China Desk — 24 Aug 2026

Xiaomi’s new chips show a strategy beyond smartphones

# GeekPark

Xiaomi unveiled three chips after investing 21 billion yuan (about $3.0 billion) over five years, using stacked memory and dedicated AI silicon to span phones, desktops and cars.

Also: Hacker News

The flagship O3 is a 3 nm mobile SoC with 24 billion transistors; Xiaomi claims 40% higher CPU performance and four times the AI performance versus O4? The supplied article compares it with O1 for some figures, so those claims need careful independent testing.
The more consequential component is O100, which stacks two memory wafers with an NPU wafer through 1.4-micrometer bonding and claims 1.22 terabytes per second of bandwidth, about 16 times the cited LPDDR5X figure.
This targets the memory wall in token generation: moving model weights and KV state can limit latency more than raw arithmetic, so packaging and data locality become first-order architecture decisions.
Xiaomi pairs the chips with a quantized MiMo model and says a prototype reached 330 tokens per second locally, but O100 is not expected to become commercial until next year and cost and yield are unknown.
D100 extends the platform toward vehicles and desktop AI, reportedly supporting up to 160 GB of unified memory and a 200-billion-parameter local model.
The strategic shape is a vertically integrated edge-compute ladder from phone to car to home, where Xiaomi can reuse silicon, models and deployment feedback across its own product surfaces.
The decisive question is whether Xiaomi can turn impressive packaging and prototype numbers into reliable, affordable volume products rather than a one-event technology declaration.

US forum debate questions whether Xiaomi really designed its new CPU Hacker News

Open at GeekPark Ask Claude ChatGPT

A humanoid robot plays live tennis against a professional

# GeekPark

Chinese coverage calls Galaxy General's autonomous televised tennis match an embodied-AI milestone: the robot returned fast balls, recovered from falls and coordinated in doubles.

Tennis tests a continuous loop of perception, prediction, whole-body control and adversarial response; a prerecorded movement sequence cannot handle changing ball speed, spin, placement and opponent behavior.
Galaxy General says its StarBrain model unifies higher-level game understanding, action planning and neural control, while the StarWorks platform converts imperfect human data into trainable examples.
The training strategy combines a virtual tennis environment with millions of multi-agent practice games, then transfers the learned policy to a physical court—a concrete sim-to-real pipeline.
Doubles play adds a social-planning problem: the robot must allocate space, infer a partner's intent and change tactics rather than merely hit the next ball.
The event is still a staged exhibition, so independent footage, scoring, safety constraints and repeatability will determine whether it marks a real capability threshold or an unusually polished demo.

Open at GeekPark Ask Claude ChatGPT

Chinese robot executives debate what blocks 10,000-unit production

# Leiphone

At China’s World Robot Conference, executives disagreed over whether embodied AI is primarily blocked by model reliability, manufacturing quality or weak customer demand.

One camp argued that a 96% task success rate is unacceptable in factories because four failures per 100 jobs can damage equipment or halt a line; they want reliability closer to 99.9% or 99.99%.
Operators countered that existing algorithms can handle bounded tasks and that the harder problems are hardware consistency, environmental reliability, manufacturing yield and the absence of mature NPI standards.
A separate commercial view said robot conversion rates are dramatically worse than those of phones and other consumer electronics, implying that many products still lack a compelling use case.
The debate is important because it separates software readiness from deployment economics: customers care about total cost, maintenance and a two-to-three-year return, not abstract model scores.
The industry’s likely near-term path is vertical systems with measurable ROI, not a general humanoid that can handle every household or factory task.
The panel also exposes a governance question for large companies: a robotics unit needs enough executive ownership and resources to behave like a company, not a small project inside a bureaucracy.

Open at Leiphone Ask Claude ChatGPT

Financial veteran Li Yun appointed Shenzhen's acting mayor

# SCMP

Li Yun, a former banking executive and Guangdong vice-governor, became Shenzhen's deputy Party secretary and acting mayor before the city hosts APEC meetings.

