China Desk

China News, Summarized 21 Oct 2026, 04:08 UTC 34 stories 21/23 sources

Today's brief

Washington hit China's AI funders and buyers in one day, and DeepSeek shipped regardless

Washington sanctions the investor, bans the model

The target is the funder. On Monday the State Department sanctioned five Chinese entities under the Protecting American Intellectual Property Act. One is Shanghai Zhisuan Technology, founded by Ding Linwei, the ex-Google engineer convicted in January of stealing AI trade secrets. The other four are MiraclePlus, the Beijing venture firm that seeded Zhisuan, and three affiliates.

Investor liability is the precedent. The theory is that MiraclePlus took equity in Zhisuan and gave Ding a stage to pitch replicating Google's infrastructure. Its founder, Lu Qi, is a former Microsoft AI executive who developed Y Combinator's China arm, then kept it going under his own brand. Every Chinese VC now has a reason to vet founders' former employers. That inference is mine.

The DeepSeek order sat for four months. Homeland Security Secretary Markwayne Mullin signed it on 26 June; it surfaced on 19 October. Civilian agencies must stop buying from DeepSeek and must remove its products, including third-party tools "developed using DeepSeek implementations". Private use remains legal. I can't tell whether surfacing it a month before the Shenzhen summit was deliberate.

Both sides are stacking leverage. Negotiators are racing to close investment deals before the leaders meet at APEC in Shenzhen next month. Congress has the MATCH Act, which would restrict immersion lithography exports. ChinaTalk's guest says China bought about 90 such systems a year in 2024–25, nearly two-thirds of ASML's immersion exports.

The coverage gap. Bloomberg's headline sells a brand: "China's Y Combinator". State's own text describes a funding chain. No mainland outlet in today's feed mentions either action. They covered DeepSeek's code release, a fired intern's funding round and Singles' Day. I found no response yet from Beijing or MiraclePlus.

WHAT WE'RE WATCHING: whether Beijing answers before Shenzhen, and whether it names the VC firm.

Go deeper on this section: ClaudeChatGPT

DeepSeek opened the plumbing anyway

DeepInfer is the serving stack. DeepSeek open-sourced DeepInfer under Apache 2.0. It runs on Nvidia's SM100 GPUs and Huawei's Ascend 950 chips. Zhidx relays DeepSeek's claims: 155 trillion tokens a day from one cluster, and $38 of revenue per card-hour at off-peak prices. Both are self-reported, and I have seen no reproduction.

The design is worth reading. Decode runs wide expert parallelism across 64 to 128 cards. KV-cache blocks live on a roughly 100 PiB SSD cluster, so any prefill node can hit any prefix, at the cost of prefix-affinity routing. "Bounded replay" recomputes only the last 128 tokens of the sliding-window layers, nearly halving prefill compute.

The catch is what's missing. The KV-cache system and load balancers are not in the release, though a complete service needs both. My read: DeepSeek published the engine and kept the cache tier behind its unit economics. A procurement ban reaches buyers, not a repository, and this landed a day after the order.

MiniMax matched it within 48 hours. MiniMax's M3.1 Flash launched Monday at 1 yuan (~$0.14) per million input tokens and 4 yuan (~$0.56) output, with a 1M-token context and 200-plus tokens a second. By MiniMax's own numbers it ties DeepSeek's V4.1-Flash-1017 on agent benchmarks, at half DeepSeek's peak price.

Chinese labs can now raise prices. TMTPost's tally: Chinese models consumed 86.74 trillion tokens on OpenRouter in the week of 5 October, 4.5 times US models, a lead 24 weeks long. Zhipu, the lab behind GLM, raised average API prices 101% without losing customers. MiniMax tried a hike, fell 15.71% in a day and halved prices. OpenRouter excludes most direct traffic to OpenAI and Anthropic, so it measures only contestable demand.

Go deeper on this section: ClaudeChatGPT

Who the AI economy actually rewards

A fired intern raised $30 million. Tian Keyu was a Peking University doctoral student and ByteDance research intern in 2024. ByteDance says he repeatedly killed colleagues' training jobs over compute allocation. It fired him, sued for 8 million yuan (~$1.1m) and won 500,000 yuan (~$70,000). This month his unnamed world-model lab, about ten people, raised nearly $30m at a $200m valuation from 5Y Capital and IDG Capital.

