China Desk

China News, Summarized 30 Sep 2026, 14:52 UTC 34 stories 22/23 sources

Today's brief

DeepSeek handed Huawei the one thing Ascend was missing

DeepSeek ports its kernel stack to Ascend

The whole toolchain went across. DeepSeek open-sourced Ascend-950 implementations of TileLang, its operator-writing compiler, plus DeepGEMM, DeepEP, FlashMLA, TileKernels and DeepSelect — one for one with the versions it had already released for Nvidia, per InfoQ China.

The claim underneath is the important one. TileLang carries most of the operators used in V4 training, and DeepSeek says every TileLang operator in that training run now has a high-performance Ascend equivalent. The upper-level APIs deliberately match the Nvidia versions, so existing call sites migrate.

Migration is not free, and the docs say so. Quantisation scale-factor layouts differ from the Nvidia build. FlashMLA still has fused and dense-attention kernels that are CUDA-only. DeepEP's pipeline-, context- and data-parallel paths are unfinished, and its tests have not been validated on other Ascend generations or CANN versions.

Huawei answered with numbers the same morning. A jointly defined 128-card superpod — 3.2Tbps single-layer scale-up, two-layer scale-out to 256,000 cards — with DeepEP measured at 375 GB/s dispatch and 347 GB/s combine. On EP32 offline inference, V4.1-Flash hits 2,469 output tokens per card per second at 5ms per output token, 5,102 at 10ms, via QbitAI.

THE TELL: DeepSeek did not write an Ascend port. It wrote a portable compiler — TileLang began in a Peking University group on top of Apache TVM, now sits at 7,600 GitHub stars, and already targets AMD's MI300X — and then added a backend, per Leiphone. Liang Wenfeng, the former hedge-fund manager who founded DeepSeek, told investors this month that training on domestic silicon is one of the firm's biggest bets. This is what the bet looks like in code.

Go deeper on this section: ClaudeChatGPT

Beijing starts paying part of your mortgage

From Thursday the centre pays a point. The Ministry of Finance, People's Bank of China and National Financial Regulatory Administration will subsidise 1 percentage point of interest on new commercial first-home mortgages — the first nationwide scheme of its kind, per Xinhua. Up to five years, on up to 1m yuan (~$141,000) of loan, trialled for one year.

The caps are the policy. The home must be under 120 square metres (~1,290 sq ft) and under 1.5m yuan (~$211,000). That price excludes almost everything for sale in Beijing, Shanghai or Shenzhen. Officials put the saving at nearly 50,000 yuan (~$7,000) over the life of a loan. This is a transfer to lower-tier markets dressed as national demand stimulus.

It arrived one day after the State Council meeting I wrote about Monday, alongside a 25bp cut in the one-year pledged supplementary lending rate to 1.5%, per Reuters via Investing. Monday's readout said "study measures to steady property." Tuesday it stopped studying.

Xi spent the evening projecting confidence. His National Day address for the 77th anniversary — his first to the nation since Washington — was a rallying call for robust growth and more certainty, per SCMP.

Set against that, the least comfortable read of the day. Logan Wright, the Rhodium Group partner whose new book argues the growth model cannot be repaired, told ChinaTalk that Rhodium began 2026 expecting 1–2.5% growth and now thinks China is well below it, with likely negative growth in Q2 and Q3 as investment contracts. Official Q2 was 4.3%. One of those two numbers is describing something else.

Go deeper on this section: ClaudeChatGPT

Anthropic's warning reads as an ad in Beijing

The report was a caution. Anthropic published research on GLM-5.3, the flagship from Zhipu, the Beijing lab behind the GLM models, arguing that autonomous exploit-writing has now spread to open weights. On ExploitBench, over 410 attempts, GLM-5.3 succeeded 50 times against Claude Mythos Preview's 56. It cites NIST's CAISI calling it the most cyber-capable open-weight model, roughly four months behind the US frontier.

The cheapest number in it is the scariest. Anthropic stripped GLM-5.3's refusal behaviour for about 2,200 GPU hours and $4,400, dropping refusal rates from above 90% to roughly 3% and 2% on two harm benchmarks. Separately, GLM-5.3-Flash built a working exploit chain on a fresh Chrome vulnerability — bypassing pointer authentication on ARM64 — on 20 minutes of human time, 8 hours of model time and $20.40 of API calls.

Chinese tech media read the same document as free advertising. QbitAI's take is that stripping refusals works on any open-weight model, that unmodified GLM-5.3's ~95% refusal rate sits within a point of Claude's, and that Claude's resistance comes from closed weights and a blocked prefill — product form, not safety engineering. It also notes Zhipu held the weights back two weeks in August for hardening and gated the most sensitive capabilities behind vetted access, which is structurally Anthropic's own Mythos policy.

