China Desk

China News, Summarized 22 Oct 2026, 08:07 UTC 34 stories 21/23 sources

Today's brief

Beijing threatens Washington over DeepSeek while its biggest platforms queue for $35 billion of Nvidia cards

Beijing answered in a day, without naming a weapon

Both Washington moves drew a reply. On Tuesday the Ministry of Commerce opposed the DeepSeek order, urged Washington to revoke it, and promised "all necessary measures". It also condemned the MiraclePlus sanctions, again naming no measure. Yesterday I asked whether Beijing would answer before the Shenzhen summit. It did, with a warning and no weapon.

THE TELL: the ministry accused Washington of ignoring understandings the two presidents had reached. That aims the blame at agencies below the White House and keeps the leaders' channel clean. Henry Huiyao Wang, founder of the Beijing think tank Centre for China and Globalisation, writes that Xi and Trump agreed a formal AI dialogue and hotline, with the next meeting in November. My inference: Beijing is protecting that meeting.

Two versions of the story. The New York Times, in Techmeme's summary, pairs China's vow of countermeasures with its rejection of FBI and AI-lab accusations of illicit distillation, meaning training on rivals' outputs. Leiphone's morning digest gives it five sentences on what Homeland Security ordered, ending with Washington "claiming" a threat. No FBI, no distillation.

Go deeper on this section: ClaudeChatGPT

Nvidia cannot fill what China ordered

The order is enormous. Reuters reported, per The Next Web, that ByteDance, Alibaba and Tencent ordered over 3 million RTX PRO 5500 cards at about $11,400 each, above $35bn. Nvidia told partners they may get 5% to 10%. Output is capped near 3 million worldwide in 2027, while Chinese orders total over 6 million. Deliveries start in the first quarter.

Why this card. It is a workstation GPU built to clear US rules: the Blackwell B300 is barred and the H200 needs licences. Leiphone's sourcing adds 84GB of memory that splits into two 42GB slices for inference, theoretical compute about double the cut-down RTX PRO 6000D, and Chinese quotes of 80,000 to 120,000 yuan (~$11,000 to $17,000).

Beijing is rationing, not only retaliating. Technology.org's write-up says China's industry ministry asked buyers in late September how many cards they wanted and why, then said it intends to approve. The state threatening countermeasures over a US procurement ban is also deciding who queues for an American supplier. At 5% to 10% of 3 million, the three firms get 150,000 to 300,000 cards. That arithmetic is mine.

WORTH NOTING: Leiphone relays unconfirmed overseas chatter that Nvidia is moving GB202 chip supply away from the RTX 5090. Chinese prices for that card are already up over 50% in a month, past 50,000 yuan (~$7,000).

Go deeper on this section: ClaudeChatGPT

DeepInfer opens the door to two vendors, mostly one

The evidence is lopsided. TMTPost's analysis of DeepSeek's DeepInfer release says Ascend and Nvidia are listed as peer build targets. But published benchmarks lean on Nvidia, Ascend side-by-sides are thinner, and the drop is framed as offline inference. Cross-node cache management and load balancing are missing. That matches yesterday's read: the engine is out, the cache tier is not.

OpenAI moved within a day. Zhidx reports that OpenAI's API now offers 12-hour cache retention on GPT-5.6 and later. Cache-write pricing rises from 1.25 to 2.5 times the uncached input price. Codex and Work users get "permanent" caching. Zhidx suspects a billing waiver rather than stored tensors, and notes DeepSeek's design targets up to three days against OpenAI's 30-minute default. Cache lifetime is becoming a price lever dressed as infrastructure.

Go deeper on this section: ClaudeChatGPT

What people are chewing on

Starbucks wants a purchase for a seat. From 22 October Starbucks China requires a purchase to sit. Boyu Capital, the private-equity firm, took 60% in April; TMTPost says nothing shows it demanded this. The free-seating promise began as damage control after a 2018 Philadelphia arrest, and the US rolled it back last year. The same day the chain pledged over 300 million yuan (~$42m) for peak-hour staff and about 1,000 refits.

The third place became a rent line. My read: it was a brand premium, and once Luckin and Cotti sold coffee at a fraction of the price, the sofa had to pay for itself.

Who burns the compute? A media claim, relayed by Leiphone, says 55% of China's AI token use goes to short drama and video generation, 24% to e-commerce ads and 15% to research. It also puts nearly 200,000 yuan (~$28,000) of compute on one 12-episode drama, ten times yesterday's TMTPost production-cost figure. Doubao's PR chief called the data unreliable, saying only about 10% of ByteDance's compute goes to video generation. Industry tokens versus one company's compute: the denial settles nothing.

