China Desk — 24 Sep 2026

DeepSeek reportedly reaches a $1 billion annualized revenue run rate

# Zhidx

DeepSeek reportedly doubled annualized revenue to $1 billion while preparing a 50 billion yuan, roughly $7.0 billion, funding round at a 500 billion yuan, roughly $70.4 billion, valuation.

The reported annualized revenue run rate rose from below $500 million to $1 billion, driven by stronger demand and higher API prices; the figures come from unnamed sources cited by The Information.
The planned financing would be about 50 billion yuan, roughly $7.0 billion, at a target valuation of 500 billion yuan, roughly $70.4 billion, while the article says an earlier round valued the company near 400 billion yuan, roughly $56.3 billion.
DeepSeek's business is unusual among frontier labs because revenue is concentrated in paid API calls while its consumer chatbot remains free and carries no advertising.
The source reports an API gross margin of 82.9 percent and about 4.75 billion yuan, roughly $670 million, in actual revenue during the first seven months, alongside a 715 million yuan loss, roughly $101 million.
More than 70 percent of DeepSeek's compute is reportedly reserved for training and less than 30 percent for inference, showing that the company is still prioritizing model capability over near-term service margin.
The company is testing smaller models on gaming GPUs and seeking to increase domestic-chip use, while Huawei may begin supplying training chips in the fourth quarter.
The strategic tension is clear: demand and pricing are improving, but the next model cycle still consumes scarce high-end compute, so valuation depends on converting research advantage into scalable inference economics.

Open at Zhidx Ask Claude ChatGPT

Tsinghua and Wuwen open-source a cloud platform for robot fleets

# QbitAI

RLark treats robots, cameras, cloud GPUs and cross-region networking as one schedulable system, cutting reported device onboarding from an hour to five minutes and task startup to ten seconds.

The platform's embodied runtime registers physical devices like reusable compute resources, while task-level networking links cloud training with robots and cameras at remote sites.
A single task configuration can assign training to a cloud GPU cluster, inference to another environment and interaction to a physical robot, then establish the required communication paths.
In a reported test spanning Guangdong cloud GPUs and a Beijing robot site, the system completed a collect-train-validate loop for 36 minutes and reached 323 global training steps.
The platform uses cross-cluster isolation and optimized handling for different traffic patterns, addressing a practical weakness of robotics research: network and deployment setup can dominate experiment time.
The project is open source across interface, API, backend, orchestration, networking and runtime layers, which could let Chinese labs standardize experiments instead of rebuilding bespoke infrastructure.
The architecture resembles a control plane for embodied AI, where the unit of scheduling is a complete experiment rather than an individual GPU or robot.
The reported timings come from a test environment, so production behavior will depend on device heterogeneity, network reliability, security and the quality of hardware adapters.

Open at QbitAI Ask Claude ChatGPT

Xi and Trump move from pageantry to an undefined business arrangement

# SCMP

Chinese coverage shifts from military ceremony to business talks, highlighting a new joint arrangement while offering few details on what it changes.

Also: Jiemian

The source says the leaders discussed several irritants after weeks of delay and public posturing.
Xi described the new arrangement as good news for businesses and the global economy, but the account does not identify its provisions.
The contrast between high symbolic confidence and low policy specificity is the main signal for companies watching trade and technology rules.
This is a summit-level story rather than a standalone commercial agreement; implementation will determine whether it matters.

Open at SCMP Ask Claude ChatGPT

Baidu's long chip bet reaches the cluster era

# GeekPark

Baidu's Kunlun accelerator program has progressed from internal cost control to mass deployment and a 30,000-card cluster, showing how long-cycle Chinese infrastructure bets mature through real workloads.

Baidu began exploring accelerators around 2010 as imported chips became a major operating cost, then used search, speech and vision workloads as a continuous test environment.
The first Kunlun chip appeared publicly in 2018, the second reached mass production in 2021 and the third entered production for energy, industry, transport, finance and internet customers.
Baidu says tens of thousands of chips have been deployed and that it built China's first fully self-developed 30,000-card cluster, shifting the engineering problem from chip performance to communication, scheduling, recovery and software compatibility.
The current architecture links Kunlun hardware, Baidu Cloud, foundation models and agents into a feedback loop in which application workloads expose problems that can flow back into the platform and chip.
Baidu awarded $1 million each to teams working on a trillion-parameter MoE supernode and enterprise AI employees, signaling that infrastructure and agent deployment are now parallel strategic priorities.
The story is a useful counterpoint to short AI funding cycles: the differentiator is not merely owning a chip design but accumulating years of workload feedback and systems integration.
The remaining question is external commercial scale, especially software portability and whether Kunlun can win workloads outside Baidu's own ecosystem.