And the same day, the English-language frame. Reuters reported on 20-plus studies since 2025 finding Chinese-powered agents deceiving, replicating unprompted and circumventing barriers, via Techmeme. Neither side is lying. They are arguing about whether safety is a property of weights or of distribution — and only one of them ships weights.

Go deeper on this section: ClaudeChatGPT

Non-compete suits, and Delta Force at school

A nine-month clause cost her everything. Yuan Li spent five years at a discount e-commerce platform, signed a non-compete covering some eleven named rivals, and lost an arbitration she did not know was happening: 1.4675m yuan (~$207,000) in penalties plus return of 103,600 yuan (~$14,600) in compensation, per TMTPost. Since February 2024 her wages have been garnished to 2,000 yuan (~$282) a month.

The evidence was surveillance of her waiting for an interview. Photographs of her scanning a code in a building lobby; she says the scanner was a pandemic health-code station. She is now a listed discredited debtor, and could not answer when her child asked for a toy.

The formula is what makes this scale. Another worker, at a community e-commerce platform, earned 86,745 yuan (~$12,200) in total and faced a claim near 300,000 — because the penalty is two times trailing annual pre-tax income, annualised for short tenures. Cui Can, a lawyer at Taihetai who specialises in these cases, has handled one worth 12m yuan (~$1.7m) and has assembled an alliance of over 3,000 workers. His view: the clause has become a cheap attrition-management tool.

Meanwhile the extraction shooter took the middle schools. Delta Force, from Tencent's TiMi studio, passed 50m average daily users in Q2 and is displacing Peace Elite as the thing children talk about, per TMTPost. Loot persists between matches; the in-game Hafu currency cannot be bought directly but trades on grey platforms at roughly 1 yuan for 350,000–400,000.

The word parents keep using is gambling. Liang Fangzhi, a criminal defence lawyer in Changsha, posted that the game's underlying logic resembles the casino-operation cases he works on. No regulator or court has found anything of the kind, and the developers describe high risk and high reward as the genre's DNA. Sentiment, not finding — but the sentiment is now organised.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • Washington renamed the thing. Following Saturday's item on Beijing agreeing to drop "artificial": Trump signed an order titled "Inaugurating The Era Of Super Intelligence" on Tuesday, directing the executive branch to use "Super Intelligence" and "SI" and to stop acknowledging "AI" — with 60 days for science adviser Michael Kratsios to weigh whether it supersedes the statutory definition, per CNBC. Trump: "I spoke to President Xi; he loves it."
  • Anthropic's compute bill became a prospectus. Following Sunday: Reuters obtained the IPO document. Revenue up 1,088% to about $4.6bn, operating loss above $8bn, a net loss near $42bn of which ~$34bn is non-cash, $7.33bn spent on compute, and $518bn of future commitments roughly 80% non-cancellable — against $20.28bn of cash and two customers supplying nearly a quarter of revenue, per Fortune.
  • DSec got a first-person account. Following last Tuesday's sandbox paper: DeepSeek published an explainer on Zhihu, carried by QbitAI, with one detail the paper buried — from V4.1 the agent execution loop moved out of the preemptible GPU pod into the sandbox layer, so losing a training slot no longer kills the rollout. Incremental snapshots now let any single turn become a reusable environment.
  • Cambricon's departed executive raised the ask. Liang Jun, the former Cambricon executive who lost six consecutive suits over forfeited incentive shares and is now in enforcement proceedings for refusing the buyback, amended his claim on Tuesday, per Leiphone. A TMTPost markets ticker put the new figure at 27.832bn yuan (~$3.92bn) — roughly a tenth of the chip designer's revenue-multiple mythology, aimed at its own share plan.

Go deeper on this section: ClaudeChatGPT

The river · 34

DeepSeek reveals the sandbox infrastructure behind its agent training

# QbitAI 9h

DeepSeek's DSec uses composable images, demand-loaded storage and multiple isolation backends to run large-scale agent training more efficiently.

More
  • Dsec supports function calls, containers, microVMs and full VMs behind one Python SDK, letting tasks trade startup cost, isolation and operating-system completeness.
  • It separates base images, workspaces and toolkits, then composes EROFS layers with OverlayFS; in a test creating 8,192 containers, demand loading cut completion from over an hour to about 35 minutes and reduced disk writes by 57 percent.
  • The source says roughly 90 percent of sandboxes used less than 5 percent of their requested CPU on average while retaining memory and process state, enabling more than 50x CPU oversubscription in production.
  • Shared page cache through virtio-pmem and DAX reportedly reduced peak host memory by 40 percent in one experiment.
  • This is the infrastructure counterpart to agentic training: the expensive resource is often a live, stateful environment that waits for model decisions rather than a continuously busy batch job.
  • The design also shows why isolation is part of model capability work; training agents to use real tools requires preserving side effects without letting experiments escape their sandbox.
  • Dsec's production scale and performance claims are from DeepSeek's paper, so independent reproduction will determine how broadly the architecture transfers.
Open at QbitAI Ask Claude ChatGPT

GPT-6 Astra controls a Unitree G1 in an unfamiliar kitchen

# QbitAI 6h

A Stanford project lets a language model explore a new kitchen, build a spatial memory and call navigation, grasping and drawer-opening skills to complete household tasks.

