China Desk

China News, Summarized 13 Oct 2026, 01:58 UTC 34 stories 22/23 sources

Today's brief

Beijing agreed to halve a forecast, which is much cheaper than halving a shipment

A halving measured against a prediction

The number finally arrived, and it is about the future. After two days with commerce minister Wang Wentao in Beijing, EU trade and economic security commissioner Maroš Šefčovič said the understanding on hybrid and plug-in hybrid exports opens the prospect of cutting Chinese shipments by more than half over four years — 700,000 fewer vehicles next year alone and up to €14bn trimmed, per SCMP.

Half of what, though. Šefčovič's own phrasing is half of the pace the EU predicts. Chinese hybrid and PHEV exports to Europe run around 50,000 a month, so four years at today's rate is roughly 2.4 million cars; halving that gives 1.2 million, not "millions". The baseline is a growth curve nobody has published, as Initium's anatomy of the deal lays out.

The curve exists because of a tariff gap. Since October 2024 Chinese battery-electric cars have faced combined duties up to 45.3% in the EU while plug-in hybrids pay the ordinary 10%. Monthly hybrid exports went from about 3,800 in October 2024 to some 50,000 by July 2026. Brussels calls that circumvention; Chinese makers call it product-market fit.

What Beijing banked is nearer-term. Sixteen consensus points, faster rare-earth and magnet licensing — Europe sources about 98% of its permanent magnets from China, and the October 2025 controls are suspended only to 10 November 2026 — advance notification of export controls, and roughly €225m of tariff relief for European exporters. Next round: March 2027.

And Brussels keeps the knife out. The goods deficit was about €360bn in 2025 and already €234bn through July; François Chimits, who runs the Europe programme at Institut Montaigne, expects it to pass €400bn and calls it a second China shock. A December toolbox is drafted: emergency import powers triggerable in 24 hours, three-supplier rules for critical sectors, a solidarity fund. Leaders meet Thursday.

Go deeper on this section: ClaudeChatGPT

The clean room got demolished in six days

REA is the loudest thing on GitHub. Reverse Engineer Anything, an MIT-licensed TypeScript project from a developer going by morluto, went from under 10,000 stars on 6 October to over 110,000 by today. It wires Ghidra, IDA, Hopper and JADX into coding agents over MCP, so one sentence gets you a decompiled answer with evidence, per GeekPark's long read.

Practitioners are unimpressed, and right. Elias Bachaalany, co-author of Practical Reverse Engineering and once at IDA's maker Hex-Rays, answered the hype wave with the word "hype". Duncan Ogilvie, who wrote the IDA Pro MCP plugin and the x64dbg debugger, says he cleared this year's Flare-On malware-analysis contest in about an hour. Virtualising packers still hold; server-side logic is untouchable.

The new thing is who is holding the tool. Reverse engineer Maurice Heumann ran a dozen Claude Code and Codex agents in parallel and rebuilt 99% of a shooter's functions as C++, 83% byte-identical — on Hex-Rays' own MCP, not REA. The skill floor, not the ceiling, moved.

Chinese coverage went straight to liability, and that is the better story. The industry's defence has been clean-room development: one team reads the binary and writes a spec, a second team that never saw the code implements it, and the wall between them is the evidence at trial. REA puts both teams inside one agent. Hacker News calls it IP laundering.

The statute cuts both ways. China's 2020 Copyright Law lists software reverse-engineering research among the permitted circumventions of technical measures — then bars supplying circumvention tools to others, which is precisely what a one-line installer does. In the US, Bowers v. Baystate upheld a licence's no-reverse-engineering clause; in the EU such clauses are void. Rockstar, Activision and Bethesda have already pulled browser ports.

The licensed version of the same idea raised money today. ModCraft, founded by Li Shuo — a Tencent and Youzu veteran who produced Game of Thrones: Winter is Coming — took a seed round in the tens of millions of yuan from SenseTime's Guoxiang Capital and L2F. Natural-language game mods across 28 Steam titles, and the run data sold to ByteDance as world-model training material, via 36Kr. Mods as a sanctioned interface into a compiled game.

Go deeper on this section: ClaudeChatGPT

Two ministers wake up on the sanctions list

Paraguay's and Belize's cabinets got personal. From Monday Beijing froze the Chinese assets of Gustavo Villate, Paraguay's ICT minister, and Francis Fonseca, Belize's foreign and education minister, after both governments signed agreements with Taipei, SCMP reports. Both countries still recognise Taipei; the signal is that individual ministers now carry the cost.

Meanwhile the diplomatic calendar is pointed at Europe's edges. Foreign minister Wang Yi attends the Baku summit of the Conference on Interaction and Confidence-Building Measures in Asia on Thursday as Xi Jinping's special representative, then Croatia, Montenegro and Bulgaria — per SCMP. The same days EU leaders sit down to judge the hybrid understanding.

Go deeper on this section: ClaudeChatGPT

Heterogeneous plumbing beat model-building for funding

Infinigence AI raised 2.3bn yuan (~$324m). The Tsinghua electronic-engineering spinout, founded 2023 by professor Wang Yu with chief executive Xia Lixue, is now at 4.3bn yuan (~$606m) cumulative and reportedly preparing an IPO. It sells neither models nor chips — it sells the conversion of mismatched silicon into tokens, per Zhidx.

