense capacity.

The exercise ended four days early after Trump called it costly and an insult to North Korea, according to the source.
South Korean President Lee Jae-myung responded by linking operational-control transfer and a nuclear-powered submarine program to stronger self-defense.
The episode exposes an alliance-management problem: reducing exercises without advance coordination can undermine the very readiness assessments needed for Seoul to assume more responsibility.
The source frames autonomy and alliance as complementary rather than mutually exclusive, a politically useful position for Seoul amid uncertainty about US commitments.
This matters to China because changes in the US-South Korea command relationship affect deterrence, regional military planning and Beijing’s room for diplomatic maneuver.

Open at Initium Ask Claude ChatGPT

China’s trade surplus keeps widening across most EU states

# Policy SCMP 6h

Chinese customs data cited by Chinese coverage show China’s surplus with 24 of 27 EU members moving higher, sharpening Brussels’ October negotiating deadline.

The source says the surplus expanded with all but three EU members in July compared with a year earlier; Sweden’s monthly gap reportedly quadrupled and Malta’s rose 166.6 percent.
The immediate issue is not simply export volume but Europe’s shrinking leverage: trade negotiations are occurring while the underlying imbalance keeps moving in the wrong direction for Brussels.
The figures come from calculations based on Chinese customs data, so the direction is more robust than any single monthly swing.
A durable response would likely involve market-access concessions, investment screening and industrial policy rather than tariffs alone.

Open at SCMP Ask Claude ChatGPT

A small model beats a larger agent across language environments

# Analysis Leiphone 17h

Research presented at IJCAI argues that agent failures across languages resemble robotics sim-to-real gaps, and that randomized prompts, actions and rewards can make a 3B model outperform a 32B model.

The framework separates observation, state, action, dynamics and reward gaps, adapting domain-randomization ideas from robotics to language agents.
The reported result is striking but underspecified: the source does not give task definitions, baselines, number of languages or statistical significance.
The practical lesson is that robustness may come from exposing agents to environmental variation and uncertainty, not simply enlarging the base model.
The speaker also argues for uncertainty estimation and human-in-the-loop escalation so agents expand capability without expanding supervision burden.
This matters for companies deploying agents across regions: translation quality is only one failure mode when workflows, tools and reward signals change.

Open at Leiphone Ask Claude ChatGPT

China and Indonesia deepen defense ties after naval drill

# Policy SCMP 8h

Chinese coverage says a joint naval exercise east of Taiwan has prompted questions about whether other Southeast Asian states will pursue similar cooperation with Beijing.

The drill was followed by a meeting of the two countries’ foreign and defense ministers, where they discussed military cooperation and food security.
The political signal is not an alliance announcement but a lower threshold for practical security cooperation with China while Indonesia maintains its non-aligned posture.
For neighboring governments, the precedent could make limited exercises politically easier without requiring a formal choice between Washington and Beijing.
The source does not establish whether other states intend to follow, so the immediate story is signaling and regional perception rather than a new bloc.

Open at SCMP Ask Claude ChatGPT

Pop Mart’s global growth hits a sharp slowdown

# Culture 36Kr 9h

Pop Mart’s first-half revenue rose 23.8 percent to 17.17 billion yuan, but overseas sales fell 11 percent and inventory days jumped as the Labubu surge cooled.

Revenue was 17.17 billion yuan, roughly $2.42 billion, and adjusted profit was 5.16 billion yuan, roughly $727 million; profit growth slowed to 9.5 percent.
Overseas revenue fell to 4.972 billion yuan, roughly $700 million, after surging 292 percent in 2025, while inventory days rose from 123 to 201.
Labubu’s family fell to 26 percent of group revenue, while the Xingxingren character rose 580.6 percent to 2.65 billion yuan, roughly $373 million.
The numbers show a consumer brand trying to convert a viral character into a diversified IP portfolio while absorbing the cost of global stores and logistics.
Pop Mart plans 2 to 5 billion yuan of buybacks, roughly $280 to $700 million, alongside dividends, prioritizing shareholder confidence and margin discipline over aggressive expansion.

Open at 36Kr Ask Claude ChatGPT

AliExpress reaches operating profitability through localization

# Analysis Leiphone 16h

Alibaba says AliExpress became operationally profitable as local fulfillment and brand sales expanded, suggesting a shift from subsidized cross-border traffic toward country-level commerce.

The company says local-product sales grew materially and managed overseas services now cover more than 30 countries, with official warehouses in Spain, France and Poland.
Brand buyers exceed 30 percent of the platform, while brand merchandise approached 40 percent of promotional gross merchandise volume, according to the source.
The strategic change is from exporting Chinese inventory through a global funnel to building local supply, warehousing and brand economics in each market.
The figures are Alibaba disclosures and the source does not provide operating margin, customer-acquisition cost or country-by-country profitability.
The result matters because logistics and local assortment, rather than a cheaper marketplace, determine whether Chinese e-commerce can retain overseas users after subsidy-driven growth slows.