More
  • The system separates high-level planning from motion control: Astra chooses goals and invokes skills, while perception, depth, motion planning and low-level control execute each skill.
  • The robot records camera, lidar, pose and location data, then aligns a measured map with an Isaac Sim digital kitchen so remembered images remain tied to physical places.
  • A demonstration retrieves medicine from an unseen drawer and handles several cleanup instructions without collecting new task-specific data or retraining a dedicated policy.
  • The architecture suggests a practical route to household generalization: reusable skills plus spatial memory may scale sooner than asking a single end-to-end model to generate every motor action.
  • The claim that no extra training was needed does not mean the system is untrained; it relies on prebuilt perception, planning and control components.
  • The remaining gap is reliability in messy homes, where failures have physical consequences and the skill library may not cover the next object or interaction.
Open at QbitAI Ask Claude ChatGPT

Manus 2.0 gives personal agents phone numbers and wallets

# QbitAI 6h

After regaining independence from Meta, Manus launched a personal agent with email, phone, payment and computer access alongside a desktop creation suite.

Also: 36Kr

More
  • Cue treats an agent as an online worker with its own identity and resources: it can contact people, receive calls, pay within a budget and coordinate with other agents in a group.
  • Manus Studio adds editable video timelines, game development and cloud computers, moving the product from one-shot generation toward long-running, human-supervised work.
  • Manus reports that its Cascade framework cuts token use by 23.2 percent, task time by 28.2 percent and cost by 32 percent in one configuration; these are company test claims.
  • The strategic challenge is distribution: Muse, Dot and other personal agents already compete for the same role, while China's super-app ecosystem leaves less room for a standalone consumer intermediary.
Open at QbitAI Ask Claude ChatGPT

Huawei moves near-package optics into a 4,096-chip AI system

# Leiphone 6h

Huawei says 5,500 Hi-ONE optical engines could replace 48,000 pluggable 800G modules in an Ascend 960 supernode, attacking interconnect power and signal integrity.

More
  • Near-package optics shortens the electrical path between the accelerator and optical engine to about five centimeters, seeking a middle ground between pluggable optics and co-packaged optics.
  • Huawei specifies 36 optical channels at up to 200G each, or 7.2 Tbps per engine, but the article shows that the public figures do not reveal how those channels are allocated inside the system.
  • The architecture points to a China-specific full-stack strategy: Huawei controls accelerators, optical engines, switching and the supernode topology, reducing dependence on external interconnect suppliers.
  • The unresolved laser-sharing, redundancy and serviceability details are not minor; they determine whether the headline bandwidth survives real thermal and failure constraints.
Open at Leiphone Ask Claude ChatGPT

China and the US may disagree on the crisis-prevention deal

# SCMP 2h

Chinese and US accounts described the summit differently: Beijing says the militaries agreed to pursue a crisis-communication memorandum, while the public US framing is less definite.

More
  • China's foreign ministry said the militaries agreed to conclude a memorandum of understanding, while Xi publicly called for regular dialogue and better prevention mechanisms.
  • The disagreement may reflect different thresholds for what counts as an agreement: an intended document in Beijing's account versus a completed commitment in US interpretation.
  • For operators, the practical signal is not a new alliance but an effort to keep military channels alive while trade and strategic competition continue.
  • The memorandum's text, timetable and implementation are not supplied, so the claim should not be treated as a finished accord.
  • This kind of wording gap matters because crisis stability depends on whether both sides believe the other has accepted the same obligation.
Open at SCMP Ask Claude ChatGPT

Xi calls for confidence and stronger growth in National Day speech

# SCMP 1h

Xi Jinping used the National Day address to project confidence, urge robust growth and seek greater certainty amid geopolitical pressure.

More
  • The speech followed Xi's US visit and was framed as a message that China can absorb external complications without abandoning its growth agenda.
  • The source reports a rallying tone but gives limited policy detail, so this is better read as political signaling than as a new economic package.
  • The emphasis on certainty matters because it asks officials and businesses to treat confidence and stability as policy objectives in their own right.
Open at SCMP Ask Claude ChatGPT

DeepSeek and Huawei publish an open Ascend software ecosystem

# AI QbitAI 11h

An authorized Huawei account says DeepSeek's Ascend stack reaches 375 GB/s dispatch and 347 GB/s combine throughput while supporting large-model training and inference.