Li spent more than two decades in financial institutions before becoming Guangdong's vice-governor in March 2025, according to the report.
His move places a finance-oriented official at the center of China's technology and manufacturing hub as Shenzhen prepares for high-profile international meetings.
The appointment illustrates how municipal leadership combines Party authority with economic-management credentials; the mayoral role is not an independent electoral mandate.
The timing suggests attention to external presentation, investment confidence and cross-border coordination around the APEC events.
The report does not explain why Li replaced Qin Weizhong or whether the move reflects a broader personnel reshuffle.
For technology executives, the relevant signal is that Shenzhen's economic governance remains tightly linked to senior Party and provincial personnel decisions.
Li's performance will be judged through growth, industrial policy and the city's ability to manage international exposure, not merely meeting logistics.

Open at SCMP Ask Claude ChatGPT

Former education CEO raises $140M for a household robot

# 36Kr

Future Near, founded after China’s education crackdown destroyed its previous company, raised nearly 1 billion yuan, about $141 million, to build lower-cost household robots.

Founder Zhang Yi says the team self-funded more than 100 million yuan, roughly $14 million, built 16 iterations and held cost near 20,000 to 30,000 yuan, about $2,800 to $4,200, before offering home trials through rentals.
Its AVLA architecture decomposes long tasks into specialist models and uses a dispatcher, while a world model predicts candidate action paths before execution.
The company says real homes provide the corner cases factories miss—misaligned furniture, children, changing object locations and clutter—and that those outcomes feed a self-evolving world-action model.
The robot has reportedly entered about 500 Shanghai homes, giving the company a potentially valuable data channel, although the source provides no independently verified success or retention rates.
Zhang’s history matters: the company was founded after the Double Reduction policy forced his education business to cut 100,000 jobs and exit the market, a reminder that Chinese policy shocks can redirect entrepreneurial talent.
The strategic bet is that reliability comes from a modular, risk-screened stack rather than one general VLA; that may sacrifice elegance but fit household deployment better.
The financing surge is itself a risk: if capital markets demand a fast buyback before the household market matures, the company could be forced to scale ahead of engineering readiness.

Open at 36Kr Ask Claude ChatGPT

Nine automakers recall 4.27 million vehicles over emergency-door risk

# Jiemian

China’s market regulator coordinated recalls covering more than 4.27 million vehicles, mostly Tesla models, because emergency mechanical releases may be hard to find after a crash.

The affected companies include Tesla, Xpeng, Geely, Leapmotor, Chery, Xiaomi, Dongfeng, BAIC BluePark and FAW; Tesla accounts for nearly 3 million vehicles.
The notices describe a failure of emergency discoverability rather than a defect affecting normal driving: handles can blend into trim when low-voltage power is lost after a severe collision.
Remedies converge on visible warning labels and over-the-air software changes that lower windows after an accident, showing how physical safety requirements increasingly reach into vehicle software.
The action anticipates China’s mandatory GB 48001-2026 standard, due to take effect in January 2027, requiring clear mechanical release capability and labeling.
For automakers, the strategic message is that a design accepted during normal operation can become a fleet-wide compliance liability when regulators formalize edge-case safety requirements.

Open at Jiemian Ask Claude ChatGPT

Alibaba’s $10.2B AI placement shares the stage with a record NAND IPO

# Jiemian

Chinese coverage pairs Alibaba’s $10.2B AI fundraising with a record proposed NAND IPO, while US coverage focuses on Alibaba’s placement as a major infrastructure bet.

Alibaba plans to raise 80 billion Hong Kong dollars, about $10.2B, and says the proceeds will all support full-stack AI capabilities and infrastructure.
The same Chinese roundup highlights an unnamed domestic NAND leader whose proposed 33 billion yuan, roughly $4.6B, listing would be the largest planned fundraising on the technology board.
The NAND company reportedly held 14 percent of global NAND shipments in the second quarter and entered the global top three, making the listing a capital-markets test of China’s memory ambitions.
The contrast matters: Chinese coverage presents an ecosystem story spanning a platform company and a strategic semiconductor supplier, whereas the US account treats the event primarily as Alibaba financing its AI buildout.

Open at Jiemian Ask Claude ChatGPT

Alibaba's liver-cancer model found 15 missed tumors in clinical use

# IT Home

Alibaba's DAMO LiON CT model found 15 previously missed malignant lesions in a two-month prospective deployment, mostly tumors about one centimeter across.