The claims are specific enough to argue with. A 70-billion-parameter model trained across Ascend, Iluvatar, MetaX and Moore Threads plus AMD and Nvidia at 97.6% utilisation; two working days to adapt a new chip; cross-domain training between Hami in Xinjiang and Shenzhen, some 2,500 miles apart, up 165%; a disaggregated prefill/decode architecture cutting first-token latency 51.5% and per-token cost 37.5%; daily token calls up 40x from December to July.

THE TELL: the Nvidia-shaped version of this business — Fireworks AI at $17.5bn, Baseten at $13bn, Together AI at $8.3bn — is worth roughly thirty times as much. Neutrality across sixteen chip families is a worse business and a more necessary one.

Huawei published the most interesting engineering of the day. Noah's Ark Lab, with UCL, cleared all 25 public ARC-AGI-3 games and 183 levels using 7,518 actions against 17,135 for humans — a perfect relative-efficiency score, on frozen Claude Opus 5 weights with no fine-tuning, per QbitAI. The same model under the official harness averages about 40.7.

The method is a versioned, executable guess about the world. Memento 3 keeps a rule book and a Python simulator of the game, replays its own actions and compares 64-by-64 frames cell by cell to falsify them, and keeps the whole thing in Git so wrong theories can be rolled back. One rule book went through eleven revisions.

Then it stopped calling the model. On Atari Pong it inferred the bounce geometry, handed control to the learned code, and won 21-0 in three different openings off roughly 9,500 frames. Recursive self-improvement with the weights nailed shut, and the artefact is auditable — which is more than most agent stacks can say.

Go deeper on this section: ClaudeChatGPT

A lawyer named the letters he can no longer read

Cai Yaqi posts other people's lives. He is a criminal-defence lawyer and a well-known instructor for China's national legal qualification exam, with 1.59m followers on Bilibili, and he regularly reads out messages strangers send him about their work. On 8 October he posted a 39-second video, now near 434,000 views, described by Pekingnology.

The content was an instruction. People working in the prison system, judicial administration or administrative reconsideration should stop submitting their accounts, because he cannot read them aloud. He would not say why. Everything else was still welcome — "as long as they don't go too far, at least for now." Scrolling comments compared it to pleading guilty for leniency.

Elsewhere the argument is about requisitions nobody files. McKinsey's August survey has 39% of respondents expecting AI to shrink headcount within a year while only 14% say it already has; last year's cohort predicted 32% and under-delivered. Gartner found around 80% of large firms piloting autonomous tech cut staff, with no link between deeper cuts and better returns, per TMTPost.

The pricing is where the shift is legible. Tencent's WorkBuddy still charges 198 yuan (~$28) per seat per month; Salesforce's Agentforce charges $0.10 per standard agent action and Intercom's Fin $0.99 per qualifying resolution. Software is starting to bill for work rather than for logins — against a budget line that used to be payroll.

One person at the sharp end. Xiao Chaohu quit a communications job at 52, crammed for about 100 days, scored 398 and is now a third-year master's student in tea processing at Fujian Agriculture and Forestry University, where he used AI to write a jasmine-scenting monitoring system with two patents in examination. At 55 he applied to be a brand director and was turned down fast, TMTPost reports.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • The Claude ban thread finally has a denominator. I have been running forum anecdote for a week. Anthropic's transparency report: 11.4m accounts banned in the first half of 2026 against 1.45m in the second half of 2025, 398,000 appeals, about 42,000 overturned — a 10.6% success rate.
  • Microsoft's Qwen base is now a company-wide test. Following yesterday: Satya Nadella says the firm is trialling Microsoft-Decision-1 across incident response, quality control and scientific discovery. Of 129 open-source decision models on HuggingFace, 84 are built on Qwen, per Leiphone.
  • CXMT answers the memory squeeze. Following Sunday's piece on the dying budget handset: president Cao Kanyu says a 4F² architecture with vertical surrounded-gate transistors ships as DDR5 server RDIMM by year end — the domestic line moving toward the contracts that caused the squeeze, via Leiphone's morning brief.
  • The hybrid cap does not touch brand entry. At Paris, Huawei-and-Seres brand AITO said it enters Europe in autumn 2027 and wants a Euro NCAP five-star rating this year, per IT Home; BYD's Denza will push its assisted-driving stack to European Z9 GTs by over-the-air update next year pending UN R171 approval.
  • The brake story is now a marketing line. Zhao Yongpo, who runs Great Wall's Wey brand, on the four-ministry design rules: the greatest luxury is not the sofa, it is the brakes.

Go deeper on this section: ClaudeChatGPT

The river · 34

Moonshot AI faces a potential $7B shareholder dispute before its IPO

# Zhidx 18h

Moonshot AI executives are reportedly negotiating with former shareholders seeking compensation equal to 10.5% of original shares, potentially worth more than $7B at the latest cited valuation.