Trending, manufactured. Sam's Club says the "quality collapse" videos that hit Weibo's top hot search were AI-generated. Products shown as pulled are on sale or were never stocked, it says, and thousands of near-identical posts appeared. That is the company's claim. Consumer outrage is now a farmable commodity.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • MiniMax was not a tie. Yesterday I said M3.1 Flash tied DeepSeek's V4.1-Flash-1017 on agent benchmarks. Leiphone's numbers show a split. DeepSeek leads coding, 77.0% to 73.7% on DeepSWE v1.1. MiniMax leads Automation-Bench. On a 3D-campus build MiniMax took 55.4 minutes against about 121, at roughly 6.5 yuan (~$0.92) each.
  • Agent gates are global. On 15 October I framed China's apps as deciding agents must be invited in. Zhidx says 155 apps refused Doubao's phone assistant, while Amazon, Yelp, eBay, Delta and JPMorgan block Meta's Muse. Muse reached 10.9 million downloads in 40 days. My frame was too local. The fight is over who owns the checkout.
  • Honor absorbs and raises at once. On 19 October it lifted the X80 Pro Max by 500 yuan. Today's gaming phone, the WIN2, starts at 4,999 yuan (~$704) after subsidy. The CEO says Honor absorbs memory costs, but Honor concedes it cannot say for how long.
  • The export draft widened. Sinocism's Sharp China describes proposed controls on AI, solar-related technology and machinery. I have not seen the solar or machinery text, so my "ledger, not gate" read covers only AI.
  • DeepSeek is investing in robots. On 14 October I said its round was filling with suppliers and customers. Leiphone reports, unverified, that DeepSeek is in the startup of Yu Tao, a former Tencent AI Lab researcher, and may lead Winverse AI at about $400m. In August it put 140.8m yuan (~$19.8m) into Unitree's IPO placement.
  • A 17-year-old may have left for Palo Alto. Chen Guangyu, co-first author of the Attention Residuals paper that Moonshot AI's Kimi K3 uses, left Moonshot on 15 October. An anonymous tracker says he joined Discovery Loop, the Jeff Dean startup reportedly seeking a $50bn valuation. Unconfirmed; the company won't comment. If true, 2-million-yuan starting packages are not what holds talent.

Go deeper on this section: ClaudeChatGPT

The river · 34

DeepSeek reportedly raises about $14 billion in a landmark funding round

# Zhidx 3h

Chinese coverage frames the reported raise as fuel for data centers, open models and a possible IPO; US coverage also highlights the scale and CATL and Tencent’s roles, but says little about that strategic agenda.

Also: Techmeme

More
  • The Chinese report, citing Bloomberg and Reuters, says DeepSeek raised about 100 billion yuan (roughly $14 billion), with CATL and Tencent among the largest investors; the details are attributed to unnamed sources.
  • The report says the fundraising target rose from 50 billion yuan (roughly $7 billion) to 80 billion yuan (roughly $11 billion) and then 100 billion yuan (roughly $14 billion) during investor discussions.
  • The article connects the financing to data-center investment, further open-model development and IPO preparation; these are reported aims, not confirmed outcomes.
  • The Techmeme account leads with the round’s size and major contributors, while the Chinese account stresses DeepSeek’s open-model strategy and capital-market implications.
  • If completed at this scale, the round would give a Chinese model developer resources to expand infrastructure while maintaining an open-weight approach; its impact depends on execution and whether the reported terms are confirmed.
Open at Zhidx Ask Claude ChatGPT

China’s proposed export controls could reshape its open-model strategy

# ChinaTalk 18h

Beijing is proposing export controls across AI technologies; the draft puts open model weights, chip design and other capabilities in tension with China’s effort to build global influence through openness.

More
  • The proposal would expand the national catalogue of technologies restricted or prohibited from export; it is still a draft open for public comment, not a law in force.
  • The analysis argues that earlier controls often targeted technologies where China held supply-chain leverage, making this a potential shift from controlling industrial processes to controlling AI capabilities.
  • The draft could affect model weights and architectures as well as chip design, memory, cybersecurity and AI content detection, according to the article.
  • A central implementation question is how licensing authority will be divided between central ministries and provincial governments; the text says that division matters to whether controls are predictable in practice.
  • The strategic trade-off is whether Beijing prioritizes limiting sensitive technology outflows or sustaining the global adoption and commercial credibility that Chinese open models have gained.
Open at ChinaTalk Ask Claude ChatGPT

Chinese team builds a single-photon detector with more than 100 channels

# SCMP 3h

A Shanghai team says a new interconnect design raises single-photon detection beyond the previous 32-channel ceiling, potentially enabling higher-bandwidth photonic quantum systems.