The model is designed to flag small, low-contrast liver lesions and liver metastases that can be obscured by cirrhosis, fatty liver or competing primary tumors.
The report says it read more than 10,000 contrast-enhanced CT scans and routed disagreements with first-read doctors to senior review or multidisciplinary discussion.
In testing, AI-assisted doctors took 27% less time and improved malignant-lesion sensitivity by 11.5%; the report also says junior doctors reached senior-level performance with assistance.
The architecture combines whole-liver context with local texture and boundary information, then fuses multiple scan phases to catch lesions that appear only briefly across the sequence.
The deployment is more informative than a benchmark because it shows a concrete human-in-the-loop escalation path rather than autonomous diagnosis.
The reported results still come from one two-month hospital deployment, so durability across institutions and disease mixes remains the key unanswered question.

Open at IT Home Ask Claude ChatGPT

Xiaomi unveils three Xring chips for phones, AI and autonomous driving

# IT Home

Xiaomi launched a phone SoC, a high-bandwidth AI accelerator and a 3-nanometer driving chip, making its chip program a broader attempt to supply AI compute across its device ecosystem.

The Xring O3 is described as a 3-nanometer flagship SoC, with a reported benchmark score above five million and a planned debut in the Xiaomi 18 Fold.
The Xring O100 targets AI acceleration with 1.22 terabytes per second of bandwidth, while the Xring D100 is positioned as a domestic 3-nanometer high-compute driving chip.
Lei Jun says Xiaomi has invested more than 21 billion yuan, roughly $3.0 billion, in chip research over five years and has nearly 3,000 chip specialists.
The architecture-level significance is portfolio reuse: one chip family is being positioned across handheld devices, vehicles, homes and factories rather than as a single phone component.
The supplied performance claims come from Xiaomi and state media; independent power, yield, software and production-volume data are not provided.
If Xiaomi can ship all three at scale, it would strengthen China's push from chip design toward an integrated device-and-AI platform.

Open at IT Home Ask Claude ChatGPT

China's EV charging network reaches 23.7 million connectors

# IT Home

China had 23.683 million EV charging connectors by the end of July, up 41.9% year over year, with private installations growing especially quickly and highway coverage above 98%.

Public connectors totaled 5.099 million, while private connectors reached 18.584 million, growing 21.3% and 48.8% respectively.
Public infrastructure had 255 million kilowatts of rated capacity and about 49.97 kilowatts per connector on average.
Public charging volume reached 58.76 billion kilowatt-hours in the first half, up 60.2%, indicating utilization is rising alongside the hardware base.
More than 98% of highway service areas reportedly have charging infrastructure, reducing one of the practical barriers to long-distance battery-electric travel.
China's fifteen-year energy plan targets more than 40 million connectors by 2030, enough claimed capacity for over 110 million EVs.
The private-public split matters: home charging is scaling faster, while public networks still determine whether EVs can serve apartment dwellers, fleets and intercity travel.

Open at IT Home Ask Claude ChatGPT

A new benchmark tries to measure whether AI can do real science

# AI QbitAI

A paper co-authored by Chinese AI-for-science company Deep Principle proposes evaluating complete discovery episodes instead of grading only an AI’s final answer.

The framework scores hypothesis formation, experimental design, execution and interpretation while preserving the full decision trajectory, making failed experiments part of the training and evaluation record.
Deep Principle says its MIRA platform links hypothesis generation, simulation, wet-lab validation and knowledge retention, and ranked first on two disclosed chemistry and drug-discovery benchmarks.
The central architectural shift is from a question-answering assistant to a system that must manage an iterative scientific loop and produce independently verifiable discoveries.
The paper was built with global academic and industry collaborators, so its China significance is the participation of a Chinese company with dry-lab and wet-lab operating data.
The reported benchmark wins come through company-affiliated coverage; independent replication and access to real experimental outcomes will determine whether the standard changes procurement decisions.

Open at QbitAI Ask Claude ChatGPT

Alibaba’s liver-cancer model finds overlooked tumors in clinical use

# AI QbitAI

Alibaba’s DAMO LiON model reportedly found 15 missed malignant lesions in a two-month prospective trial covering more than 10,000 CT patients.