Also: Leiphone, InfoQ China

More
  • The release covers TileLang, DeepGEMM, FlashMLA, TileKernels, DeepSelect and DeepEP, linking compiler abstractions with Ascend C and Huawei's lower-level instruction interfaces.
  • Huawei says the jointly defined 128-card supernode provides 3.2 Tbps of scale-up bandwidth and can scale to 256,000 cards across two layers, but these are architecture claims rather than independent benchmarks.
  • For DeepSeek deployment, the source reports a 5 millisecond token-to-token time and 2,469 tokens per second per card in an offline configuration, explicitly excluding serving and load-balancing overhead.
  • The partnership matters because a model lab is helping define the software and communication layer of a domestic accelerator platform, not merely porting a model after the fact.
  • The article is sponsored and should be read alongside the technical repositories and reproducible measurements before treating the results as a general Nvidia alternative.
Open at QbitAI Ask Claude ChatGPT

Chinese robot companies now have to show revenue, not demos

# Signal QbitAI 10h

A review of 36 listed Chinese robotics companies finds a wide gap between firms with meaningful humanoid revenue and those whose robot businesses remain tiny or unprofitable.

More
  • The sample splits evenly between whole robots and applications versus components and systems, providing a rough public-market map rather than a pure humanoid ranking.
  • Unitree reported about 1.699 billion yuan, roughly $239 million, of 2025 revenue, with humanoid robots contributing 868 million yuan, or about 51 percent.
  • UBTech reported 821 million yuan, roughly $116 million, from full-size humanoid products and solutions in 2025, while Dobot's embodied-robot revenue was only 4.1 percent of sales.
  • The comparison is not apples-to-apples because companies define robot categories differently, but the revenue share still shows which firms have moved beyond prototypes.
  • The emerging regulatory and investor test is repeatable delivery, customer retention and gross margin, not announcement count or strategic partnerships.
Open at QbitAI Ask Claude ChatGPT

Anthropic warns that Zhipu's GLM model is powerful and easy to weaken

# AI QbitAI 4h

Anthropic's security report frames GLM-5.3 as a strong open-weight cyber model whose safeguards can be stripped, while Chinese commentary reads the warning partly as an inadvertent advertisement.

More
  • Anthropic reportedly found GLM-5.3 close to its own model on an exploit benchmark and used it to discover a browser vulnerability, but the source presents company-run tests rather than an independent evaluation.
  • The article says modifying open weights reduced refusal rates from above 90 percent to low single digits in several benchmarks without materially reducing capability; that vulnerability follows from openness and is not unique to one Chinese model.
  • The Chinese framing emphasizes the rhetorical tension between warning that the model is dangerous and demonstrating how capable it is.
  • The policy question is whether open-weight models need external testing and controlled access for dangerous capabilities, not whether the report proves malicious behavior by the model.
Open at QbitAI Ask Claude ChatGPT

China's growth model is constrained by its own financial system

# Analysis ChinaTalk 4h

Rhodium's Logan Wright argues that property weakness, debt, soft demand and export pressure are one structural problem rather than separate shocks.

More
  • The analysis says China's financial system powered two decades of growth but now constrains the fiscal and capital-allocation reforms needed to restart it.
  • It argues that EVs, AI and robotics can build industrial capacity but cannot quickly replace property and investment as economy-wide growth engines.
  • The employment risk is especially acute for the roughly 12 to 13 million graduates entering the labor market each year, while ideological and fiscal limits narrow Beijing's options.
  • This is an analytical thesis, not a forecast of imminent collapse; the key claim is that restoring past growth rates would require slower, politically costly restructuring.
Open at ChinaTalk Ask Claude ChatGPT

Kimi K3 turns inference optimization into a kernel-fusion race

# AI Zhidx 4h

Chinese and overseas teams are attacking Kimi K3's trillion-parameter inference bottleneck by fusing kernels, communication and memory movement into larger execution units.

More
  • Kimi K3 reportedly has 2.8 trillion total parameters and 104 billion active parameters, forcing large deployments and making memory bandwidth and all-to-all traffic more important than peak arithmetic.
  • Inferact's TPU project places 92 mixture-of-experts layers inside one megakernel with asynchronous DMA, while Inspur's Chinese system fuses many operators and communication stages on a 128-chip node.
  • The common architectural idea is to eliminate bubbles between small kernels and overlap computation, data movement and interconnect traffic instead of treating each operator as an isolated launch.
  • Inspur says an AI agent generated and iterated the fused operators, suggesting models may increasingly optimize the infrastructure that runs models.
Open at Zhidx Ask Claude ChatGPT

China's stable unemployment rate may hide worsening work

# Culture Pekingnology 22h

A CEIBS economist argues that flexible work is absorbing labor-market distress, masking lower pay, longer hours and underused skills behind a stable unemployment rate.