More
  • The report describes an interconnect scheme supporting more than 100 detection channels, compared with a previously reported maximum of 32.
  • Single-photon detectors are key components in photonic quantum technologies, where increasing channel count can expand parallel signal detection.
  • The article says the design maintains detection accuracy, but the supplied account gives no performance numbers, fabrication details or independent validation.
  • The result concerns a component bottleneck, not a complete quantum network; sources, switching, transmission loss and system integration remain necessary.
  • Its importance will turn on whether the channel count and accuracy hold in larger integrated systems rather than only in a laboratory demonstration.
Open at SCMP Ask Claude ChatGPT

Chinese team reports a full silicon-28 gas-to-wafer process

# IT Home 13m

A Zhejiang University team reported high-purity silicon-28 gas and a four-inch epitaxial wafer, extending domestic isotope work into a key substrate step for silicon quantum devices.

More
  • The team reported silicon-28 abundance of 99.99576% in its gas and 99.9937% in the wafer’s epitaxial layer; chemical purity of the gas was reported separately at 99.99998%.
  • Silicon-28’s nuclear spin is zero, making it a promising substrate for silicon quantum devices that are sensitive to nuclear-spin noise.
  • Earlier domestic work described in the article achieved isotope enrichment; the new report adds conversion into a semiconductor-grade epitaxial layer.
  • The results establish material measurements, not a working quantum processor: coherence time, gate fidelity, interface defects and device fabrication remain untested in this report.
  • The useful milestone is a domestic path through both gas production and wafer growth; whether it reduces dependence on external material supply will depend on reproducible scale and device performance.
Open at IT Home Ask Claude ChatGPT

Platforms are blocking personal AI agents, including a Chinese phone assistant

# Analysis Zhidx 23h

Chinese coverage links global platform blocks on agents such as Meta’s Muse to China’s 155-app refusal of Doubao’s phone automation, exposing a clash over user delegation, platform control and commercial access.

More
  • The article says Amazon and several other services have restricted agents, while China’s Doubao phone assistant was refused automated operation by 155 apps, including major social, commerce and payment services.
  • Platforms cite terms of service and security; agent providers argue that users should be able to delegate actions, while platforms also face reduced control over ads, recommendations and transaction flows.
  • The Chinese example is distinctive because app restrictions are presented as a concrete limit on system-level phone automation, rather than only website access by cloud agents.
  • The article describes a proposed personal-agent protocol effort, but says major providers have not all joined and the legal boundary remains unsettled.
  • The core engineering and business issue is delegated identity and authorization: an agent needs permission to act for a user without becoming an unaccountable bot that bypasses the platform’s own controls.
Open at Zhidx Ask Claude ChatGPT

A Chinese startup brings a predictive world model to real robots

# AI QbitAI 1d

Tsinghua-linked Qianjue says its Polibrain model predicts task-relevant physical state changes rather than generating every future frame, and is being adapted across several robot forms.

More
  • Polibrain is presented as a predictive world model that links environment state, action prediction, planning, execution and feedback.
  • Instead of reconstructing full future images, the system aims to encode task-relevant variables such as object location and the robot’s relation to a target, reducing computation on irrelevant visual detail.
  • A demonstration described in the article has a quadruped robot locate and pick up litter using a manipulator; the report does not provide a full benchmark or independent evaluation.
  • The company says it is adapting the same intelligence across quadruped, humanoid, wheeled and dual-arm platforms, an attempt to separate reusable world understanding from robot-specific control.
  • The key validation question is whether latent state predictions generalize to new environments and recover reliably when actions fail.
Open at QbitAI Ask Claude ChatGPT

A Chinese economist argues public investment must lead the demand push

# Policy Pekingnology 15h

A Chinese policy commentary makes the case for debt-funded public investment in services, housing and infrastructure, arguing that weak spending can worsen rather than cure debt dynamics.

More
  • The essay argues that China still has room to invest in public services, affordable housing and maintenance of existing infrastructure, not just new construction.
  • It challenges the idea that lower borrowing automatically improves debt sustainability, arguing that slower growth can weaken incomes and push debt ratios higher.
  • The argument follows official calls to accelerate projects across six infrastructure networks, including power, computing, logistics and urban utilities.
  • The piece represents a policy argument, not an announced spending package or a settled government position.
  • The practical question is whether investment can raise household demand and productivity rather than add capacity that lacks durable use or returns.
Open at Pekingnology Ask Claude ChatGPT

China is developing identity standards for workplace AI agents

# Policy Leiphone 1d

A Shanghai platform has begun issuing digital identities for AI agents, reflecting a broader Chinese push to define how agents are authorized, monitored and held accountable inside organizations.