The missed lesions were generally about one centimeter and difficult to distinguish from liver tissue, especially in patients with fatty liver, cirrhosis or postoperative anatomy.
The model combines whole-liver context with local texture and boundary features and fuses multiple CT phases to detect transient pixel-level differences.
Alibaba reports that AI assistance cut reading time 27% and raised malignant-lesion sensitivity 11.5%, with junior doctors reaching senior-doctor performance in the study.
The deployment used disagreement escalation: cases where the model and first reader differed went to senior radiologists and, when needed, a multidisciplinary team.
The result is more consequential than a retrospective image benchmark because the system changed treatment decisions, but the study’s design and independent validation still matter before broad clinical adoption.

Open at QbitAI Ask Claude ChatGPT

DeepSeek adds a low-cost multimodal API

# AI GeekPark

DeepSeek launched an experimental vision API that claims 384 tokens per image, unchanged text performance and near-frontier multimodal agent capability at its existing Flash price.

The model accepts inline images, URLs and uploaded files, while its Files API avoids retransmitting the same image across repeated agent calls—a practical systems improvement for visual workflows.
DeepSeek attributes the low token count to a visual-primitive representation based on spatial coordinates rather than dense image patches, targeting lower context and inference cost.
A community test generated an HTML game from a film still in about 36 seconds, showing the product combination of visual understanding and code generation rather than image recognition alone.
The performance and cost claims come from DeepSeek and a media test, with no independent benchmark or production reliability data in the supplied account.
Adding vision closes a conspicuous capability gap for one of China's most widely watched open-model providers, especially for agents that must inspect screenshots, diagrams and interfaces.

Open at GeekPark Ask Claude ChatGPT

Chinese humanoid robots claim record-breaking athletic performance

# AI Leiphone

At China's humanoid-robot sports event, the Tiangong line reportedly won five golds, two silvers and six bronzes, including a claimed 9.32-second 100-meter run.

The event tests high-speed balance, recovery and gait control, but the reported times exceed human records so dramatically that hardware configuration, course rules and verification matter as much as the headline.
The system combines high-power-density joints and cooling with a large-model brain, world model and lower-level controller, illustrating the industry's move toward layered embodied architectures.
The platform emerged from a consortium involving UBTECH, Xiaomi and other industrial partners, with shared patents and a split between foundational research and commercial deployment.
Athletic demonstrations can expose control limits, but they do not establish useful manipulation, safety or economic performance in factories and homes.
The important signal is institutional: China is using a public sports event as a testbed, recruiting arena and national showcase for a robot platform.

Open at Leiphone Ask Claude ChatGPT

Why phone memory is becoming more expensive than HBM in AI servers

# AI Leiphone

AI racks are pulling low-power LPDDR memory out of phones and into servers, where modular SOCAMM designs trade serviceability against bandwidth and supply pressure.

The cited estimate puts a next-generation Nvidia rack at 124.4 TB of HBM4E costing about $2.5 million and 216 TB of LPDDR5X costing nearly $2.8 million, roughly 12% more.
LPDDR offers high bandwidth per watt, but soldered or tightly integrated designs make capacity changes and field replacement difficult for long-lived servers.
SOCAMM2 adapts a low-profile compression-attached module to make LPDDR more serviceable while preserving short signal paths and thermal headroom.
The system-level constraint matters: GPUs can consume most of a node's power budget, leaving CPU memory to deliver capacity and bandwidth without a large additional energy bill.
Agent workloads intensify the need for CPU-side memory because tool calls, code execution and database access generate repeated, parallel data movement.
The supply-chain implication is important for China as well as the US: AI demand is changing the economics of mobile memory, potentially tightening supply for handset makers and shifting bargaining power toward a small set of memory vendors.

Open at Leiphone Ask Claude ChatGPT

China's Science Exploration Award names 50 young researchers

# Signal Leiphone

The eighth Science Exploration Award gives 50 Chinese researchers five years of unrestricted support, with a record 10 women and the program's first recipient born after 1995.