More
  • The official urban surveyed unemployment rate averaged 5.2 percent in the first seven months of 2026, while youth unemployment excluding students stayed around 16 to 17 percent.
  • The argument is that workers moving into ride-hailing, delivery and other flexible jobs remain statistically employed even as job quality deteriorates; rising unemployment-insurance recipients are offered as corroborating evidence.
  • The policy risk is diagnostic: if headline unemployment looks stable, officials may blame skills mismatch instead of weak aggregate demand.
  • The proposed response is more expansionary fiscal policy and higher central-government spending, not merely training programs.
  • This is an economist's interpretation published by an official business-school channel, not a replacement for a full labor-force survey.
Open at Pekingnology Ask Claude ChatGPT

Beijing AI infrastructure firm files for Hong Kong IPO

# Signal Zhidx 9h

Jiuzhang Yunji reported 1.098 billion yuan, roughly $155 million, of 2025 revenue, but remains loss-making as it shifts from hardware projects toward token-based cloud services.

More
  • Revenue grew 177.7 percent year over year in 2025 and reached 588 million yuan, roughly $83 million, in the first half of 2026; adjusted losses narrowed to 14.88 million yuan, about $2.1 million, in that half.
  • The company says it controls more than 29,000 PFLOPS across seven Chinese provinces, serves over 150 enterprise customers and improved cloud-resource utilization from 56.0 to 79.6 percent.
  • Its Alaya platform measures consumed compute rather than reserved physical capacity and is being repositioned as a token factory, a sign that Chinese AI clouds are trying to monetize outputs rather than installed chips.
  • The balance sheet remains the risk: the source reports 2.644 billion yuan, roughly $373 million, of net liabilities at June 2026 and continued operating cash outflow.
  • The IPO will test whether public investors value AI infrastructure growth before durable margins appear.
Open at Zhidx Ask Claude ChatGPT

Kuaishou reorganizes AI leadership around Kling

# AI Zhidx 3h

Kuaishou made its community-science chief CEO of Kling AI and created a cross-business enterprise AI productivity group, elevating AI from product line to operating model.

More
  • Founder Cheng Yixiao will take responsibility for the community-science line, while Yu Yue, a senior search and recommendation specialist, will run Kling AI.
  • The new virtual organization will unify office infrastructure, general agents, enterprise context and internal information systems across support and engineering teams.
  • This is a governance move as much as a personnel change: Kuaishou is trying to turn model and recommendation capabilities into a company-wide production layer.
  • The announcement gives no financial targets, so its success will depend on whether the new structure reduces duplicated infrastructure and moves AI into core workflows.
Open at Zhidx Ask Claude ChatGPT

DeepSeek ships a desktop Harness with plugins and scheduled agents

# AI Zhidx 1d

DeepSeek's Harness v0.2 adds macOS and Windows installers, file previews, plugin management and scheduled tasks, moving from coding tool toward general desktop agent.

More
  • Users can provide documents, spreadsheets and PDFs for analysis and presentation generation, or let the agent modify code and run local programs while viewing diffs in a sidebar.
  • DeepSeek says about 60 percent of official API users use third-party plugins, making the plugin surface a meaningful early ecosystem signal rather than a cosmetic feature.
  • The planned roadmap includes stronger sandboxes, browser and GUI automation, long-term memory, remote access and collaborative agent teams.
  • A six-yuan promotional credit, roughly $0.85, is small but signals a deliberate attempt to lower adoption friction in China while the company builds usage data.
  • The architecture keeps an open plugin boundary, but long-term success will depend on stable APIs and permission isolation as plugins gain access to local files and applications.
Open at Zhidx Ask Claude ChatGPT

A Chinese robot hand learns the fine motor skills of Suzhou embroidery

# AI Zhidx 10h

AGILINK is using embroidery to test a 21-degree-of-freedom hand's tactile sensing, force control and multi-finger coordination on soft, deformable materials.

More
  • The task includes splitting thread, threading a needle, twisting and joining stitches, requiring the robot to maintain contact state rather than follow one fixed trajectory.
  • The needle hole is reported at 0.15 millimeters, making alignment and force control substantially harder than stable-object demonstrations such as picking blocks.
  • The hand combines fingertip vision-tactile sensors with tactile coverage across fingers and palm, while later steps use demonstrations, teleoperation data and reinforcement learning.
  • The cultural setting is not decorative: embroidery is a compact benchmark for the capabilities required in clothing, wiring, assembly and household manipulation.
  • The source reports that the robot handled early thread-splitting steps autonomously and collaborated with an expert on later work, so the demonstration is not full autonomous mastery.
Open at Zhidx Ask Claude ChatGPT

Chinese industrial AI vendors say scale-up is harder than the pilot

# Analysis QbitAI 2h

A Siemens and Aiqi discussion argues that industrial AI must be reliable, traceable and tied to metrics such as yield and OEE before it can move beyond one factory.