More
  • The dispute concerns the company’s spinout from Circles and allegations involving data transfers, deleted data, and undisclosed equity arrangements; the claims have not been independently verified.
  • The reported amount is a calculation based on a $30B valuation, not an awarded damages figure, and the parties are negotiating through Hong Kong arbitration.
  • For a company preparing a Hong Kong listing, the central risk is disclosure and governance: investors will scrutinize who owned the data, who received equity, and whether liabilities were fully represented.
  • Moonshot’s reported valuation may be around $50B in a current financing process, making the dispute potentially material even if it does not formally block the IPO.
  • The episode shows how China’s founder spinouts can carry unresolved intellectual-property and shareholder obligations into a new AI entity.
  • A settlement before listing could contain the issue; prolonged arbitration would make ownership, capitalization, and internal controls part of the investment story.
Open at Zhidx Ask Claude ChatGPT

Chinese AI-infrastructure firm Wuwen raises a record $3.2B

# Zhidx 23h

Wuwen Xinqiong raised 2.3 billion yuan (about $324M), taking cumulative funding to 4.3 billion yuan (about $606M) as it builds a neutral layer for China’s fragmented AI chips.

More
  • The company’s core problem is heterogeneous compute: turning Huawei, AMD, Nvidia, and other accelerators into one usable training and inference platform.
  • Wuwen says it has trained a 70-billion-parameter model across six chip families at 97.6% utilization, can adapt to a new chip in two working days, and has cut adaptation costs more than 70%.
  • Its cross-region systems reportedly connect clusters more than 4,000 kilometers apart, while a PDD inference architecture cuts first-token latency 51.5% and per-token cost 37.5%.
  • The company’s strategic position is deliberately neutral: it does not build foundation models or chips, but mediates between many models and many accelerator types.
  • The reported platform touches more than 37,000P of compute across 16 chip types and says daily token volume rose 40-fold, though these are company disclosures awaiting independent validation.
  • A potential IPO would make AI infrastructure a public-market category in China and test whether orchestration software can capture value as domestic hardware fragments.
  • The broader signal is that China’s chip constraints are creating demand for software-defined portability and scheduling, not merely more silicon.
Open at Zhidx Ask Claude ChatGPT

China’s new-productivity policy ties innovation to national laboratories

# Sinocism 2h

A Central Committee and State Council document reportedly sets 19 measures for developing new productive forces, emphasizing core technologies, original innovation, and higher total-factor productivity.

More
  • The policy treats productivity growth as the central test of new productive forces and calls for production relations better suited to technology-led change.
  • The emphasis on national laboratories and breakthroughs in key core technologies places industrial policy, research institutions, and enterprise deployment in one state-led framework.
  • For technology executives, the important signal is not a single subsidy but a planning logic that links frontier research to industrial upgrading and institutional reform.
  • The document likely provides a policy umbrella for AI, advanced manufacturing, energy, and strategic materials, although the supplied excerpt does not enumerate implementation mechanisms.
  • Chinese coverage presents the document as an organizing principle for the next development phase, not merely as another technology program.
Open at Sinocism Ask Claude ChatGPT

China and the EU reach a temporary truce over hybrid-car exports

# Initium 14h

Beijing agreed in principle to limit hybrid and plug-in-hybrid exports to Europe, potentially cutting projected shipments by more than half while securing progress on rare earths and market access.

Also: SCMP

More
  • The agreement reportedly covers millions of vehicles over four years, but the baseline and enforcement mechanism remain undisclosed, making the headline reduction impossible to interpret precisely.
  • China’s concession came before EU safeguards or tariffs were imposed, a notable shift from the familiar investigation, tariff, retaliation, and renegotiation cycle.
  • Beijing frames the arrangement as WTO-compliant and reciprocal, preserving room to manage quotas administratively while avoiding an immediate tariff fight.
  • The trade-off includes faster Chinese licensing for rare earths and permanent magnets, improved access for selected European goods, and continued discussion of electric-vehicle tariff commitments.
  • For Chinese automakers, the episode shows how tariff gaps can redirect exports toward hybrid models and trigger new defensive measures even when battery-electric trade is already restricted.
  • The deal is a truce, not structural rebalancing: the trade imbalance, subsidies dispute, and uncertainty over the quota baseline remain unresolved.
  • European officials describe the agreement as a test of whether dialogue produces measurable restraint, leaving the EU ready to use trade instruments if implementation disappoints.
Open at Initium Ask Claude ChatGPT

China is using AI microdramas as a new propaganda channel

# SCMP 17h

Beijing is scaling AI-generated short dramas for public-opinion work, combining popular melodrama formats with policy messaging for domestic and overseas audiences.

More
  • The strategy borrows the distribution logic of commercial microdramas—short episodes, familiar emotional tropes, and rapid production—but redirects it toward culture, policy, and national narratives.
  • AI lowers the cost of producing many localized variants, allowing propaganda teams to test formats and themes at a volume traditional television cannot match.
  • The important shift is not synthetic video alone; it is the integration of generation, platform distribution, ideological guidance, and audience targeting.
  • Chinese media ecosystems already reward sensational short-form content, so state messaging can travel through entertainment conventions rather than formal news channels.
  • The approach raises authenticity and disclosure questions, especially when foreign viewers encounter culturally coded persuasion without clear state attribution.
  • This is a window into how China treats AI as an instrument of ideological work, not just an efficiency technology.
Open at SCMP Ask Claude ChatGPT

China is building a utility-scale molten-salt and CO2 power plant

# SCMP 17h

China has begun building the Ruitan project, combining supercritical CO2 generation with large molten-salt storage at a coal station to reduce waste and stabilize the grid.