Each recipient receives 3 million yuan over five years, roughly $423,000, with no stated restriction on how the money is spent.
The winners span 30 universities, hospitals and research institutions in 13 cities; seven institutions received their first awardees, widening the geography of recognized research.
Ten women won, the first time the award reached double digits, while a 30-year-old Beijing Institute of Technology researcher became its first post-1995 recipient.
The award is designed to fund future exploration rather than only reward completed work, giving early-career researchers flexible runway outside ordinary project grants.
The program introduced an interval year for applicants who have failed three consecutive times, an attempt to reduce repetitive application pressure and encourage more selective submissions.
Organizers say international reviewers now make up more than 40% of peer reviewers and that more precise field matching has been added; these are process claims from the award organization.
The award's structure shows how Chinese philanthropy and major technology wealth are being used to build an alternative talent pipeline alongside state funding, while the geographic spread suggests research capacity is broadening beyond traditional centers.

Open at Leiphone Ask Claude ChatGPT

Shengshu Technology maps world models from video to orchestration

# AI Leiphone

Shengshu Technology's five-stage roadmap runs from coherent video generation to interactive worlds, robot action, autonomous agents and multi-agent orchestration.

The proposed model stack combines modality-specific transformer parameters with shared attention so perception, state prediction and action can be trained as one loop.
The company reports a 96.1 RoboTwin 2.0 score, tenfold faster inference than Motus and adaptation to new robot bodies with 50 to 100 human demonstrations.
The roadmap makes an important distinction between a model that renders plausible motion and one that can act, receive physical feedback and preserve state over long tasks.
Levels four and five—autonomous goal pursuit and coordination across robots, people, tools and software—remain research directions rather than demonstrated products.

Open at Leiphone Ask Claude ChatGPT

China briefing: Xi, the economy, discipline and US-China ties

# Policy Sinocism

A Sinocism briefing packages Xi Jinping’s meeting with Jordan’s king, economic conditions, a CCDI plenum, food-safety concerns and US-China developments from Chinese official coverage.

The Jordan meeting readout emphasizes ceasefire, a two-state solution, Palestinian governance and humanitarian responsibility, presenting China as a diplomatic advocate of political settlement.
The briefing also flags July fiscal revenue and a CCDI Sixth Plenum, placing foreign policy beside domestic economic management and party discipline rather than treating them as separate beats.
A formaldehyde-treated cabbage case shows how food-safety scandals remain a recurring test of local enforcement and public trust.
Because this is a roundup, its value is the selection and framing of issues inside Chinese official and semi-official information flows, not a single new event.
The supplied text is truncated, so the economic data and US-China sections cannot be assessed in full.

Open at Sinocism Ask Claude ChatGPT

Chinese AI music startup chooses a native-language training path

# Culture GeekPark

Chinese AI music company Yinchao is rebuilding around Chinese-language musical structure and listener intent as generative-song volume overwhelms discovery systems.

Deezer reportedly receives 75,000 AI songs a day, about 44% of new uploads, but AI tracks account for only 1% to 3% of listening—an output-to-attention collapse.
The company argues that Chinese is not just translated English: tonal pronunciation, phrasing and emotional conventions require different data, objectives and model behavior.
Its team pairs programmers with trained musicians and makes every iteration pass a human listening review, treating perceived musical life as a product metric alongside generation speed.
Version 4.0 reportedly focuses on translating vague emotional requests into usable internal representations, rather than merely generating a song from a short prompt.
Training on domestic Biren GPUs is presented as supply-chain independence, though the source offers no quality comparison or cost evidence against foreign systems.
The broader market question is whether AI music products will compete on controllable authorship and cultural fit once one-click generation becomes abundant.

Open at GeekPark Ask Claude ChatGPT

Chinese team develops a compact real-time X-ray ghost camera

# Signal SCMP

Researchers in China demonstrated tabletop X-ray ghost imaging for high-speed see-through video with potentially lower radiation and simpler hardware than conventional methods.

The technique reconstructs moving objects from correlated measurements rather than forming a conventional frame directly, allowing it to track dynamic structures.
The reported applications include rotating aircraft engines and biological motion such as a beating heart, where motion blur and radiation exposure complicate ordinary imaging.
The source does not provide resolution, frame rate, dose, reconstruction latency or a comparison against clinical or industrial systems.
If the claims survive replication, the architectural advantage would be a smaller sensing setup for dynamic imaging rather than merely sharper still images.