More
  • The article says a Shanghai agent-identity platform launched with 30 participating organizations; one identity code is intended to correspond to one agent.
  • National guidance described in the report calls for agent registration, identity management, capability declarations and differentiated controls based on permissions and autonomy.
  • Identity is only one layer: companies still need to define which person or organization an agent represents, what systems it may access and which actions require human approval.
  • The examples range from factory planning assistants to business-data agents, illustrating a move from individual productivity tools toward agents assigned end-to-end job responsibilities.
  • The governance test is traceability across agent handoffs: logs, authorization boundaries and named human owners are needed if errors or unauthorized actions occur.
Open at Leiphone Ask Claude ChatGPT

Tencent opens its in-house search stack to agent developers

# AI Zhidx 19h

Tencent has made Yuanbao Search available through APIs and agent integrations, turning an internally used retrieval system into a developer service for grounding answers in fresh web and structured data.

More
  • The service is offered through an API, skill and MCP integration, according to the report, and is said to draw on a large index of public and Tencent-owned content.
  • The article emphasizes structured vertical results for categories such as weather and stock prices, which can spare agents from scraping pages and extracting fields.
  • Search quality is especially consequential for agents because stale or incorrect retrieval can propagate into later tool calls and actions.
  • External developers trade control of their own crawl and ranking pipeline for dependence on Tencent’s source coverage, uptime and result quality.
  • The report’s testing is editorially conducted, not an independent quality benchmark; the long-term differentiator will be freshness, source transparency and reliability.
Open at Zhidx Ask Claude ChatGPT

A Chinese startup proposes teams of AI doctor avatars for hospitals

# AI QbitAI 20h

Tsinghua AIR spinout Zijing Health is testing a hospital system where specialist AI avatars advise clinicians and surface disagreements rather than replacing the doctor’s final judgment.

More
  • The system is described as supporting single-specialty consultations and multi-agent collaboration for cases involving several departments.
  • One example has a respiratory specialist agent flag medication concerns for a patient being seen for hyperthyroidism with an asthma history.
  • The product presents differing agent opinions to the physician rather than forcing a single consensus answer, potentially making uncertainty more visible.
  • The article describes pilots and product design but supplies no controlled clinical outcomes or evidence of improved patient safety.
Open at QbitAI Ask Claude ChatGPT

China is shifting robot training from labs to real production sites

# Analysis Leiphone 1d

A Chinese industry report describes a national push for embodied-AI training at real workstations, where production failures can feed back into models instead of staying isolated in demonstrations.

More
  • The article cites a gap between the millions of hours of real interaction data said to be needed for general-purpose robots and the much smaller volume currently available.
  • It reports that more than 70 embodied-AI training facilities were operating in China by June 2026, with another 46 planned or under construction; the bottleneck is converting these sites into reusable data and deployment pipelines.
  • A joint government initiative is described as using real production and service workstations for training, testing and validation rather than relying only on simulated environments.
  • The report gives an automotive-line example in which a humanoid robot performed battery-cell handling, with nearly 80,000 workpieces completed and a reported 98% uptime.
  • The infrastructure challenge is standardization: data formats, safety responsibility, transfer from simulation to real sites and shared validation facilities all determine whether the training economy scales.
Open at Leiphone Ask Claude ChatGPT

Chery sees motorcycles as China’s next family vehicle

# Culture GeekPark 17h

Chery is entering recreational motorcycles through a new venture, betting on premium leisure riding and applying car technologies while emphasizing rider culture and safety rather than gadget-led features.

More
  • The article describes a market shift in China: overall motorcycle sales remain large, but domestic demand is moving from commuting toward leisure bikes above 250cc.
  • Chery’s Zero-Dimension venture is targeting premium recreational motorcycles, where the company says Chinese manufacturers have less brand and profit power than overseas incumbents.
  • Its prototypes include a fast-charging electric scooter and a range-extender cruiser; the company is also adapting automotive thermal management, electronics and hybrid systems.
  • The team combines car-industry engineers with motorcycle specialists, and its executives say understanding riders has been a bigger challenge than transferring technology.
  • A proposed cornering leaderboard was rejected because it could encourage dangerous riding, an example of safety considerations shaping product design rather than merely adding sensors.
Open at GeekPark Ask Claude ChatGPT

Beijing and Taipei clash over mainland travel for Sun Yat-sen events

# Policy SCMP 7h

Beijing condemned Taipei’s ban on officials attending mainland commemorations of Sun Yat-sen’s 160th birthday, turning a shared historical figure into a fresh point of cross-strait political dispute.