More
  • The project is significant because it combines a high-efficiency power-conversion cycle with long-duration thermal storage rather than treating generation and storage as separate systems.
  • Locating it at a coal-fired plant suggests a pragmatic Chinese strategy: reuse grid and industrial infrastructure while testing technologies that could later support lower-carbon or flexible generation.
  • Molten salt can shift electricity across demand periods, while supercritical CO2 systems promise compact turbomachinery and potentially higher conversion efficiency.
  • The project is a demonstration, not proof of commercial economics; materials durability, control complexity, and round-trip efficiency will determine whether it scales.
  • The account calls it the world’s first utility-scale combination, making construction progress and operating data more important than the headline design.
Open at SCMP Ask Claude ChatGPT

China’s embodied-AI boom is creating a low-paid data workforce

# TMTPost 16m

A fast-growing market is paying ordinary workers to record physical tasks for robot training, exposing a new labor chain built on low wages, weak consent, and uncertain data ownership.

More
  • The supplied account says embodied models need vastly more real-world manipulation data than currently exists, driving investment in data collection, labeling, and brokerage companies.
  • Workers use camera helmets, gloves, and phones to record slow, precise household or industrial actions; reported pay is often 3,000–4,000 yuan monthly (about $420–560) or roughly 25 yuan (about $3.50) per hour.
  • Crowdsourcing is replacing expensive robot-equipped collection centers, allowing platforms to recruit workers across dozens of Chinese cities and pay only for accepted data.
  • The labor model resembles gig work but adds hidden quality-control and surveillance burdens: a worker’s movement, biometric signals, and workplace context may become training data.
  • The account describes consent clauses buried in employment paperwork and workers rarely knowing who will use recordings or how long they will persist.
  • A particularly sharp tension is that people may help train robots that could eventually replace their own jobs, making the data supply chain a labor-policy issue rather than merely an AI cost problem.
  • Reported resale prices and circular purchasing between hardware and data firms also raise concerns that financing narratives may overstate genuine production demand.
Open at TMTPost Ask Claude ChatGPT

ByteDance is positioning AI for science as a separate business

# TMTPost 11m

TikTok CEO Shou Zi Chew predicts AI-driven scientific breakthroughs within three years; ByteDance’s Anew Labs spinout, open models, and financing show a concrete strategy behind the claim.

Also: Jiemian

More
  • ByteDance reportedly spun out its five-year drug-discovery team as Anew Labs, retaining 56% while allowing the unit to raise $290M at a $1.5B post-money valuation.
  • The spinout separates internet-scale iteration from drug development’s long validation cycles and creates a structure that can recruit pharmaceutical talent and capital.
  • Anew’s stated lead program is an IL-17 small-molecule inhibitor designed to block three dimer forms, while ByteDance’s open Protenix-v2 targets protein, DNA, RNA, and ligand-complex structure prediction.
  • The reported Apache 2.0 release strategy gives researchers and smaller drug companies deployable weights, potentially making ByteDance’s model an ecosystem interface rather than merely an internal tool.
  • The commercial bottleneck is validation: AI can compress candidate discovery, but clinical trials and manufacturing still impose years of cost and uncertainty.
  • ByteDance also reportedly secured a $29.6B unsecured loan for non-China AI projects, showing how its balance sheet can support long-horizon research even as the company keeps scientific work organizationally separate.
Open at TMTPost Ask Claude ChatGPT

DeepSeek reportedly doubles its planned funding to $141B without raising valuation

# IT Home 10m

DeepSeek reportedly plans a 1,000 billion yuan (about $141B) round at a 5,000 billion yuan (about $704B) pre-money valuation, drawing more industrial investors and moving toward an IPO.

More
  • The reported round is roughly twice the earlier plan, while the unchanged valuation suggests investors are prioritizing strategic access and capacity over a higher entry price.
  • New backers reportedly include semiconductor, automotive, and industrial funds, while existing investors such as Tencent and CATL are also participating; CATL may add about 10 billion yuan (about $1.4B).
  • The investor mix points to a model company being financed as strategic infrastructure for chips, vehicles, cloud, and data rather than as a standalone software startup.
  • The reported valuation is about $704B before the round, an extraordinary figure that makes governance, revenue quality, compute costs, and eventual public-market disclosure material risks.
  • The company is also reportedly preparing a domestic listing, with investor-structure work aimed at avoiding a shareholder-count problem; these claims remain based on unnamed transaction sources.
  • DeepSeek’s first reported external round was about 51 billion yuan (about $7.2B), making the new financing a continuation of unusually state-and-industry-heavy capital formation around Chinese foundation models.
Open at IT Home Ask Claude ChatGPT

Huawei agent beats 25 puzzle games through recursive self-improvement

# AI QbitAI 20h

Huawei and UCL researchers say Memento 3 solved 183 ARC-AGI-3 levels with 56% fewer actions than humans by maintaining a rulebook and executable world model.

More
  • The system reportedly achieved a perfect score without updating model weights or fine-tuning for the games, using code as a mutable model of the environment.
  • It continually revised a written rulebook and simulator, replaying actions against 64-by-64 game frames to catch incorrect hypotheses before planning further.
  • The approach resembles software development: versioned rules, simulation, regression checks, and reusable knowledge let the agent improve through experience rather than retraining.
  • On three games, keeping the rulebook reduced actions about 9% and agent rounds about 18%, suggesting explicit external memory can improve planning efficiency.
  • The result is meaningful for agent architecture, but pixel puzzles offer unusually clean feedback; transfer to open-world tasks with ambiguous rewards remains unproven.
  • The key lesson is that recursive self-improvement may be more about maintaining executable state and verification loops than about autonomous weight changes.
Open at QbitAI Ask Claude ChatGPT

MOSI raises more than $140M to build always-on multimodal AI

# AI Zhidx 21h

Shanghai startup MOSI raised more than 1 billion yuan (about $141M) to commercialize real-time multimodal interaction across devices and physical environments.