Open at SCMP Ask Claude ChatGPT

Pinduoduo reports 112.4 billion yuan in quarterly revenue

# Signal Leiphone

Pinduoduo reported second-quarter revenue of 112.4 billion yuan, roughly $15.8 billion, and profit of 27.2 billion yuan, while expanding platform governance and supply-chain programs.

The company says non-GAAP research spending reached 4.3 billion yuan, roughly $606 million, up 40%, and that a new Xiongan operation is focused on data processing and industrial modernization.
Its platform governance program now spans product entry, food and drug safety, advertising, intellectual property, livestreaming and AI misuse, shifting from broad rules to category-specific enforcement.
The company says image and keyword databases now cover 47 million images and 9.5 million keywords, with intellectual-property complaints targeted for resolution within 24 hours.
The agricultural strategy uses traffic, logistics and supplier support to scale regional products, showing how Pinduoduo treats platform governance and industrial policy as part of the same supply-chain engine.
All performance, governance and rural-development claims come from the company's earnings communication and are not independently audited in the supplied text.

Open at Leiphone Ask Claude ChatGPT

A Chinese startup bets on immersive real-time video

# AI GeekPark

Chinese startup Xingjie argues that real-time, camera-driven worlds should let the model adapt to users rather than force users to learn prompts and keyboard controls.

The company distinguishes offline video generation, interactive streamed worlds and a third immersive route in which gestures, expressions and posture drive the environment.
Its Rise series aims to make the user an embodied participant, lowering the interface barrier from prompt engineering to ordinary movement and camera input.
The technical and commercial challenge is much harder than producing a short clip: the system must maintain a world state while responding to continuous human actions at interactive latency.
Investors are reportedly betting on games, content and eventual links to robotics, but the startup was founded only recently and the supplied evidence is mostly company positioning.
The broader signal is that Chinese AI-video competition is moving from image quality toward control loops, latency and interface design.

Open at GeekPark Ask Claude ChatGPT

Taiwan reveals Strong Bow missile details while seeking $1.13B

# Policy SCMP

Taiwan disclosed technical details of its Strong Bow anti-ballistic missile as it seeks $1.13 billion for mass production amid pressure from Beijing.

The locally developed system uses hypersonic speed, two-stage solid-fuel propulsion and an active guidance system, according to Taiwan’s government-funded weapons institute.
Releasing specifications while seeking legislative funding is also a deterrence signal: Taipei wants China and outside supporters to see a domestic air-defense production path.
The excerpt does not establish deployment numbers, interception performance or the requested procurement schedule.

Open at SCMP Ask Claude ChatGPT

Chinese government-ship activity near Taiwan hits a third monthly record

# Policy SCMP

Taiwan recorded 118 sightings of mainland coast guard, survey and other official vessels in the first 24 days of August, exceeding July and June totals.

The reported figures were 117 sightings in July and 111 in June, indicating sustained rather than episodic maritime pressure.
The vessels are described as official Chinese government craft, so the pattern represents a law-enforcement and administrative footprint as well as a military signal.
The activity reportedly expanded east of Taiwan after Tokyo and Manila announced maritime-boundary talks, linking the patrol pattern to Beijing’s concern about outside coordination.
Sightings are not the same as an attempted blockade or military operation, but repeated presence can normalize Chinese jurisdictional claims and raise the cost of regional cooperation.

Open at SCMP Ask Claude ChatGPT

Visa says trusted agents could reshape Chinese commerce

# Analysis QbitAI

Visa’s China president argues that AI agents will move from recommending products to making authorized purchases, making identity, permission and liability central to digital commerce.

The essay’s main distinction is between AI that generates content and agents that participate in decisions, browse merchant systems and execute transactions within user-defined limits.
China is presented as an early test market because payments, retail, travel and local services are already deeply digitized and connected through large platform ecosystems.
The proposed control plane has three parts: prove which agent is acting, define and revoke what it may do, and establish shared rules across banks, merchants, platforms and payment networks.
This is a payment-industry argument rather than evidence of a deployed standard, but it identifies the governance layer that must exist before autonomous purchasing can scale.

Open at QbitAI Ask Claude ChatGPT

Alibaba raises $11.3B for AI as China debates layoffs and platform control

# Policy Leiphone

A Chinese roundup pairs Alibaba’s 80 billion Hong Kong dollar AI financing with a reported graduate-labor dispute, police action over a false listing rumor and a warning that China’s auto market is entering its final consolidation.