Alibaba says AliExpress is profitable Leiphone

Open at Leiphone Ask Claude ChatGPT

China prepares Chang’e 7 mission to search for lunar ice

# Signal Solidot 7h

China plans to launch Chang’e 7 toward the lunar south pole, using a rover and hopping vehicle to inspect permanently shadowed craters for water ice.

The mission is scheduled to launch on a Long March 5 from Wenchang on August 24 and target a November landing near Shackleton crater, according to the source.
Its unusual hopping vehicle is designed to leave illuminated areas after charging and enter shadowed terrain, where conventional solar-powered rovers face a severe energy constraint.
The mission’s value is scientific and infrastructural: confirming the location and accessibility of polar ice would shape later lunar-resource and sustained-presence plans.
The source says the spacecraft will spend about two months preparing in lunar orbit, so the outcome will depend on a long sequence rather than launch alone.

Open at Solidot Ask Claude ChatGPT

Europe’s sovereign AI stack may still depend on China

# Analysis SCMP 10h

Chinese coverage highlights the irony that Europe’s sovereignty push around Mistral AI may rely on Z.ai models, exposing the limits of regional self-sufficiency.

Mistral plans to offer other companies’ models through European infrastructure while presenting the service as sovereign, according to the source.
The strategic contradiction is architectural: sovereign hosting can control data location and operations without eliminating dependence on foreign model weights, chips or research.
Chinese coverage uses the arrangement to question Europe’s ability to separate geopolitical autonomy from practical access to competitive AI components.
Whether this matters commercially depends on licensing, model quality and whether European buyers value jurisdictional control enough to accept a mixed-origin stack.

Open at SCMP Ask Claude ChatGPT

Taiwan tests mobile-network slowdown for civil defense

# Policy Initium 8h

Chinese-language coverage examines Taiwan’s exercise that reduced mobile speeds to 1 percent across 14 counties, affecting nearly 17 million people, and the debate over resilience and legal authority.

The exercise treated network degradation as a civil-defense scenario alongside evacuation and traffic controls, making communications continuity a mass-participation preparedness issue.
The important question is not the drill’s procedure but whether Taiwan can preserve emergency coordination and public access when commercial mobile infrastructure is disrupted.
The source highlights a policy tension between national-security preparation and legal or civil-liberties concerns over deliberately impairing communications.
This is a Taiwan story with a substantive China dimension because cross-strait crisis planning is shaping how the island thinks about network dependence and civilian resilience.

Open at Initium Ask Claude ChatGPT

JPMorgan sees Alibaba Cloud’s mature AI returns rising

# Signal Leiphone 10h

JPMorgan argues Alibaba Cloud’s current 12 percent margin understates the eventual return on AI infrastructure as GPU utilization rises and newer assets mature.

The report models current AI infrastructure at roughly 60 percent utilization, with new asset cohorts producing about 6 percent first-year return and mature assets approaching 20 percent return on invested capital.
Its vintage model says weighted returns could rise from about 6 percent to nearly 16 percent even without better unit economics, while AI infrastructure free cash flow could turn positive around year three.
This is an analyst model, not Alibaba guidance; the key risks are demand durability, accelerated hardware depreciation and domestic chip performance.
The argument matters because it treats cloud AI economics as a cohort-utilization problem rather than assuming today’s low margins are permanent.

Open at Leiphone Ask Claude ChatGPT

Baidu Cloud restructures its partner ecosystem for AI

# Analysis Leiphone 14h

Baidu Cloud is narrowing and reclassifying partners around products, leads, delivery and resale, reflecting a shift from project-by-project cloud channels to repeatable AI solutions.

The company says its first-quarter growth reached 79 percent and that 10 percent of core partners grew faster than the cloud business, but these figures are company disclosures.
Its new partner model tests industry expertise, product thinking and AI collaboration rather than rewarding simple resale volume.
The strategic change recognizes that enterprise AI buyers want a delivered business solution, not just compute or access to a model.
Baidu’s three-year plan moves from restructuring to validating repeatable scenarios and then scaling partner revenue, but one quarter cannot prove the ecosystem is durable.
The interview is useful as a candid description of how Chinese cloud vendors are trying to make partners invest in capability instead of waiting for project referrals.

Open at Leiphone Ask Claude ChatGPT

A robotics founder predicts a market for physical tokens

# Analysis Leiphone 17h

A Chinese robotics company argues that hardware may become a low-margin distribution layer while value shifts to models that charge for successful actions in the physical world.