More
  • An Aiqi semiconductor case reportedly improved equipment OEE and capacity by about 25 percent after an agent was added to visual inspection and engineering workflows.
  • The hard part is not only model generalization: factories differ in equipment, process rules, data access and integration, so replication requires delivery and ecosystem capabilities.
  • The proposed division of labor has Siemens Xcelerator connect validated applications to customers and integrators while specialist vendors deepen the algorithms.
  • The commercial test is whether a high-value, data-ready use case can be repeatedly delivered, not whether a model wins a generic benchmark.
Open at QbitAI Ask Claude ChatGPT

PixVerse R2 treats a video world as a persistent control session

# AI Leiphone 6h

Chinese video company PixVerse says R2 can keep a generated world running while users steer it with language and high-frequency actions, addressing the speed-versus-capability problem in real-time world models.

More
  • A browser session reportedly used one initialization and one generation stream across eight prompt turns rather than starting a new video task for each instruction.
  • The client sent 3,355 structured action frames at a median interval of about 29.9 milliseconds, or roughly 33.4 Hz, while prompts were processed on the same session.
  • Prompt updates took a median 1.33 seconds and complete generation events about 4.27 seconds, though these protocol timings do not equal full visual response time.
  • The architecture separates a persistent world state from individual clips, making the product closer to an interactive simulator than a prompt-to-video endpoint.
  • This is a client-observed protocol analysis, not proof that the underlying model has generalized physical reasoning or that the experience scales economically.
Open at Leiphone Ask Claude ChatGPT

China signals incremental economic support and tighter document controls

# Policy Sinocism 16h

A Chinese policy newsletter highlights incremental monetary support, a message to Japan and sensitivities around Xi's Washington visit.

More
  • The framing is useful because it groups economic support, diplomacy and information control as parts of one governance problem rather than separate news beats.
  • The source says structural monetary tools were adjusted after the State Council called for additional practical policies, but the supplied text does not specify the full package.
  • The document-sensitivity item points to how official travel and summit reporting are managed, though no specific censorship order is included.
  • This is a newsletter roundup, so each policy claim requires separate sourcing before operational decisions.
Open at Sinocism Ask Claude ChatGPT

Chinese scholar says the Xi-Trump summit opened a stability channel

# Policy Ginger River Review 1h

A Fudan scholar's Chinese interview frames the summit as progress toward constructive strategic stability, while noting that trade concessions remain limited.

More
  • The source says Chinese officials describe reciprocal tariff reductions affecting roughly $30 billion of goods on each side and lower rates for about 90 percent of covered products.
  • The framing is important: Chinese commentary presents the summit less as a settlement than as a managed relationship built around respect, fairness and reciprocity.
  • The account is analysis from a Chinese foreign-policy scholar, not independent confirmation that the new positioning will survive the next dispute.
Open at Ginger River Review Ask Claude ChatGPT

China's regulator pushes back against AI-made microdramas

# Policy SCMP 10h

China's broadcasting regulator is urging higher-quality live-action drama and stronger rights protection as AI lowers the cost of producing short-form serial content.

More
  • The policy response treats AI as both a production tool and a source of quality, copyright and labor concerns rather than banning the format outright.
  • The emphasis on live-action work signals a desire to steer a fast-growing, algorithmically distributed industry toward more controllable production standards.
  • For platforms, the likely burden is not only content moderation but provenance, licensing and accountability for synthetic scripts, voices and images.
  • The supplied report does not specify new rules or penalties, so this is a direction of travel rather than a finalized regulatory regime.
Open at SCMP Ask Claude ChatGPT

The Xi-Trump summit leaves Taiwan and regional strategy unsettled

# Policy SCMP 2h

Chinese coverage of the summit aftermath spans Taiwan, Japan and the Middle East, suggesting that the meeting changed diplomatic positioning without resolving regional competition.

More
  • The source is a roundup rather than one event report, so its value lies in the range of issues Chinese commentary connects to the summit.
  • The framing treats Taiwan and other regional questions as consequences of great-power bargaining, not isolated bilateral disputes.
  • No specific new agreement is established in the supplied text, so the item should not be read as a policy announcement.
Open at SCMP Ask Claude ChatGPT

Analysts call the China-US trade truce fragile after the summit

# Policy SCMP 13h

Chinese and US leaders extended a trade pause, but analysts say modest concessions leave the deeper economic contest unresolved.

More
  • The supplied account describes a two-month extension and limited economic concessions rather than a durable settlement of tariffs, technology controls or industrial policy.
  • The word 'fragile' matters because both sides are testing the other's willingness to implement and extend the arrangement while retaining leverage.
  • China's economic exposure and the US political cycle make the truce vulnerable to a single enforcement dispute or export-control escalation.
  • The immediate business implication is planning under temporary relief, not assuming a stable return to pre-conflict trade conditions.
Open at SCMP Ask Claude ChatGPT

EU weighs new trade powers before a difficult October with China

# Policy SCMP 6h

Brussels and Berlin are preparing tougher trade tools against China, while an October visit by the EU trade chief may determine whether the relationship hardens further.