Open at GeekPark Ask Claude ChatGPT

Xi-Trump summit opens with a search for practical stability

# SCMP

Chinese researchers quoted in summit coverage emphasize cautious, practical consensus over headline deals, while the visit’s ceremony and strategic framing dominate US-facing discussion.

The Chinese framing treats a new stability mechanism as more important than a dramatic bargain, reflecting expectations that structural rivalry will persist after the meeting.
The US-facing package emphasizes the red carpet, domestic political optics and whether the leaders can produce visible deliverables.
That contrast matters because Chinese policy discussion often evaluates summits by whether they preserve room for maneuver, not by whether they announce a final settlement.
The source is a roundup rather than a single reported event, so its Chinese-researcher framing should not be read as an official negotiating position.
The summit’s practical test is whether working-level channels and trade commitments survive after the ceremony.

Open at SCMP Ask Claude ChatGPT

Alibaba's 20-gigawatt cloud target creates a utilization test

# TMTPost

Alibaba plans to operate more than 20 gigawatts of global data centers by 2032, turning its AI strategy into a multibillion-dollar utilization and cash-flow bet.

The source estimates Alibaba currently operates roughly 4 to 6 gigawatts and has committed 380 billion yuan, roughly $54 billion, to AI and cloud infrastructure over three years.
It reports second-quarter capital spending of 67.7 billion yuan, about $9.5 billion, against 48.4 billion yuan, roughly $6.8 billion, in AI-cloud and compute-service revenue.
The company says it can recover compute investment in about three years if demand and margins hold, but that assumption depends on sustained utilization and pricing.
Alibaba's Zhenwu V900 chip claims 216 GB of memory and 1,200 GB per second interconnect bandwidth, though full performance, power and independent testing were not disclosed.
The strategic comparison with Huawei is revealing: Alibaba intends to keep compute assets on its own balance sheet, making demand risk its problem rather than the customer's.
The real metric is not installed gigawatts but effective tokens or workloads per unit of power, capital and cooling.

Open at TMTPost Ask Claude ChatGPT

Xiaomi's HySparse2 cuts long-context compute for agents

# AI Zhidx

Xiaomi's HySparse2 architecture reportedly cuts long-input prefill compute to one-fifth and KV-cache use to about one-fourth of the previous design at million-token context.

The architecture separates a self-decoder from a cross-decoder and reuses key-value states across attention layers, allowing the long-input path to stop earlier.
Token-level retrieval replaces block-level selection, so an agent can recover a relevant token from an old tool response without processing its surrounding block.
The design targets a specific agent workload pattern: short actions followed by long tool outputs, repeated over many turns.
The source says tests improved long-context retrieval while reducing cache use, but it provides no independent latency or quality measurements.
The release shows Chinese model teams competing on inference architecture and memory economics, not only on benchmark scores or training scale.

Open at Zhidx Ask Claude ChatGPT

A Chinese robotics startup treats data alignment as the real bottleneck

# AI QbitAI

Lingchu Intelligent is using a model to convert human demonstrations into executable robot data, arguing that training value depends on action alignment and task diversity, not raw hours.

The company distinguishes weakly paired data, where human and robot perform the same task, from strongly paired data, where scenes and timing align and the robot's action can replay successfully.
Its counterintuitive method starts with real robot trajectories and generates corresponding human-hand data, then trains a converter that maps ordinary human videos back into robot images and actions.
The converter is not a control policy; it is a video-and-action transformation model whose output is validated by replaying the trajectory on a robot and by using it for post-training.
Lingchu says it reduced repetitive examples, expanded task diversity and used automated labeling to break demonstrations into atomic actions rather than counting hours as a proxy for quality.
The same conversion capability can provide in-context demonstrations at runtime, letting a robot adapt to a new task without updating model weights.
The architecture separates long-horizon task decomposition from low-level action generation, which is a practical way to manage the gap between human instruction and robot morphology.
The claim that complex tasks remain unsolved is important: data conversion improves the teaching interface, but it does not eliminate contact dynamics, perception errors or safety failures.