Also: 36Kr

More
  • MOSI’s front-end and back-end model split is designed to preserve low-latency conversation while a separate system performs deeper reasoning and actions.
  • The company is pursuing glasses, cars, home devices, and robots, where continuous audio-visual context matters more than a one-shot prompt.
  • Its funding and university lineage show Shanghai institutions backing a differentiated path between general models and device-level interaction.
Open at Zhidx Ask Claude ChatGPT

A Chinese legal celebrity hints at the boundaries of online speech

# Culture Pekingnology 10h

A popular Chinese legal educator says workers in prisons, judicial administration, and administrative review should stop sending him stories he cannot read aloud, a matter-of-fact glimpse of censorship.

More
  • The creator’s large Bilibili audience turns personal submissions into a public channel for experiences that might otherwise remain invisible.
  • His warning is notable for its narrow, institutional boundary: he does not describe a general ban, but signals that stories touching specific state systems have become unusable.
  • The episode shows how censorship often appears through self-editing and creator discretion rather than a visible takedown order.
  • Chinese online life therefore depends on informal knowledge of what can be said, by whom, and in which format, even when the platform remains open.
Open at Pekingnology Ask Claude ChatGPT

Cursor’s S3-backed WAL makes Git storage linearly scalable

# AI InfoQ China 0m

Cursor replaces replica coordination with an S3 write-ahead log, reporting strong consistency and more than 300 pushes per second in synthetic tests.

More
  • S3 is the authoritative store while local NVMe repositories act only as warm caches, shifting consistency from replica coordination to durable object storage.
  • The design uses conditional writes, rendezvous hashing, gossip, and on-demand repository regeneration; reads can scale with replica count, but bandwidth and Git compression become bottlenecks.
  • Cursor reports more than 300 pushes per second on S3 Express One Zone, but the figures have not been independently validated.
  • The architecture resembles a database-backed Git service, making force-push transactions, point-in-time recovery, and failure semantics central design questions.
  • Chinese coverage emphasizes the architectural break from GitHub-style multi-replica coordination; the practical significance is a possible template for agent-generated repositories and high-churn CI workloads.
Open at InfoQ China Ask Claude ChatGPT

Chinese developers debate whether AI reverse-engineering lowers the legal barrier

# Policy GeekPark 13h

The REA project’s explosive popularity is making binary analysis accessible to ordinary AI developers, while Chinese coverage stresses the gap between technical copying and lawful reuse.

More
  • REA connects coding agents to tools such as Ghidra, IDA, Hopper, and Android unpackers through MCP, lowering setup and expertise costs rather than inventing reverse engineering itself.
  • The reported growth from under 10,000 to more than 110,000 stars in six days reflects AI-community adoption more than a new technical breakthrough.
  • Public demonstrations work best on unprotected or lightly protected clients; server-side logic, heavy virtualization, dynamic analysis, and context limits remain barriers.
  • Chinese legal framing distinguishes lawful study of a legitimately obtained product from trade-secret theft, copyright infringement, or violating license terms.
  • The strategic risk is distribution: reverse engineering is moving from a specialist security practice into a general-purpose agent capability, changing who can replicate software behavior.
Open at GeekPark Ask Claude ChatGPT

China’s secrecy industry is expanding from laser shields to phone detectors

# Policy SCMP 3h

China’s Ministry of State Security is warning agencies about espionage risks while a commercial secrecy industry sells laser-window defenses, phone detectors, and related controls.

More
  • The threat model includes optical eavesdropping through windows, where a laser can recover room vibrations and transmit meeting content.
  • The commercial response shows security becoming an everyday procurement category for government departments and sensitive organizations, not only a classified technical capability.
  • The story reveals how Chinese state security messaging creates demand: warnings from a central agency can normalize surveillance countermeasures across bureaucracies.
  • The expansion also implies a more defensive and compartmentalized operating environment for companies handling sensitive data or working near government clients.
  • Chinese coverage presents the industry as a necessary response to hostile intelligence collection, leaving little space for debate over cost, proportionality, or false positives.
Open at SCMP Ask Claude ChatGPT

Zhipu quietly tests a model claiming 490 tokens per second

# AI Zhidx 22h

Zhipu is reportedly gray-testing GLM-5.3 HighSpeed inside ZCode, with one user seeing more than 490 tokens per second, though scale and production stability remain unknown.

Also: QbitAI

More
  • The reported speed is several times faster than cited DeepSeek and standard GLM API figures, which could materially reduce latency across multi-step coding-agent loops.
  • The architecture is not disclosed, but Zhipu’s earlier high-speed model used static compiled GPU kernels, dynamic batching, request merging, and KV-cache scheduling.
  • The model is available only to a random subset of subscribers, so the observed speed may reflect a narrow test path rather than a service-wide capability.
  • For coding agents, throughput matters because dozens of model calls can turn small per-call delays into many minutes of end-to-end latency.
Open at Zhidx Ask Claude ChatGPT

China’s FAST telescope confirms a rare pulsar triple system

# Signal Solidot 20h

FAST data helped confirm J0435+3233 as only the second known pulsar triple system, with a millisecond pulsar, a white dwarf, and a distant solar-mass companion.