More
  • The report says Taipei barred Taiwanese officials from attending events in Beijing marking the anniversary of Sun Yat-sen’s birth.
  • Beijing criticized the decision, framing the commemoration as a matter of shared historical respect.
  • Sun’s legacy carries political meaning on both sides of the Taiwan Strait, making an anniversary event a vehicle for competing claims about identity and legitimacy.
  • The immediate dispute is about official travel, but the deeper signal is how historical symbols remain active in cross-strait diplomacy.
Open at SCMP Ask Claude ChatGPT

Long-term memory is emerging as a core problem for embodied AI

# AI Leiphone 5h

A Chinese startup’s BetaMem system and an IROS award-winning study both tackle how robots should preserve changing object histories, rather than treating each observation as a fresh snapshot.

More
  • The article describes IROS 2026’s best-paper award going to LT-Mem, which tracks object-location changes across repeated visits instead of retaining only the latest observation.
  • BetaMem, from Chinese startup Beta Infinite, is described as combining a temporal state graph, an append-only event log and a limited archive of original sensor data.
  • That separation lets a robot maintain its current belief while retaining evidence that could support correction; an incorrect remembered fact can otherwise contaminate later decisions.
  • The article reports benchmark-leading results for a 9-billion-parameter model on four memory tests, but does not provide enough benchmark methodology here to assess the comparison independently.
  • Memory remains a systems problem involving what to store, how to retrieve it, how to resolve contradictory observations and how to trace a decision back to sensory evidence.
Open at Leiphone Ask Claude ChatGPT

US universities are ending more partnerships with Chinese institutions

# Policy SCMP 12h

A Pentagon push for tighter oversight is prompting US universities to unwind China-linked programs, with Virginia Tech accelerating the closure of a joint degree program with Xidian University.

More
  • The article reports that Virginia Tech is moving up the end of a joint undergraduate program with Xidian University, previously expected to run until 2029.
  • It places the decision within more than 80 partnership-termination actions across 30 institutions following Pentagon pressure for greater scrutiny of research ties.
  • The change indicates that security policy is moving from research compliance toward institutional partnerships and student programs.
  • The article does not establish that every partnership has the same security rationale; universities’ risk assessments and federal guidance may vary.
  • For Chinese universities, the trend narrows access to US academic collaboration and may encourage more domestic or non-US partnerships.
Open at SCMP Ask Claude ChatGPT

China says it allowed the Philippines to resupply a grounded warship

# Policy SCMP 11h

Beijing said it permitted a Philippine resupply mission at Second Thomas Shoal, while emphasizing that its coast guard supervised the operation amid continued diplomatic tensions.

More
  • China’s coast guard said the Philippines delivered necessities to the grounded BRP Sierra Madre with Chinese permission and under Chinese supervision.
  • The report places the event amid high tensions over a controversial comment by China’s ambassador.
  • The statement’s emphasis on permission and supervision asserts Chinese control over the encounter while avoiding an immediate physical confrontation.
  • The report does not include the Philippine account in the supplied text, so the operational details and framing are one-sided here.
Open at SCMP Ask Claude ChatGPT

AI-assisted research finds PHP flaws that trigger government site shutdowns

# Analysis GeekPark 1h

Chinese coverage says AI helped identify two severe PHP memory-management vulnerabilities, prompting several governments and organizations to take websites offline before patching.

More
  • The report says the flaws could enable unauthenticated remote code execution and affected a broad range of PHP installations.
  • It traces the discovery to AI-assisted code review and says a reusable search pattern helped researchers find a second issue.
  • The article frames AI as reducing the cost of finding bugs that previously required extensive manual work; independent technical validation remains important.
Open at GeekPark Ask Claude ChatGPT

A startup sells first-person video data to robotics labs

# Analysis QbitAI 1h

Chinese coverage claims Maxinsights has built a large human-action video business supplying robotics and AI labs, highlighting the growing market for embodied-training data.

More
  • The article says the company has delivered more than two million hours of data and collected over 1.5 million hours with hand tracking and language annotation.
  • It names several prominent labs and robotics companies as customers, but the claims are not independently verified in the supplied report.
  • First-person human video can provide varied interaction examples more cheaply than collecting every behavior on robots, but transfer to robot action remains a technical challenge.
  • The company’s reported commercial traction suggests training data is becoming a distinct Physical AI supply-chain business.
Open at QbitAI Ask Claude ChatGPT

Taiwan’s new attack drone is designed to strain mainland defenses

# Policy SCMP 18h

A flight-tested Taiwanese drone modeled on low-cost attack systems is being discussed as a way to saturate PLA air defenses, though analysts differ on its likely role in a conflict.