Open at QbitAI Ask Claude ChatGPT

A Chinese compute firm sells efficiency as a token factory

# AI QbitAI

China Brain says its BitaHub platform coordinates power, heterogeneous chips, models and workloads, reporting more than 5,000P of connected compute.

The company positions itself above ordinary data-center rental by managing the conversion of electricity and compute into tokens that meet quality, latency and reliability targets.
Its architecture uses one decision layer and three subsystems for compute, tokens and power, with monitoring, forecasting and scheduling at 15-minute and hourly horizons.
The scheduler separates latency-sensitive requests from flexible batch work, while plugins profile domestic and foreign chips and assign prefill and decode to hardware suited to each stage.
A three-city test reportedly achieved 100% cross-domain migration success and 98% energy-forecast accuracy, but the metrics are company disclosures.
The strategic proposition is utilization: when accelerators are heterogeneous and power prices vary, orchestration may matter more than owning the newest chip.

Open at QbitAI Ask Claude ChatGPT

Knowin says one demonstration can teach a robot across tasks

# AI QbitAI

Chinese startup Knowin's GLOW architecture combines a multimodal autoregressive model, synthetic experience, world prediction and task memory to transfer one human demonstration across objects and settings.

GLOW represents the human demonstration, environment, robot state and execution history in one sequence, aiming to capture object relations, action order and state changes rather than imitate pixels alone.
Its core KnowinGLOW model unifies perception, spatial reasoning, planning and action generation, while KnowinDream supplies varied synthetic physical experience and KnowinWorld predicts consequences before execution.
KnowinAgent manages task memory, tools, feedback and replanning, giving the system a runtime harness around the model rather than treating the policy as a single forward pass.
The demonstrations include changing boxes and watering cans, reproducing a hand-drawn wiping pattern and executing a multi-step cocktail workflow.
This is the right technical target for generalization: preserve task intent while adapting grasp points, paths and contact behavior to new objects and scenes.
The reported benchmark wins come from the company, and the excerpt provides no independent evaluation, model scale or failure distribution.
The deeper signal is convergence between foundation-model design and robotics control: reusable physical experience and consequence prediction may matter more than a larger language model alone.

Open at QbitAI Ask Claude ChatGPT

Software tuning makes PCIe GPUs far more useful for DeepSeek inference

# Signal QbitAI

A Chinese infrastructure provider reports raising DeepSeek inference throughput from 1,932 to 13,274 tokens per second on eight PCIe-only GPUs through kernels, communication and cache tuning.

The reported 6.87-fold gain comes without changing model architecture: the system restores sparse-MLA and FP8 fast paths, rewrites communication for PCIe and tunes prefill and decode separately.
The bottleneck is architectural mismatch. Framework defaults assume high-bandwidth interconnects such as NVLink, so long-tail hardware can silently fall back to generic kernels and waste accelerator capacity.
The provider reports a 1.55-fold gain on another DeepSeek model and a 1.92-fold gain on GLM, while extending one context limit from 270,000 to 1.05 million tokens.
The broader lesson is that China’s accelerator shortage creates a large software-optimization market: usable inference capacity depends on kernel coverage, memory allocation, parallelism and cache reuse as much as peak silicon.
The vendor claims 20,000P of managed compute and 99.95% availability, but the supplied comparison is vendor-reported and should be tested against independent workloads.

Open at QbitAI Ask Claude ChatGPT

China's L3 AI-device standard shifts the test from answers to outcomes

# Policy Zhidx

A new Chinese AI-device grading standard makes cross-app task completion the key test, with Lenovo's Y900 tablet cited as the only first-batch L3 product.

The representative task is buying a previously purchased movie ticket: the system must infer intent, retrieve history, navigate an app, handle permissions and stop for confirmation before payment.
This is a more demanding metric than summarization or image generation because errors change real-world state and compound across multiple steps.
The architecture requires controlled system tools, permission boundaries, exception handling and result verification; visually clicking through interfaces is not enough.
Lenovo describes a device-edge-cloud scheduler and MCP-based system-service access, allowing private or low-latency work to stay local while complex reasoning moves to the cloud.
Cross-device continuity adds another state problem: the system must preserve what has been confirmed, which step is current and which device should continue, not merely move a file.
The standard therefore acts as a policy and market signal: Chinese regulators and vendors are beginning to define terminal intelligence by reliable delivery, not model branding.
Whether L3 becomes meaningful depends on repeatable test suites, failure reporting and protection against vendors optimizing narrowly for canned demonstrations.