More
  • The pulsar spins every 3.2 milliseconds and was monitored 192 times over five years, allowing researchers to link its unusual spin change to the third star’s gravitational influence.
  • The inner pulsar-white-dwarf pair orbits every eight days, while the solar-like companion follows a 73.5-year eccentric orbit.
  • The system is valuable for testing stellar evolution and multi-body dynamics because all three stars appear to have formed from one primordial cloud.
  • FAST’s contribution is both observational and infrastructural: a domestic radio telescope combined with international optical and high-energy archives to identify the faint companion.
Open at Solidot Ask Claude ChatGPT

Beijing sanctions Paraguayan and Belizean ministers over Taiwan

# Policy SCMP 19h

China imposed sanctions on ministers from Paraguay and Belize after both governments deepened ties with Taiwan, using personal financial restrictions to enforce its diplomatic red line.

More
  • The sanctions reportedly freeze assets in China and restrict transactions, targeting officials rather than imposing broad measures on the two states.
  • The move shows Beijing using selective personal sanctions to raise the cost of maintaining formal or substantive ties with Taipei.
  • It also demonstrates how Taiwan policy reaches into technology, trade, and diplomatic networks across smaller states, not only cross-strait military signaling.
  • Chinese policy framing treats support for Taiwan as interference in China’s sovereignty rather than an ordinary bilateral choice.
Open at SCMP Ask Claude ChatGPT

A former Robotaxi operator is building open-world Physical Agents

# AI Leiphone 23h

Former Pony.ai executive Liu Nianqiu raised nearly $100M to build robots for open environments, applying lessons from mapless driving to data collection and physical-agent generalization.

More
  • Feimoyi is avoiding a single tightly defined job and instead wants robots to operate in offices, homes, commercial districts, and public spaces as adaptable companions.
  • The founder argues that mapless driving required rebuilding the architecture around broader perception and data production; similar integration may be needed for robots that cannot rely on fixed task boundaries.
  • The company plans a deployment-to-data-to-training loop, but it acknowledges the large gap between occasional demo success and dependable service whenever a user asks.
  • The strategic bet is on transfer across tasks and environments, whereas much of the sector still optimizes for narrow factory, hotel, warehouse, or delivery scenarios.
  • The near-$100M seed round shows investors are willing to fund open-world autonomy before the reliability standard has been solved.
Open at Leiphone Ask Claude ChatGPT

Shanghai will test robots on real jobs, not staged demos

# AI Leiphone 11h

A Shanghai embodied-AI competition will evaluate robots across 23 tasks in nine industries using real work settings, task execution, data collection, autonomy, and innovation.

More
  • The competition uses a common scoring framework and different coefficients for humanoid, quadruped, wheeled, autonomous, and teleoperated systems.
  • Its design treats deployment data and task completion as part of the product, linking training, evaluation, financing, and commercialization in one pipeline.
  • The event reflects a state-backed attempt to turn robotics demonstrations into standardized, comparable evidence tied to industrial demand.
  • A competition is not proof of generalized capability, but real job constraints should expose reliability gaps that staged demos hide.
Open at Leiphone Ask Claude ChatGPT

US executives are lukewarm on China’s Shenzhen APEC events

# Analysis SCMP 58m

Chinese coverage says senior US executives are unenthusiastic about Shenzhen’s APEC summit, citing weak scheduling, few expected breakthroughs, and limited room for meaningful engagement.

More
  • The account suggests Beijing is struggling to convert a high-profile summit into substantive private-sector engagement, despite inviting leaders who joined Trump’s China trip.
  • The issue is not attendance alone: it reflects uncertainty over whether political symbolism can overcome US-China business friction and unclear policy deliverables.
  • For US executives, the opportunity cost and lack of concrete meetings appear to matter more than the summit’s formal diplomatic status.
Open at SCMP Ask Claude ChatGPT

The US is pursuing cheaper hypersonic missiles aimed at China’s defenses

# Analysis SCMP 13h

The US Navy is developing Castelion’s lower-cost Blackbeard hypersonic missile for carrier aircraft, emphasizing mass over exquisite cost as a challenge to Chinese coastal defenses.

More
  • The weapon is designed to exceed Mach 5 and target ships and land installations, creating a distributed threat rather than relying only on a few expensive missiles.
  • The China dimension is substantive: the reported operational concept is explicitly about stressing defenses around Chinese warships and coastal bases.
  • For Chinese planners, affordability may matter more than peak performance because it changes the attacker’s ability to saturate sensors and interceptors.
  • The article is expert commentary on a development program, not evidence that the missile is deployed or combat-proven.
Open at SCMP Ask Claude ChatGPT

AI microdramas now consume more Chinese tokens than software development

# Signal TMTPost 34m

Industry data cited by Chinese media put AI dramas and video generation at 55% of domestic token use, exposing entertainment as the largest near-term demand source for video models.