More
  • The report says the Taiwanese maker claims the drone could reach military targets on the mainland; the system has been flight-tested but is not described as operationally deployed.
  • Analysts see potential value in low-cost drones that can force defenders to expend more costly interceptors, drawing on lessons from Ukraine and the Middle East.
  • Experts differ on whether the drone would be most useful against PLA forces or other targets in a conflict scenario.
  • The item is an assessment of military utility, not evidence that Taiwan has a finalized operational plan or that the system can evade defenses.
Open at SCMP Ask Claude ChatGPT

China Telecom opens a specialized model for intent classification and extraction

# AI QbitAI 56m

China Telecom’s Xing4.0-29B-Intent combines intent selection with structured parameter extraction, targeting fast, schema-constrained tasks such as customer-service routing and device commands.

More
  • The model and training materials are being open-sourced, including 2.75 million training examples, according to the report.
  • Its unified schema-driven parsing approach returns structured fields such as intent, entities and parameters instead of free-form prose.
  • The company claims hundred-millisecond response and reports stronger scores than tested general models on a four-task evaluation; the benchmark is presented by the model’s developers.
  • A specialized model can be cheaper and easier to validate for narrow, high-volume workflows, but evaluation quality and deployment-specific error rates matter more than headline scores.
Open at QbitAI Ask Claude ChatGPT

A China newsletter tracks policy, youth unemployment and AI export controls

# Analysis Sinocism 9h

Sinocism’s roundup highlights a Long March anniversary, tech-industry integration guidance, youth unemployment and AI export controls, offering a broad agenda of current Chinese policy concerns.

More
  • The newsletter points to a planned anniversary commemoration and Xi Jinping speech as part of a larger propaganda push.
  • It flags policy debate over tech-industry integration, youth unemployment and proposed AI export controls.
  • As a roundup, it identifies topics rather than providing full reporting or analysis of each.
Open at Sinocism Ask Claude ChatGPT

Chinese team livestreams the training of an embodied-AI model

# AI QbitAI 22h

Shengshu Technology is publicly streaming Motubrain 2.0’s training and evaluation process before an expected open release, offering a rare view into how a robotics model is iterated.

Also: Leiphone

More
  • The planned livestream covers pretraining, task-focused post-training, agent execution and experiments in recursive self-improvement.
  • The team says the goal is to show failures and iteration as well as final robot demonstrations, rather than presenting only polished outcomes.
  • The model aims to generalize across tasks, environments and robot bodies, problems that require more than language-model-style scaling.
  • The research remains exploratory and the report supplies no independent results establishing the model’s generalization.
Open at QbitAI Ask Claude ChatGPT

Community developers build plugins around DeepSeek Harness

# AI Zhidx 22h

Chinese coverage maps a growing plugin ecosystem around DeepSeek Harness, from context inspection to file management, suggesting developer demand for extensible agent workspaces.

More
  • The article says the open-source project has more than 250,000 GitHub stars and describes a dozen third-party plugins and integrations.
  • Tools address context usage, sidebars, browser operation and plugin discovery, filling gaps before a more complete official marketplace.
  • The report notes that community plugin availability is not a security endorsement; source review remains important when granting agent tools access.
Open at Zhidx Ask Claude ChatGPT

Tesla’s profit falls as Musk promises FSD will lead in China

# Analysis QbitAI 53m

Chinese coverage pairs Tesla’s falling operating profit with Musk’s claim that FSD will lead Chinese automakers, while Shanghai production and exports provide the clearest China-specific business signal.

More
  • The article reports Tesla’s operating profit fell 56% year over year despite a 9% rise in gross profit, with higher operating expenses cited as a major factor.
  • FSD paying users reportedly reached 1.74 million globally, with monthly subscriptions exceeding half of the total for the first time.
  • Musk claimed Tesla’s driver-assistance technology would lead Chinese automakers, but the article gives no timetable for full FSD availability in China.
  • The Shanghai factory reportedly shipped 95,366 vehicles in September, its highest monthly total of the year, including strong export growth.
  • China-specific regulatory approval and local performance are the key tests of the FSD claim; global subscription growth does not establish domestic-market competitiveness.
Open at QbitAI Ask Claude ChatGPT

Chery lays out a broad technology agenda from chips to farm robots

# Analysis GeekPark 22h

Chery used its innovation conference to position itself as a technology company, presenting automotive chips, an enterprise model, driver-assistance systems, electric motorcycles and autonomous tractors.