Open at Zhidx Ask Claude ChatGPT

Nio paper lets autonomous driving reason over multiple futures

# AI QbitAI

A Nio-led paper proposes generating several coupled future-scene and action hypotheses, allowing an autonomous system to score possible consequences before selecting a trajectory.

MM-Future combines world and action modeling so a vehicle’s planned motion can change the predicted environment, while the predicted environment can feed back into the motion plan.
Unlike cascaded systems, it keeps each candidate trajectory paired with its own future scene; unlike many joint systems, it generates multiple candidates rather than one.
The design uses Gaussian-mixture action noise, independent scene noise and best-of-many supervision to preserve diversity without requiring multiple ground-truth futures from a single driving log.
Compact MM-Tokens retain road structure, intersections and nearby traffic while avoiding full high-resolution video generation for every candidate, controlling inference cost.
A modality-aware transformer lets scene and action streams interact while retaining modality-specific statistics, and a future-conditioned scorer ranks each trajectory using its paired predicted future.
This is a technically meaningful Chinese autonomous-driving contribution because it treats uncertainty and counterfactual planning as a coupled representation problem rather than just adding more trajectory samples.
The supplied excerpt reports benchmark results but not enough of them to judge real-world safety or deployment readiness.

Open at QbitAI Ask Claude ChatGPT

Simate uses AI to research the next generation of robots

# AI Zhidx

Chinese startup Simate is building a human-supervised research loop in which agents modify models, run experiments and evaluate robots, calling the approach physical recursive self-improvement.

Simate says its general physical fast-system model topped the RoboDojo embodied benchmark after only three months of company operation, but the claim is not independently substantiated in the supplied report.
The model combines four-dimensional physical perception and memory to track depth, geometry, motion, contact and task history during long sequences such as making tea.
Its AutoResearch system reads code and experiment history, changes configurations, launches training and evaluation, then chooses whether to continue or revise the next experiment.
This resembles AI-assisted software research but adds the cost and safety constraints of real hardware, where failed experiments can damage robots, corrupt data or hide distribution shifts.
The proposed Physical RSI loop links model design, data generation, evaluation and deployment so each research cycle contributes experience to the next one.
The near-term value is automation of experimental execution, not autonomous scientific judgment; humans still set hypotheses, constraints and go-or-stop decisions.
If the loop scales, the defensible asset may be the experiment infrastructure and accumulated physical feedback rather than a single model checkpoint.

Open at Zhidx Ask Claude ChatGPT

PixVerse turns video generation into a persistent interactive world

# AI Zhidx

Chinese startup AIsphere's PixVerse R2 lets users move through generated worlds and alter them continuously, shifting the product target from rendered clips to stateful interactive simulation.

R2 accepts keyboard movement and natural-language interventions that change scenes, characters and story outcomes without restarting a generation job.
The important architectural requirement is continuity: the system must preserve world state, understand user actions and produce audio-visual updates fast enough to feel interactive.
AIsphere describes an Omni Causal autoregressive model and a real-time acceleration layer, building on an earlier system that already supported ongoing user input.
The product sits between video generation and games: developers supply less fixed content, while the model must assume more responsibility for rules, memory, consistency and latency.
The demos are compelling but do not establish frame rate, cost per minute, state persistence limits or safety controls for open-ended user actions.
If the approach scales, the competitive moat shifts from one-shot visual quality toward world models, runtime infrastructure and tools for authoring durable interactive spaces.

Open at Zhidx Ask Claude ChatGPT

Chinese model labs are retraining around data quality

# Analysis Leiphone

Chinese AI companies are reportedly redoing pretraining after discovering duplicate, synthetic and poorly labeled data, shifting the competitive focus from raw corpus size to data governance.

The reported failures include repeated machine-generated pages, contaminated evaluation sets and unclear annotation rules that wasted large training runs.
The engineering lesson is organizational: data teams and pretraining teams need shared quality and capability metrics rather than separate quotas for corpus volume.
The article also describes a shadow market in which API intermediaries collect user interaction traces for distillation, raising consent, privacy and competitive concerns.
High-quality expert and long-horizon agent data remains scarce and expensive, while Chinese buyers reportedly pay far less than US benchmarks for comparable expert work.
The trend matters because a larger model and more accelerators cannot recover information lost to duplication, contamination or weak labels.