More
  • The reported workload combines scripting, storyboards, character design, video generation, voice, and editing, with repeated generation making discarded clips a major cost.
  • The account estimates a 120-minute AI live-action drama can cost 30,000–40,000 yuan (about $4,200–5,600) in model compute, roughly half of total production cost.
  • Video providers are gaining recurring business-to-business API revenue as studios industrialize production, while content producers bear the risk that most generated works will not monetize.
  • The sector also shows classic Chinese platform dynamics: models, production tools, copyright, distribution, and traffic are being recombined inside large-company ecosystems.
  • The figures come from industry disclosures and media estimates, but the direction is strategically important: consumer entertainment may be a larger compute sink than coding.
Open at TMTPost Ask Claude ChatGPT

China’s foreign minister travels to four countries to reinforce trade corridors

# Policy SCMP 13h

Wang Yi will visit Azerbaijan, Croatia, Montenegro, and Bulgaria as Beijing seeks stronger trade links and geopolitical influence across the Caucasus and Balkans.

More
  • The itinerary combines a regional security forum with bilateral visits, using the foreign minister’s travel to connect diplomatic presence with corridor and trade interests.
  • The choice of smaller European and Eurasian states suggests Beijing is working below the level of major-power summits to build durable commercial and political access.
  • The visit is a core China-policy signal even though the supplied account gives few concrete agreements or economic commitments.
Open at SCMP Ask Claude ChatGPT

China’s office-agent market is consolidating around three ecosystems

# AI TMTPost 26m

Tencent, Alibaba, and ByteDance are consolidating office agents, but headline user numbers are not comparable; enterprise payment and workflow depth matter more than reach.

More
  • The reported market reached 102 million monthly active users and 1.387 billion monthly uses by July, while native apps and PC clients grew far faster than web access.
  • Tencent reports 20 million monthly users for WorkBuddy, Alibaba reports 30 million cumulative public-test users but only about 1.58 million actual monthly users, and ByteDance’s 382 million figure covers its entire app.
  • WorkBuddy is the only one of the three described with paid metrics: about 14,000 paying seats and enterprise retention near 80%, though the source gives a wide range for customer count.
  • Tencent is betting on ecosystem distribution and an open platform; Alibaba is betting on Qwen, cloud, and organizational data to reach deep enterprise workflows.
  • The market is not yet a clean three-way contest: each company is measuring a different layer of adoption, and the durable winner may be the one that owns permissions, documents, and business processes rather than chat volume.
Open at TMTPost Ask Claude ChatGPT

Gasoline cars fall below half of global new-car sales

# Signal Solidot 18h

Gasoline vehicles accounted for 49% of global new-car sales in the first half of 2026, with China showing the steepest decline in gasoline demand among major markets.

More
  • Pure-electric sales rose 12% to 6.87 million units, while conventional hybrids reached 7.27 million and 18% share.
  • China’s gasoline-car sales reportedly fell 26%, compared with 13% in Europe, while Chinese electrification trends help explain the global crossover point.
  • The statistic is a hard market signal rather than a forecast: powertrain substitution has become visible in the sales mix even as total vehicle demand weakened.
  • The supplied figures come from an industry dataset and should be checked against official registrations before being used for investment planning.
Open at Solidot Ask Claude ChatGPT

Tencent puts a 28-year-old former OpenAI researcher at its AI controls

# Analysis TMTPost 37m

Tencent has concentrated its model and infrastructure units under former OpenAI researcher Yao Shunyu while sharply increasing capital spending, signaling a more aggressive AI phase.

More
  • Yao reportedly leads both AI infrastructure and large-model teams, and Tencent has folded earlier research groups into a centralized foundation-model organization.
  • Tencent’s first-half capital expenditure reached 84.72 billion yuan (about $11.9B), up 81.9% year on year and above the prior full-year figure.
  • The governance choice is notable: Tencent is combining young research leadership with its distribution, messaging, and enterprise ecosystems rather than creating a stand-alone research lab.
  • WorkBuddy is the visible product bet, but the deeper shift is capital allocation and organizational consolidation after Tencent judged its earlier AI response too cautious.
Open at TMTPost Ask Claude ChatGPT

White Rhino raises $100M to scale autonomous delivery fleets

# Analysis 36Kr 22h

Chinese L4 delivery-robot company White Rhino raised $100M to improve end-to-end autonomy, expand city operations, and build a vehicle-and-service ecosystem.

More
  • The round combines industrial, government, and financial capital, reflecting how Chinese autonomous delivery companies need roads, deployment sites, manufacturing, and financing—not just better models.
  • White Rhino’s RX vehicle reportedly completed more than 200,000 kilometers of validation and targets an eight-year, 300,000-kilometer design life, signaling a move toward automotive-grade products.
  • Its larger R24 vehicle and fleet-as-a-service model extend the business from selling robots to selling adaptable delivery capacity across customers and demand cycles.
  • The remaining barriers are local road permissions, cross-city replication, and long-tail safety in rain, darkness, and mixed traffic.
  • The company’s strategic challenge is closing the loop between autonomy, vehicle design, public-road access, and operating economics.
Open at 36Kr Ask Claude ChatGPT

Taiwan considers keeping fighter pilots until age 58

# Policy SCMP 15h

Taiwan plans to raise the retirement age for military pilots as 66 US-built F-16Vs arrive, highlighting a personnel shortage that affects the island’s defense readiness.