More
  • The company announced a chip roadmap including a central-compute processor it says will reach 1,200 TOPS per chip and enter vehicles in 2027; these are company claims and future targets.
  • Its CheryMind model is said to connect hundreds of internal systems and be used by more than 60,000 employees, showing a strategy that starts with internal operations.
  • Chery also presented a world-model stack for driving and electric tractors with autonomous operation, extending its technology narrative beyond passenger cars.
  • The breadth is strategically notable, but the article is an event report and most capabilities remain company-reported or future plans.
Open at GeekPark Ask Claude ChatGPT

Pakistan and US policy shape China’s pitch for the J-35 fighter

# Analysis SCMP 10h

Pakistan is the only publicly identified interested buyer for China’s J-35 stealth fighter, while a US upgrade package for Pakistani F-16s keeps Washington central to the export calculus.

More
  • The report says the US approved a $686 million F-16 upgrade package for Pakistan, including data links, encryption equipment and service-life work.
  • Pakistan is described as the only declared interested buyer of China’s J-35 stealth fighter, though the supplied text does not report a purchase agreement.
  • The juxtaposition highlights how export controls, interoperability and the US-Pakistan security relationship complicate China’s effort to sell advanced aircraft.
  • A sale would be strategically significant, but training, sustainment, weapons integration and financing matter as much as the platform itself.
Open at SCMP Ask Claude ChatGPT

China’s weak car sales meet an AI-driven memory-chip squeeze

# Analysis TMTPost 54m

Chinese coverage links a sharp September car-sales decline to expensive, scarce memory chips, showing how AI infrastructure demand may be spilling into the cost base and launch timing of another industry.

More
  • Passenger-car retail sales in China fell 28% year over year in September, while exports rose 73.7%; exports helped support wholesale volumes despite weak domestic demand.
  • The article says new-energy vehicle penetration reached 67% even as those sales declined, while gasoline-car retail fell 43%, illustrating a shift in mix beneath the overall downturn.
  • It attributes slower new-model launches and higher vehicle costs partly to memory-chip scarcity as AI demand competes for semiconductor capacity.
  • Automotive margins are thin, so even a few thousand yuan of additional memory cost per vehicle can matter to manufacturers; the article uses Xiaomi as an example of reported pressure on margins.
  • The story connects two supply chains, but the causal weight of memory prices versus subsidies, the prior-year sales base and consumer demand remains uncertain.
Open at TMTPost Ask Claude ChatGPT

Huawei’s auto unit says autonomy will shift driving responsibility to systems

# Policy Leiphone 5h

Huawei’s automotive executive argues that the move from driver assistance toward higher automation changes who bears safety responsibility, as China advances L3 testing and standards.

More
  • Jin Yuzhi said China’s L2-or-higher driver-assistance fitment rate reached 70% by September, citing Ministry of Industry and Information Technology data.
  • The article says an L3 national standard is scheduled to take effect in July 2027, while L3 approval testing is underway.
  • Huawei’s automotive unit presents itself as a supplier rather than a vehicle competitor, an important positioning claim in its partnerships with carmakers.
  • The executive said its systems have accumulated nearly 17 billion kilometers of assisted-driving use and that safety data is publicly disclosed.
  • The regulatory transition matters because responsibility moves toward the system at higher automation levels, increasing the need for clear validation, reporting and liability rules.
Open at Leiphone Ask Claude ChatGPT

China’s token economy needs a better measure than token counts

# Analysis Leiphone 1d

A Chinese industry analysis argues that tokens are becoming a pricing interface for AI, but useful compute economics must account for task quality, latency, power and hardware as well as model usage.

More
  • The report cites a national daily token-call volume above 500 trillion as of June 2026, while cautioning that token counts differ across models and modalities.
  • It uses examples of time-of-day pricing and cache-sensitive input costs to show how usage metering is becoming more granular.
  • For inference infrastructure, it argues customers should measure delivered work, throughput and cost per useful task rather than accelerator counts alone.
  • In robotics, tokenized visual inputs are only one piece of the cost: inference time, energy, success rate and equipment costs also matter.
  • The analysis is useful as a framework, but its examples and vendor performance claims do not establish a universal pricing standard.
Open at Leiphone Ask Claude ChatGPT

iPhone Duo apps show what a foldable screen changes in China

# Culture GeekPark 7h

A hands-on review of Chinese apps on Apple’s new foldable finds the clearest gains in side-by-side shopping, navigation and document work rather than novelty alone.