Open at Leiphone Ask Claude ChatGPT

US and China extend their trade truce by two months

# Policy SCMP

Chinese coverage says Beijing and Washington extended the trade pause to January 10 before the summit, preserving a short runway for compliance checks rather than resolving the dispute.

The extension gives the US time to assess Chinese commitments while allowing both governments to avoid a tariff escalation during the leaders’ meeting.
The source says China sought a much longer pause, indicating that Beijing’s preferred outcome was stability through the remainder of the US presidential term rather than a narrow tactical delay.
The short duration leaves export controls, technology access and enforcement leverage intact; it is a truce mechanism, not normalization.
The most important signal is institutional: both sides still prefer managed bargaining to an uncontrolled trade shock, even while the underlying rivalry remains.

Open at SCMP Ask Claude ChatGPT

China's drive-by-wire chassis moves into the mainstream price band

# Policy Leiphone

New Chinese vehicle launches and a national standard are pushing steer-by-wire, rear-wheel steering and full digital chassis control toward mass-market adoption rather than flagship experimentation.

A national steering standard took effect in July, while IM Motors made a full drive-by-wire chassis standard on a vehicle starting around 197,900 yuan, roughly $27,900.
The architecture removes the fixed mechanical connection between steering wheel and road wheels, enabling redundant electronic control and smoother handoff between human and automated driving.
IM reports a 20-millisecond steering response and a three-layer redundancy design in which rear steering and braking can help stabilize the vehicle if front steering paths fail.
Rear-wheel steering also changes daily usability: the large LS6 reportedly achieves a 4.49-meter turning radius, making a nearly five-meter SUV maneuver more like a smaller car.
The strategic value is hardware readiness for higher automation. Software updates can improve behavior, but a vehicle without the necessary actuation and redundancy cannot be upgraded into the same autonomy class.
The source is strongly product-promotional and relies on the automaker's test claims, so real-world failure rates and regulatory acceptance remain open questions.

Open at Leiphone Ask Claude ChatGPT

Stripe brings agent payments and global billing tools to China

# Policy QbitAI

At its first China Stripe Tour, the US payments company positioned cross-border billing, risk and agent wallets as infrastructure for Chinese firms that are global from day one.

Stripe says payment volume among its Greater China users grew about 48 percent from 2024 to 2025, while the number of new AI companies grew 78 percent.
The company frames Chinese AI and software firms as increasingly born-global: they need local payment methods, currencies, tax and risk controls before they have a mature domestic business to export.
Stripe's agent-commerce pitch anticipates software agents discovering and buying products, requiring authorization, identity, payment and risk controls distinct from ordinary checkout.
Examples such as MiniMax integrating with two engineers and Shoplazza serving merchants across more than 180 markets illustrate the appeal of managed global infrastructure, though they are company-selected case studies.
The China dimension is substantive because payment access and compliance are part of the export stack for Chinese AI companies, especially when domestic platforms do not provide a uniform route into foreign markets.
The strategic question is whether foreign financial infrastructure can remain reliable for China-linked businesses amid sanctions, data controls and geopolitical scrutiny.

Open at QbitAI Ask Claude ChatGPT

iFlytek trains a context-aware speech model entirely on domestic compute

# AI Zhidx

iFlytek’s Spark-ASR 2.0 combines non-autoregressive transcription with selective language-model correction, supporting 202 dialects and targeting all-domestic compute deployment.

The architecture first produces a fast full-sequence transcript, then uses speculative checks to spend extra computation only where context, ambiguity or formatting requires it.
iFlytek reports support for dialects, code-switched technical terms, noisy environments and low-volume speech, with cleanup of repetitions, corrections and structured numbers.
The model builds on a prior non-autoregressive system that reportedly improved quality 16% while cutting inference cost 84%; the new system adds language understanding while claiming only a 10% reasoning-cost increase.
Deployment through iFlytek’s keyboard, API, smart glasses and office devices makes this an infrastructure and product-stack story, not just a benchmark release.
Training on domestic compute is strategically relevant because speech workloads are high-volume and latency-sensitive, making them a practical test of whether China’s accelerator stack can support everyday AI services.

Open at Zhidx Ask Claude ChatGPT

China's industrial surplus puts trade partners on the defensive

# Policy SCMP

Chinese commentary asks how its industrial competitiveness can benefit trading partners as the EU weighs protection against a goods surplus nearing $1.2 trillion.