Alibaba’s first share placement since its 2019 Hong Kong listing is intended to fund AI infrastructure and full-stack capability, making capital allocation itself a strategic signal.
The reported auto-parts labor dispute describes new master’s graduates allegedly being pushed to resign or accept factory-floor roles, reflecting a sharper mismatch between educational credentials and industrial hiring demand.
Shanghai police detained a social-media operator over an unverified claim that Xiaohongshu’s listing had failed, showing how financial rumors can become an administrative-order matter rather than merely a civil dispute.
Nio CEO William Li’s forecast that the surviving auto players will be clear within three to five years reflects the shift from feature competition toward full-stack competition across supply, manufacturing, service and capital.
The roundup also notes that Tesla’s published supervised-driving availability list no longer includes China, a reminder that regulatory access can change independently of product capability.
As an aggregator, the item is valuable for the Chinese framing of AI finance, labor discipline, rumor control and industrial consolidation, but each component needs its own source review.

Open at Leiphone Ask Claude ChatGPT

How Xi's anti-corruption campaign became a central Party policy

# Analysis SCMP

The South China Morning Post's data project examines how Xi Jinping's anti-corruption campaign moved from a purge to one of the Communist Party's central governing policies.

The campaign is presented as more than a sequence of prosecutions: it became an organizational mechanism for discipline, political legitimacy and control over senior officials.
A data-driven treatment can reveal target selection, timing and institutional reach that official case announcements conceal, but the supplied excerpt gives no actual dataset or results.
The strategic question is whether anti-corruption remains primarily punitive or has become a permanent operating system for Party supervision and cadre behavior.

Open at SCMP Ask Claude ChatGPT

Enflame begins its Shanghai STAR Market offering

# AI Leiphone

Chinese AI-chip maker Enflame started its STAR Market issuance after revenue reached about $158M in the first half, using a full-stack approach to challenge the domestic accelerator gap.

Enflame reported first-half revenue of 1.12 billion yuan (about $158M), after 2025 revenue of 990 million yuan (about $139M) and a three-year revenue compound growth rate of 81.32%.
It says 2023 to 2025 research spending totaled 3.676 billion yuan (about $518M), or more than 180% of cumulative revenue, underscoring the capital intensity of domestic accelerator development.
The company has developed four architecture generations and five cloud AI chips, alongside accelerator cards, clusters and a programming platform.
Its GCU-CARE compute unit, GCU-LARE chip-to-chip interconnect and TopsRider software are intended to reduce migration costs from mainstream AI workloads.
Enflame plans to raise 6 billion yuan (about $845M) for fifth- and sixth-generation chips and hardware-software co-design; it expects to reach consolidated profitability in 2026 or 2027.
The strategy is broader than making a chip: cluster deployment and software compatibility are being used to turn domestic silicon into an operational alternative for large AI systems.
The key risk is whether revenue growth can become profitable as product generations, software ecosystems and customer support all consume cash; an IPO gives runway but does not solve compatibility.

Open at Leiphone Ask Claude ChatGPT

US opens Venezuelan telecom investment while excluding Chinese suppliers

# Policy SCMP

Washington is easing Venezuelan telecom sanctions while explicitly excluding Chinese companies, targeting a market where Huawei and ZTE are deeply embedded.

The carve-out gives Western network vendors a policy opening as Venezuela expands connectivity, turning sanctions relief into a tool for supplier substitution.
The measure is a substantive China story because it targets the installed position of Huawei and ZTE rather than merely commenting on Chinese technology.
The long-term effect depends on financing, maintenance and political durability: excluding Chinese vendors on paper is easier than replacing an existing ecosystem.

Open at SCMP Ask Claude ChatGPT

Xiaomi’s humanoid robot returns after two quiet years

# AI Leiphone

Xiaomi publicly showed its 66-degree-of-freedom humanoid robot Tie Da and said factories, not homes, are the near-term target because reliable embodied intelligence remains unsolved.