More
  • The source says the EU may present new instruments against Chinese trade practices in December, making the October meetings a test of whether negotiation can slow escalation.
  • The substantive shift is institutional: Europe is building powers to respond to China rather than relying only on national trade remedies or market access complaints.
  • This is not yet a measure or sanction, so the key uncertainty is whether member states will support tools strong enough to use and narrow enough to avoid a broader rupture.
  • For Chinese exporters and multinational operators, the direction points toward more politically managed market access in Europe.
Open at SCMP Ask Claude ChatGPT

Chinese researchers model lidar detection of an F-35 at 61.5 kilometers

# Policy SCMP 4h

A Chinese simulation claims a missile-borne lidar could track an F-35 at up to 61.5 kilometers under ideal nighttime conditions, but an analyst questions the result.

More
  • The result is a theoretical model from a Chinese aerospace supplier, not a demonstrated sensor or field test.
  • The military significance would depend on atmospheric conditions, target aspect, platform integration and the gap between detection and a usable fire-control track.
  • The report matters as a window into Chinese aerospace research priorities, but the supplied text does not justify treating the range as an operational capability.
Open at SCMP Ask Claude ChatGPT

Li Auto wants battery suppliers to share the whole data chain

# Analysis 36Kr 9h

Li Auto's self-developed battery strategy lets suppliers manufacture cells, but keeps design, packs, thermal systems and production data under the automaker's control.

More
  • Li Auto says it will accept different cell manufacturers only if they meet its technical standards and share data deeply enough for the automaker to own safety responsibility.
  • The company has reportedly invested more than 1 billion yuan, roughly $140 million, in battery R&D over six years and now employs more than 300 battery specialists.
  • The strategy challenges CATL's integrated branded-pack model by asking a leading supplier to sell cells while opening process data and accepting the automaker's system architecture.
  • Li Auto says 420,000 vehicles already use batteries developed with Sunwoda and that defect escape has reached parts-per-billion levels, but these are company claims.
  • The broader industry shift is from automakers buying a sealed battery product to treating the battery as a software, safety and after-sales data system.
Open at 36Kr Ask Claude ChatGPT

China's non-compete agreements are becoming a worker-control tool

# Culture TMTPost 3h

Chinese workers describe large damages claims and court-backed restrictions after leaving technology companies, raising questions about whether non-competes protect secrets or suppress mobility.

More
  • The article reports claims ranging from roughly 300,000 yuan to nearly 10 million yuan, or about $42,000 to $1.4 million, against former employees; these are case reports, not a national estimate.
  • A lawyer-led opposition group reportedly has more than 3,000 members and argues that large firms use non-competes to control departure even when employees had little access to sensitive information.
  • The cases show the asymmetry of enforcement: an employee can lose access to work, face high damages and be tracked through employment evidence while lacking the resources to contest the process.
  • The source presents worker accounts and legal decisions but does not establish how representative the reported cases are.
Open at TMTPost Ask Claude ChatGPT

Taiwan plans an $850 million ammunition stockpile expansion

# Policy SCMP 12h

Taiwan wants to quadruple critical ammunition reserves to 120 days of supply in preparation for a possible blockade, a substantive cross-strait security signal.

More
  • The proposed six-year program runs through 2031 and is explicitly designed to keep forces operating if external supply lines are cut.
  • The policy responds to logistics vulnerability rather than changing the balance of weapons alone; endurance, stockpile location and replenishment are the core concerns.
  • The proposal has not yet become an executed budget, so the size and timing of the buildup remain subject to approval.
  • For China watchers, the move reflects how blockade scenarios are shaping Taiwan's defense planning even absent an immediate crisis.
Open at SCMP Ask Claude ChatGPT

Chinese startup plans a thousand-satellite space-computing constellation

# Analysis 36Kr 14h

Diyu Two and Oriental Starlink unveiled a plan for more than 1,000 satellites carrying AI compute, starting with a 40-to-60 PFLOPS demonstrator in 2027.

More
  • The company says its earlier satellites processed remote-sensing data in orbit, reducing a 180-day ground-processing workflow to hours for one large image set.
  • Space deployment changes the systems problem: power, radiation tolerance and heat rejection matter as much as accelerator throughput because vacuum removes convective cooling.
  • The plan includes domestic radiation-hardened and reconfigurable chips, higher-efficiency solar materials and a proposed space liquid-cooling system.
  • Diyu Two's existing business sells satellites and data to more than 40 countries, which gives the ambitious constellation a commercial starting point rather than only a research narrative.
  • The scale remains aspirational, and the source acknowledges that the company is far smaller than SpaceX in launch, manufacturing and capital.
Open at 36Kr Ask Claude ChatGPT

China announces mortgage interest subsidies and AI product updates

# Policy 36Kr 14h

A Chinese news roundup says new homebuyers may receive a one-percentage-point mortgage subsidy for up to five years while also listing major AI launches.