More
  • The review tests Chinese services including Taobao, Amap, Xiaohongshu, Feishu and WPS across folded, unfolded and tabletop positions.
  • Amap’s dual-view driving interface and WPS side-by-side document comparison show how app-level adaptation can make a foldable useful beyond split-screen multitasking.
  • Some features remain demonstrations or limited tests, and the review does not establish long-term user adoption.
Open at GeekPark Ask Claude ChatGPT

Google and Meta open-source a PyTorch stack for Google TPUs

# Zhidx 21h

Google and Meta have released TorchTPU to make it easier to move PyTorch workloads onto Google accelerators, a challenge to Nvidia’s software lock-in.

More
  • TorchTPU connects standard PyTorch workflows to Google TPU hardware and is being integrated with tools including vLLM, SGLang and Hugging Face.
  • Google says the stack has been tested on as many as 8,000 Ironwood TPUs and reports 52% model FLOP utilization in a Qwen3-8B training test.
  • The report also cites a customer claim of more than fourfold inference throughput, but does not supply enough test detail to generalize it.
  • This is a major software-ecosystem move, but the article has no substantive China dimension despite mentioning Chinese models as workload examples.
Open at Zhidx Ask Claude ChatGPT

Xiaohongshu merges its AI group with core infrastructure

# Analysis TMTPost 19m

Xiaohongshu is tightening the link between model development, compute infrastructure and product teams, signaling a strategy centered on embedding AI in community search rather than competing as a general model lab.

More
  • The company’s internal letter announced that its Dots AI division would be integrated with the core infrastructure team, the second major AI reorganization in six months.
  • The stated priority is connecting model research, infrastructure and business needs, while the longer-horizon Hi Lab remains focused on exploratory work.
  • Xiaohongshu’s AI search products use platform posts to produce cited answers and summaries, tying model value to the platform’s user-generated lifestyle content.
  • This is a product-distribution strategy as much as a research strategy: the company is positioning AI to reinforce search and consumer decision-making.
  • The open question is whether AI-generated answers can improve utility without weakening the community’s dependence on original posts and creator participation.
Open at TMTPost Ask Claude ChatGPT

Sinocism surveys export controls, growth and China’s next policy meeting

# Analysis Sinocism 23h

A China-focused newsletter links proposed export controls with third-quarter growth, the upcoming Fifth Plenum and US chip-policy debate, but provides a show-notes summary rather than full analysis.

More
  • The topics include AI, solar and machinery export controls, rare-earth access concerns and China’s recent growth data.
  • The newsletter also flags speculation about the Fifth Plenum agenda and changing US lobbying around chip controls.
  • This is a curated agenda, not original reporting on the outcomes of those policy debates.
Open at Sinocism Ask Claude ChatGPT

Starbucks China tells customers they must buy to use seats

# Culture TMTPost 19m

Starbucks China’s nationwide seating rule has sparked online debate over the meaning of its “third place” as local coffee competition, customer price sensitivity and pressure on store economics intensify.

More
  • The new guidance asks customers to make a purchase before using seating, while preserving flexibility for pregnant people and customers with limited mobility.
  • The article connects the change to crowded high-traffic stores, long seat occupancy and pressure on average ticket size, rather than presenting it only as a conduct rule.
  • Online discussion turned the policy into a debate about whether Starbucks is abandoning its social-space identity or simply clarifying a commercial bargain.
  • Starbucks China is also adjusting store formats, staffing and products as it pursues expansion beyond major cities.
  • The rule’s effect will depend on customer response and whether higher seat turnover improves store economics without eroding the brand experience.
Open at TMTPost 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 1/20
Ginger River Review 200 0/20
Sinocism 200 2/20
QbitAI 200 7/10
36Kr 200 3/30
Solidot ERR 0/0 ConnectionError: HTTPSConnectionPool(host='www.solidot.org', port=443): Max retries exceeded with url: /index.rss (Caused by NewConnectionError("HTTPSConnection(host='www.solidot.org', port=443): Fail
Leiphone 200 14/20
GeekPark 200 6/30
IT Home 200 12/60 +37 over cap
SCMP 200 10/50 +8 over cap
Initium 200 6/15 +2 over cap
Jiemian 200 8/30 +16 over cap
China Digital Times 200 0/15
The Wire China 200 0/50
Techmeme 200 5/15
Hacker News 200 0/20
V2EX 200 8/50 +36 over cap
Zhidx 200 10/20 +3 over cap
InfoQ China 200 8/20 +12 over cap
New Wisdom 200 8/15 +2 over cap
TMTPost 200 8/19 +11 over cap