More
  • The proposed retention scheme addresses a familiar military bottleneck: acquiring aircraft faster than experienced crews can be trained.
  • The China dimension is direct because Taiwan’s air-force readiness is shaped by the prospect of conflict with Beijing, though the supplied article does not report a Chinese response.
  • Keeping veteran pilots longer can buy near-term capacity but may also reveal limits in recruitment, training pipelines, and retention incentives.
Open at SCMP Ask Claude ChatGPT

China’s jewelry boom is turning into a liquidity and governance crisis

# Culture TMTPost 21m

Chinese jewelry chains are failing despite high gold prices as weak retail traffic, leverage, unpaid wages, and illiquid inventory expose the limits of brand-led expansion.

More
  • Qianye Jewelry reportedly had only about 930,000 yuan (about $131,000) in cash against more than 1.5 billion yuan (about $211M) of inventory, showing why rising gold prices do not guarantee liquidity.
  • The company shifted from premium direct stores to franchising and then online sales, but e-commerce revenue later collapsed, suggesting channel migration did not repair the underlying brand economics.
  • The broader lesson is that gold retailers earn on turnover, trust, and working-capital discipline; inventory appreciation can coexist with an inability to pay staff or publish accounts.
  • The account places the failures in a wider pattern of Chinese listed jewelry companies facing delisting, alleged financial irregularities, and control problems.
Open at TMTPost Ask Claude ChatGPT

China’s space-compute startups surge after Musk’s satellite push

# Analysis 36Kr 22h

Chinese investors are rushing into orbital AI-compute projects after SpaceX publicized a space data-center plan, despite limited evidence that the business model is ready.

More
  • The pitch combines constant solar power, vacuum radiation cooling, and satellite networking to escape terrestrial energy and data-center constraints.
  • Chinese startups have reportedly raised more than 100 million yuan (about $14M) rounds, while some projects use the space-compute label mainly to lift valuations.
  • The more credible near-term use case may be processing Earth-observation data in orbit—ocean, agriculture, carbon, and disaster monitoring—rather than replacing terrestrial data centers.
  • China’s commercial-space sector still lacks a clear recurring-revenue model, so orbital compute is being asked to supply the next economic narrative after launch, manufacturing, and constellations.
  • The strategic importance is real if in-orbit compute reduces downlink costs and enables autonomous sensing, but launch cadence, radiation, thermal design, and networking remain hard constraints.
Open at 36Kr Ask Claude ChatGPT

Orbbec abandons lawn-mower robot manufacturing to focus on sensors

# Analysis Leiphone 18h

Chinese 3D-vision company Orbbec is reportedly dropping its lawn-mower robot ODM business after only several thousand units shipped, while keeping the sensing-module line.

More
  • The retreat reflects three constraints: tighter US authorization rules for foreign-made advanced robots, a crowded lawn-care market, and a customer whose volume could not support an ODM product line.
  • Orbbec is moving back toward its capability boundary as an upstream perception supplier rather than owning the full vehicle, channel, compliance, and service stack.
  • The case is a useful counterweight to China’s enthusiasm for embodied AI: sensing expertise does not automatically translate into viable branded hardware.
Open at Leiphone Ask Claude ChatGPT

China’s power retailers face a forced test of risk-management skills

# Policy TMTPost 7m

Losses across several provinces are exposing electricity retailers that relied on fixed-price bets; new guarantees and margin-sharing rules favor firms with hedging and service capabilities.

More
  • The supplied analysis describes a transition from open-ended price arbitrage toward credit requirements, contract discipline, and operational capability.
  • Performance increasingly depends on load forecasting, spot-price risk management, hedging, and services such as storage, demand response, green certificates, and carbon management.
  • The policy direction matters beyond power trading: it shows Chinese regulators tightening entry and survivability standards as markets become more volatile.
  • The account is an industry analysis rather than a formal regulatory notice, so the precise timing and national scope of the changes remain unclear.
Open at TMTPost Ask Claude ChatGPT

Chinese excavator sales rise 14.4% as exports lead the recovery

# Signal Jiemian 16m

China sold 22,717 excavators in September, up 14.4% year on year; exports rose 20.7%, outpacing domestic growth and reflecting stronger overseas competitiveness.

More
  • Domestic sales rose 7.16% to 9,911 units, while exports reached 12,806 units, making overseas demand the larger contributor for the month.
  • Through September, total sales rose 23.1%, with exports up 30.4% versus 16.3% domestically.
  • The data point to infrastructure and industrial-equipment exporters offsetting weak property demand at home, helped by price, improving products, and service networks.
  • Excavator sales are an imperfect economic proxy, but the export mix is a useful signal of where Chinese heavy-equipment firms are finding growth.
Open at Jiemian Ask Claude ChatGPT
Source ledger
ChinAI 403 0/0 HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk 200 0/20
Pekingnology 200 1/20
Ginger River Review 200 0/20
Sinocism 200 1/20
QbitAI 200 2/10
36Kr 200 7/30
Solidot 200 10/20
Leiphone 200 12/20
GeekPark 200 1/30
IT Home 200 12/60 +35 over cap
SCMP 200 10/50 +3 over cap
Initium 200 5/15
Jiemian 200 8/30 +8 over cap
China Digital Times 200 0/15
The Wire China 200 0/50
Techmeme 200 1/15
Hacker News 200 0/20
V2EX 200 8/50 +38 over cap
Zhidx 200 5/20
InfoQ China 200 8/20
New Wisdom 200 8/15 +7 over cap
TMTPost 200 8/16 +8 over cap