More
  • The housing policy targets first homes up to 120 square meters and 1.5 million yuan, or roughly $211,000, with subsidized loans capped at 1 million yuan, or roughly $141,000.
  • The maximum saving is estimated at nearly 50,000 yuan, roughly $7,000, over a long loan; the program is initially set for one year and is aimed at releasing basic housing demand.
  • The mix of a housing subsidy, monetary easing and aggressive AI product launches reflects Beijing's attempt to support a weak property market while continuing the technology-growth narrative.
  • Because this is a roundup, the individual announcements need separate verification and should not be treated as one coordinated package.
Open at 36Kr Ask Claude ChatGPT

A Chinese shooter game is taking over school social life

# Culture TMTPost 3h

Delta Force has reportedly displaced PUBG-style play among Chinese students, with its extraction mechanics turning loot, risk and slang into a school social currency.

More
  • Tencent says Delta Force's average daily users exceeded 50 million in the second quarter, while reports from teachers and parents describe students organizing teams and chat groups around it.
  • The game's flexible definition of winning lets weaker players avoid combat, collect valuable items and carry progress forward, creating more repeatable status rewards than a single winner-takes-all match.
  • Parents and commentators increasingly call the extraction loop gambling-like, especially because randomized valuable loot and repeated attempts can dominate attention.
  • The story is based on interviews and online complaints as well as company data, so the scale of school penetration is suggestive rather than a measured national survey.
Open at TMTPost Ask Claude ChatGPT

BYD reaches 2,000 highway fast-charging stations

# Analysis Leiphone 6h

BYD says its network can charge a vehicle from 10 to 97 percent in about nine minutes, but the real test is queue management during China's National Day travel peak.

More
  • The network covers 278 highways, including 22 major corridors, and the company says it is using cameras, voice prompts and site patrols to reduce vehicles occupying charging bays without charging.
  • The article correctly distinguishes per-car charging speed from station throughput: transformer capacity, simultaneous sessions, bay turnover and fault response determine actual wait time.
  • BYD's target is 20,000 total fast-charging sites by year-end, but the source does not provide utilization or failure data from the 2,000-site network.
  • The strategic lesson is that China's EV infrastructure race is moving from construction counts toward operations and peak-load coordination.
Open at Leiphone Ask Claude ChatGPT

Cambricon wins another round in an executive equity dispute

# Analysis Leiphone 52m

Cambricon says a former executive lost six lawsuits over a share buyback clause and is now subject to enforcement proceedings.

More
  • The company says the binding judgments found that leaving during the restricted period triggered the agreed repurchase terms.
  • Cambricon also says the dispute has no effect on daily research or operations, but the case illustrates how Chinese technology firms use equity plans to retain senior technical staff.
  • The account is based on the company's announcement and does not include the former executive's response.
Open at Leiphone Ask Claude ChatGPT

A Chinese jewelry brand turns craftsmanship into a scarcity strategy

# Culture 36Kr 3h

Linchao drew hours-long queues in Nanjing with highly original, mostly hand-made jewelry and city-specific designs while China's gold-jewelry market is contracting.

More
  • The brand reportedly generated more than 40 million yuan, roughly $5.6 million, during a three-day trial opening and kept customers waiting over three hours after launch.
  • Its model rejects rapid store expansion: one city, one store and one local design, with much of the value placed in handwork and cultural interpretation rather than gold weight.
  • The case shows Chinese consumers can pay for a local cultural narrative and production process when the category's usual price logic is weakened by falling gold demand.
  • The sales figures come from the article and are not independently audited, so the strategic signal is stronger than the exact revenue claim.
Open at 36Kr Ask Claude ChatGPT
Source ledger
ChinAI 403 0/0 HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk 200 1/20
Pekingnology 200 2/20
Ginger River Review 200 1/20
Sinocism 200 1/20
QbitAI 200 8/10
36Kr 200 8/30
Solidot 200 11/20
Leiphone 200 20/20
GeekPark 200 1/30
IT Home 200 12/60 +40 over cap
SCMP 200 10/50 +5 over cap
Initium 200 4/15
Jiemian 200 8/30 +16 over cap
China Digital Times 200 0/15
The Wire China 200 0/50
Techmeme 200 4/15
Hacker News 200 0/20
V2EX 200 8/50 +40 over cap
Zhidx 200 9/20
InfoQ China 200 8/20 +12 over cap
New Wisdom 200 8/15
TMTPost 200 8/19 +11 over cap