The proposed progression runs from selling machines to selling packaged solutions and finally charging for effective physical actions, analogous to software token economics.
The speaker defines generalization as training cost: a robot that transfers skills across tasks reduces deployment and onboarding expense for each customer.
The company claims a unified autoregressive model handles zero-shot generalization, general grasping and long-horizon tasks, with reinforcement learning raising some precision to 99.9 percent.
The business model is speculative and depends on cheap, reliable hardware, standardized robot interfaces and customers accepting usage-based pricing for physical work.
It is nevertheless a useful Chinese industry thesis because it exposes the sector’s central economic question: whether intelligence can capture more value than the machine that carries it.

Open at Leiphone Ask Claude ChatGPT

Bilibili relaunches an international app

# Culture Solidot 13h

Bilibili is preparing an English-language international app without identity verification, but its overseas expansion faces the moderation and data-security problems familiar from TikTok.

The source says the company is hiring community managers in Los Angeles, London, Mexico City, Sao Paulo, Istanbul and Tokyo.
The strategic challenge is cultural and regulatory: a Chinese content platform must export its creator graph while satisfying foreign rules on moderation, privacy and data handling.
Bilibili reportedly has 376 million monthly active users on its Chinese service and has invited overseas creators, but the source does not say how much of that audience or content can travel.

Open at Solidot Ask Claude ChatGPT

Kuaishou bets over $169M on a major fantasy MMO

# Culture 36Kr 8h

Kuaishou’s self-developed Lord of Mysteries game launched after reported spending above 1.2 billion yuan, testing whether a declining MMO category can be revived by a major web-fiction IP.

The reported budget is more than 1.2 billion yuan, roughly $169 million, and the project has over 600 staff on Unreal Engine 5.
The game uses a conventional MMO structure modeled on proven Chinese designs while adding an original post-novel storyline and a faithful recreation of the source novel’s plot.
That choice reduces systems risk but creates a strategic compromise: preserve the IP for existing fans without making the product feel like a reskinned incumbent.
The game reportedly topped the iOS free chart after 2.1 million TapTap reservations, but launch ranking is not evidence of retention or monetization.
For Kuaishou, the project is a high-cost test of whether its platform can become a durable games publisher rather than merely a traffic source.

Open at 36Kr Ask Claude ChatGPT

Former TikTok product manager raises nearly $10M for remixable video

# Culture GeekPark 14h

Wapoo, founded by a former TikTok product manager, raised nearly $10 million for a social video platform where users branch and rewrite one another’s clips with AI.

The product treats a published video as the start of a version tree: users can change characters, camera views or plot lines, and each remix remains linked to the original.
Its strategic bet is that relationships, not generation quality, create retention; a friend’s invitation to continue a video is intended to make editing feel like replying to a message.
The founder’s TikTok and Kuaishou experience informs a focus on Gen Alpha users accustomed to Roblox and Minecraft, but the platform launched in the US and has only about ten staff.
The hard problems are moderation, consent, rights to derivative content and controlling generative abuse across many branches, none of which the source resolves.
This is a venture signal rather than a proven product-market fit story, but it shows Chinese founders exporting social-product lessons into a new market.

Open at GeekPark Ask Claude ChatGPT

Former Douyin executive builds AI for 3D manufacturing

# Analysis 36Kr 16h

Former Douyin leader Ren Lifeng raised nearly $50 million for a company combining a 3D-generation model with factory production, choosing supply-chain integration over another consumer AI app.

The company says its Hi3D V3.0 model reaches 2,048-cubed voxel resolution and has built a factory in the Pearl River Delta to test whether generated designs can be manufactured reliably.
The founder reportedly took investors into jewelry factories before accepting capital, emphasizing process data, tooling and production constraints rather than celebrity-founder valuation.
The architecture targets the full path from design generation to printable or manufacturable output, where broken surfaces, tolerances and material behavior expose errors that a screen-only benchmark misses.
This is strategically interesting because China’s manufacturing clusters can provide proprietary feedback data that consumer model companies cannot easily reproduce.
The source is an interview and company announcement; it does not establish model quality, factory throughput or customer revenue.

Open at 36Kr 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	10/10
36Kr	200	9/30
Solidot	200	9/20
Leiphone	200	20/20
GeekPark	200	4/30
IT Home	200	12/60	+44 over cap
SCMP	200	10/50	+11 over cap
Initium	200	6/15	+1 over cap
Jiemian	200	8/30	+12 over cap
China Digital Times	200	0/15
The Wire China	200	0/50
Techmeme	200	4/15
Hacker News	200	1/20
V2EX	200	8/50	+40 over cap

Two verdicts reached past the defendants — to the sons, and to the organisation itself

Life for Xu Jiayin, prison for his two sons

Hong Kong convicts a group, not only its leaders

A typhoon toll quadrupled six weeks late

Yangtze Memory wants $4.6bn at the top of the cycle

Clouds are now losing money selling Zhipu's models

Labubu cooled and the champions took over

Pekingnology audits the phrase "barbarian handler"

Threads we are pulling