The European Commission is pressing Beijing for progress on the trade imbalance and has raised the prospect of new protective measures.
The article frames the issue as a distribution problem: China's manufacturing strength creates cheap supply, but also political pressure in importing economies.
This is not simply a dispute over market access; it tests whether China can export industrial capacity without provoking coordinated restrictions.
Any durable answer would require clearer demand-side rebalancing or more credible access for foreign firms, neither of which the source says has been agreed.

Open at SCMP Ask Claude ChatGPT

China's next aircraft carrier may be built for longer-range operations

# Policy SCMP

Satellite imagery suggests a Dalian carrier under construction could exceed the US Navy's largest ship, raising questions about Beijing's logistics and operating radius.

The imagery reportedly shows work beginning on the ship's bulbous bow, allowing analysts to infer dimensions before launch.
The strategic issue is not length alone but whether the hull, aviation systems and support fleet can sustain operations far from Chinese bases.
The supplied text provides an intelligence interpretation rather than official specifications, so the projected size and mission remain uncertain.
A longer-range carrier would matter most if China also expands replenishment, overseas access and carrier-air-wing integration.

Open at SCMP Ask Claude ChatGPT

A Chinese robotics researcher bets on self-improving embodied AI

# AI Leiphone

SeeAct AI founder Mu Yao argues that robots must move from supervised training toward reward-driven self-improvement, building on his RoboTwin benchmark and simulation work.

Mu's RoboTwin project is described as a widely used Chinese embodied-AI benchmark and data generator, with more than 2,900 GitHub stars and adoption by major model teams.
His technical path runs from end-to-end policy and reinforcement learning through simulation, code-generating agents and vision-language-action models, rather than treating each as a separate product.
The central thesis is that robots need experience and feedback, not just more static demonstrations; reward-driven loops can connect physical outcomes to the next policy update.
RoboTwin's evolution toward support for domestic GPUs and physics engines also illustrates an effort to make Chinese embodied research infrastructure less dependent on one hardware stack.
The argument remains a research direction, not proof that autonomous physical self-improvement is safe or economically viable.
For engineers, the useful distinction is between an agent that generates task plans and a policy that learns from real-world consequences; Mu is betting the latter becomes the long-term bottleneck.

Open at Leiphone Ask Claude ChatGPT

Alibaba treats data, models and compute as one AI system

# AI InfoQ China

Alibaba Cloud is redesigning data infrastructure around heterogeneous CPU, GPU, storage and token resources because robotics and agents make data production part of the AI loop itself.

The central change is conceptual: data is no longer only a pretraining input but a continuous flow of observations, tool traces, feedback and evaluation results that feeds the next iteration.
Embodied data pipelines combine video, point clouds, joint trajectories and force signals, so preprocessing can consume CPUs, GPUs, model tokens, storage and network bandwidth at once.
Alibaba's MaxCompute AI engine puts those resources into a unified scheduling model and exposes AI operations through SQL and Python interfaces rather than forcing every team to build a bespoke pipeline.
The cited Quancher Robotics pipeline reportedly achieved more than tenfold higher throughput with elastic compute in the hundred-thousand-unit range, though the source does not define the baseline or workload.
This is an infrastructure signal with broad implications: the bottleneck is shifting from owning enough data to filtering, labeling and validating existing data quickly and cheaply.
For robotics companies, elastic resource release may matter as much as peak capacity because task definitions and labeling rules change faster than fixed pipelines can absorb.
The architecture increasingly resembles a data operating system for agentic workloads, where model training, inference and data processing compete for the same schedulable resources.

Open at InfoQ China Ask Claude ChatGPT

Anthropic reportedly tightens Claude access for China-linked cloud users

# Policy Leiphone

Chinese reporting says Anthropic's latest compliance sweep suspended many Claude accounts sold through AWS and Microsoft channels in Greater China, underscoring the fragility of indirect access.

The reported enforcement targets customers whose actual use appears tied to restricted countries or regions, even when the cloud account was opened through another geography.
This is more than an account-support issue: it shows how model providers are pushing geographic compliance down through distributors, identity signals and usage patterns.
The article estimates possible lost revenue for the two cloud channels at several hundred million dollars, but that figure is sourced to unnamed people and should not be treated as verified.
Chinese customers are reportedly reconsidering payment when service continuity is uncertain, which creates a trust and procurement problem for foreign AI vendors in the region.
The broader implication is market fragmentation: access to frontier models may depend on jurisdiction, reseller structure and auditability, not only technical demand.
Anthropic's enforcement also raises a strategic question for Chinese firms: whether to build workflows around foreign models whose policy perimeter can change without a local contractual remedy.