The demonstration used model-driven autonomous interaction for tasks such as handing over flowers and responding to handshakes, rather than relying on backstage teleoperation.
Xiaomi’s robotics chief says household adoption may be three, five or ten years away because the limiting factor is AI reliability, not only mechanical design.
The company is prioritizing factories, where fixed workstations, measurable throughput and private deployment make robot value easier to validate than in unpredictable homes.
Xiaomi says it began robotics work in 2020 and deliberately reduced public releases after 2023 while continuing internal investment.
The strategy reflects a broader global sequence: use controlled industrial workflows to build data and reliability before attempting the safety, privacy and social demands of domestic robots.

Open at Leiphone Ask Claude ChatGPT

HiDream proposes a native multimodal world model for robotics

# AI QbitAI

HiDream introduced an interactive world model that combines text, images and direct control, claiming long-horizon spatial and physical consistency for simulation and embodied AI.

The model uses a unified-transformer architecture and supports roaming, editing and interaction across realistic and stylized environments.
The proposed data-model-agent-body loop treats internet video, simulated failures and real action feedback as complementary sources rather than assuming any one dataset is sufficient.
The company says the model topped a WBench sub-ranking on its first evaluation, but the source is an authorized presentation and gives no independent test details.
Potential applications include interactive games, robot simulation and structured 3D-scene creation; the hard problem is transferring simulated competence to messy physical environments.

Open at QbitAI Ask Claude ChatGPT

Space thermal-management supplier RayLite raises a Series A-plus round

# Signal 36Kr

Chinese space thermal supplier RayLite raised several tens of millions of yuan, roughly $4 million to $14 million, after three years of in-orbit liquid-cooling operation.

The company says its liquid-cooling systems hold 90% of China's market for computing satellites and that one system has operated continuously for more than three years.
High-power satellites push passive radiators and heat pipes toward their physical limits, making pumped fluid loops a core design constraint as onboard computing reaches tens or hundreds of kilowatts.
RayLite combines pumps, fluid loops, radiators, heat pipes, coatings and system integration, then applies commercial manufacturing methods to reduce space-hardware cost and delivery risk.
The strategic bottleneck is evidence: in-orbit lifetime, sealing and corrosion data matter more than ground simulations because repairs are impossible after launch.
The company expects five high-power computing satellites using its systems by the end of 2026, but that remains a forecast rather than a delivered constellation.

Open at 36Kr Ask Claude ChatGPT

Shein's Everlane purchase faces a US national-security review

# Policy SCMP

US authorities are reviewing Shein's proposed $80 million purchase of Everlane, adding national-security scrutiny to the China-founded retailer's expansion in America.

Also: Techmeme

The review by the Committee on Foreign Investment in the United States shows how data, supply chains and ownership concerns can follow a Chinese-founded company into an apparently ordinary retail acquisition.
The case is not evidence that the transaction will fail; it demonstrates that approval risk is now part of the operating model for Chinese consumer companies buying US assets.
The supplied account gives no public detail on the specific data or governance concerns under review.

Shein sought an unusual CFIUS review for its Everlane purchase Techmeme

Open at SCMP Ask Claude ChatGPT

Shein opens a Hong Kong offering targeting $1.85B net proceeds

# Signal Leiphone

Shein began a Hong Kong global offering expected to raise about 13.12 billion Hong Kong dollars, roughly $1.85 billion, to fund technology, branding, global expansion and corporate responsibility.

The offering covers about 280 million shares at 47.6 to 49.5 Hong Kong dollars each, with trading expected to begin September 1.
Shein reports 2025 revenue of $41.8 billion, net profit of $2.06 billion and service to roughly 273 million active customers across 160 markets.
Its small-batch, automated replenishment model links trend detection, design, supplier production, fulfillment and customer feedback into one data-driven supply chain.
The Hong Kong listing gives a China-founded global retailer a new capital base while it faces scrutiny over labor, sourcing, data and market access in the US and Europe.
The strategic question is whether the company can extend its supply-chain operating system to third-party brands without losing speed, compliance or the low-price proposition.

Open at Leiphone Ask Claude ChatGPT

Hong Kong funded one China today and marked the other one down

Alibaba asks shareholders for the compute bill

Shein is worth $27bn, and the discount is the tariff

Xiaomi's three chips meet a memory squeeze

Shenzhen gets a banker, TikTok pays $400m

A master's degree or the assembly line

A censored post names the mechanism, not the man

Threads we are pulling