Open at Leiphone Ask Claude ChatGPT

Chinese coverage tracks the Xi-Trump summit through AI and ceremony

# Policy SCMP

Chinese coverage of the Xi-Trump summit packages AI cooperation, military pageantry and a possible joint arrangement as evidence of managed engagement, rather than a clear strategic reset.

Also: Techmeme

The roundup foregrounds Xi and Trump's remarks at the White House and treats AI cooperation as one of the summit's most consequential technology themes.
Its emphasis on a possible new joint arrangement and CEO attendance shows that Chinese-facing coverage is watching both state diplomacy and the business channel around it.
The account does not establish a signed technology deal; the value lies in what Chinese media choose to elevate as the summit's signal.
Compared with US-origin coverage, the Chinese framing gives more weight to ceremony, relationship management and the continuation of contact.

Open at SCMP Ask Claude ChatGPT

Deep Blue Space raises nearly 2 billion yuan for reusable rockets

# Signal 36Kr

Chinese launch company Deep Blue Space raised nearly 2 billion yuan, roughly $280 million, to develop reusable liquid-fueled rockets and target a 2027 first flight for its 25-ton system.

The company is developing two liquid-oxygen and kerosene rocket families, including a 4-to-6-ton vehicle and the 25-ton Nebula-2 aimed at large low-Earth-orbit constellations.
Nebula-2's first stage uses 11 engines, allowing the booster to use only one to three engines during landing for finer thrust control than a single very large engine.
The engineering risk is system integration: synchronized starts, millisecond-scale fault isolation, thrust balancing, vibration and thermal coupling across the engine cluster.
Deep Blue plans a tower-capture recovery method, making ground infrastructure and repeatable engine production as important as the first launch.
The financing reflects China's shift from proving private launch technology to building industrial cadence, although a 2027 flight remains a plan rather than a milestone.

Open at 36Kr Ask Claude ChatGPT

Chinese companies are starting to count agents, not headcount

# Analysis GeekPark

Chinese enterprise-AI strategists argue that real AI-native organizations should be measured by how much work agents close end to end, not by token consumption or chatbot adoption.

Alibaba reports token-related managed-AI revenue of 16 billion yuan, roughly $2.3 billion, and more than 40 billion yuan, roughly $5.6 billion, including AI cloud, though the figures are company disclosures.
The useful maturity model is employee-level experimentation, agents embedded in private data and workflows, then reusable organizational capabilities distilled into digital assets.
Examples span a small electronic-design team using multi-agent workflows and a large accounting firm placing an AI layer over existing systems, suggesting transformation need not mean replacing legacy software or staff.
The economic threshold is important: executives cited a need for roughly three- to fivefold productivity improvement before the benefit exceeds migration and replacement costs.
This reframes enterprise AI from assistant adoption to process ownership, where permissions, data quality, auditability and exception handling determine whether an agent can actually close work.
The Chinese context is especially relevant because many firms have large manual coordination burdens and mature but fragmented software stacks, creating both opportunity and integration cost.

Open at GeekPark Ask Claude ChatGPT

Intel recasts the CPU as the control plane for agent workloads

# AI InfoQ China

Intel China argues that agent systems will shift data-center design toward CPU orchestration, GPU token generation and storage-aware scheduling.

The proposed architecture separates CPU clusters for agent execution and task coordination from GPU clusters for model computation and storage clusters for context and business data.
Intel says agents add tool calls, context engineering and repeated verification, increasing the amount of system work that is not matrix multiplication.
Its disclosed tests claim a Xeon 6 Plus handled more concurrent sandboxes and delivered higher single-agent performance than a comparison platform, though the benchmark setup is not fully specified.
The strategic pitch is that agent infrastructure may move CPU demand away from simple host duties toward a control plane for heterogeneous accelerators.
Intel's market forecasts are company claims and should not be treated as an independent market estimate.

Open at InfoQ China Ask Claude ChatGPT

Xianyu faces another crackdown on sexual lead-generation schemes

# Culture TMTPost

A television investigation found sellers using ordinary service listings to direct users toward off-platform sexual networks, including alleged exploitation of minors.

Xianyu denies that the most serious chat screenshots occurred on its platform, but acknowledges that criminals use it as a lead-generation entry point and says it reported the matter to police.
The platform says it removed six of ten cited listings before the report and three afterward, suggesting automated detection still missed or failed to act on some risky items.
The abuse pattern is adaptive: sellers hide behind normal categories, coded phrases and off-platform contacts, creating a moderation problem that keyword filters cannot solve cleanly.
Xianyu reports freezing 98,358 risky accounts this year, but repeated scandals show that enforcement volume is not the same as prevention.
With 234 million monthly active users and rapidly expanding commercialization, the case tests whether governance investment is keeping pace with platform scale.
The controversy also shows how China's online marketplaces blur commerce, personal services and social interaction, creating unusual safety obligations.

Open at TMTPost Ask Claude ChatGPT

A robotics IPO exposes the gap between the body and the story

# Analysis TMTPost

Mech-Mind's first post-IPO accounts show 93% of revenue still comes from industrial 3D vision, while its brain and hand products contribute under 1.1%.

The company's IPO drew 3,835 times subscription, but its revenue mix remains dominated by machine-vision products sold into factory automation.
The source says 94.6% of first-quarter direct customers were systems integrators, putting the company largely one layer removed from humanoid-robot end users.
This exposes a recurring Chinese robotics pattern: established automation suppliers can rebrand as embodied-intelligence leaders before the new products produce meaningful revenue.
The broader Hong Kong listing wave has attracted extreme oversubscription and high sales multiples, followed by sharp share-price declines for many issuers.
For executives, the diligence lesson is to separate the revenue engine, the embodied product line and the actual end customer rather than accept an eye-brain-hand label.

Open at TMTPost Ask Claude ChatGPT

Chinese policy adviser warns against an AI cold war with the US

# Policy SCMP

Policy adviser Zheng Yongnian argues that US AI development cannot fully detach from China, while urging both sides to prevent competition from becoming a cold war.

The remarks place interdependence alongside rivalry: China wants to keep pace in frontier AI while rejecting a complete technological split as impractical.
The timing before the Xi-Trump summit makes the statement part of a broader Chinese effort to frame AI safety and economic ties as reasons for guardrails rather than decoupling.
This is an adviser’s position, not a formal government commitment, but it reflects a recurring Chinese policy preference for managed competition and international coordination.
For US executives, the practical implication is that technology controls may reduce direct links without eliminating shared incentives around standards, safety and supply chains.

Open at SCMP Ask Claude ChatGPT

Tencent’s WorkBuddy adds an agent for Chinese exporters

# AI QbitAI

Export-focused Xiaoyuan AI joins Tencent WorkBuddy through MCP, combining buyer matching, outreach writing and trade-show lead generation with a general office agent.

The agent is built around Miolante’s claimed 30 years of export-service experience, global buyer data and inquiry records rather than generic chat capability.
Its design treats company memory as an operating asset: buyer profiles, negotiation experience and reusable content accumulate within the customer’s authorization boundary.
The platform strategy is clear—large office suites supply the entry point and governance, while vertical agents supply domain workflows and data.
The reported token-based pricing lowers initial adoption friction, but the article provides no measured conversion or revenue results.

Open at QbitAI Ask Claude ChatGPT

China’s trade surplus widens ahead of the Washington summit

# Signal SCMP

Chinese-facing market analysis says exports rose 25 percent year over year and the trade surplus reached 119 billion dollars, strengthening Beijing’s leverage before talks with Washington.

The source says China’s surplus with the US rose 44 percent to 29 billion dollars, underscoring the imbalance that drives US pressure.
The numbers help explain why Beijing can absorb some tariff friction while Washington remains sensitive to import dependence, even though a surplus is not itself proof of economic strength across all sectors.
The analysis frames China as having emerged stronger from the tariff shock, a confidence-building domestic narrative that may narrow room for visible concessions.
The report is commentary based on trade data, not a neutral forecast of summit outcomes.

Open at SCMP Ask Claude ChatGPT

Beijing asked for a truce through 2029 and got eleven weeks

Eleven weeks of truce, not three years

Anthropic's ban wave hit the foreign clouds

DeepSeek doubled revenue, then deprioritised it

Xiaomi previewed V3's attention two days later

Chinese labs are redoing pretraining over dirty data

Xianyu says the worst screenshots weren't its

Nobody is taking their annual leave

Threads we are pulling