The anger is about the price gap. After a 12 October podcast where he called it "plain justice", five ex-colleagues published a rebuttal. They allege worse damage and an open model passed off as in-house. Those are allegations. TMTPost's commentary reads the fury as arithmetic: a penalty worth about three years of intern pay, a reward of $200m. The investors have said nothing.

Same question, two enforcers. Washington sanctions the investor who backed a founder carrying someone else's secrets. China's internet interrogates its investors by shame. The pairing is mine; neither source draws it.

The drama factory broke. Per TMTPost, Soy Sauce Culture, an AI-drama studio founded by Huang Haonan, grew past 1,200 staff in late 2025. This year it halved targets of 1bn yuan (~$141m) revenue and 200m yuan (~$28m) profit. About 150 finished titles sit unreleased, costing 12m yuan (~$1.7m), and staff are down to 600-odd.

The unit economics are brutal. TMTPost's film-industry piece says an AI-made drama costs 10,000 to 20,000 yuan (~$1,400 to $2,800) against about 300,000 (~$42,000) for live action. But revenue per ten thousand plays has fallen from 30–100 yuan (~$4–$14) to 5–10 yuan (~$0.70–$1.40).

Yuewen opened its vault. The web-novel publisher reported first-half short-drama and animated-drama revenue of 4.3bn yuan (~$606m), triple last year. It says it has adapted under one in a thousand of its stories. Huang told reporters no AI-drama studios will remain after 2026, only one-person companies. My read: when a few people can make a drama in days, margin moves to whoever owns the rights and the shelf.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • The export draft got a rulebook. On 16 October I called Beijing's draft a gatekeeper, and on 18 October I couldn't say why labs weren't flinching. Sinocism's summary of Friday's proposal covers 23 AI technologies, from models to compute clusters. Public technology is excluded, new open releases could get automatic licensing, and API and SaaS delivery generally sit outside. Registration applies immediately, including to transfers underway. I now read it as a ledger more than a gate. DeepInfer on Ascend tests the cluster category.
  • Unitree warned on itself. After the slide from 1,100 yuan to about 420 that I tracked on 15 October, the robot maker's shares rose enough over three sessions to trip the Shanghai exchange's abnormal-trading rule. Its notice cites a trailing price-to-earnings ratio of 388.25 against 39.51 for its industry, and warns of valuation reversion.
  • The home car market kept shrinking. On 19 October I argued that Chen Deming, the former commerce minister, had diagnosed weak demand yet advised going abroad. 36Kr now cites industry data showing domestic car retail down 28% in September. Li Auto, Wenjie and Xiaomi swung to first-half losses. Wenjie will enter Europe in autumn 2027 with Bosch supplying its driver assistance.
  • CATL's profit lagged its sales. The battery maker, reportedly adding about 10bn yuan to DeepSeek's round, posted third-quarter revenue up 58.58% to 165.2bn yuan (~$23.3bn). Net profit rose only 14.54%, to 21.2bn yuan (~$3.0bn).

Go deeper on this section: ClaudeChatGPT

The river · 34

DeepSeek open-sources its large-scale inference framework

# Zhidx 17h

DeepSeek released DeepInfer and a production tokenizer component, describing the system behind V4.1 Flash as a distributed inference stack that runs on Nvidia and Huawei Ascend hardware.

Also: Hacker News

More
  • DeepInfer includes distributed inference, speculative decoding, structured output and performance kernels; DeepSeek says its production service uses the framework.
  • The report describes separated prefill and decode paths, expert parallelism and asynchronous scheduling to improve utilization across different workload phases.
  • For KV-cache reuse, the design stores blocks in a separate SSD cluster and accepts weaker prefix affinity in exchange for allowing any prefill machine to retrieve a matching prefix.
  • A short-window replay strategy replaces persistent storage for some sliding-window attention state, reportedly cutting repeated prefill computation.
  • DeepSeek says the open release excludes some cache-management and load-balancing components tied to its internal infrastructure.
  • The release makes a meaningful part of a Chinese lab's production inference stack inspectable, but the disclosed performance and revenue figures remain company claims.
Open at Zhidx Ask Claude ChatGPT

China faces renewed calls to release detained US scholar Min Zin

# China Digital Times 1d

Rights groups say Min Zin, a US citizen and Burmese democracy advocate detained in Kunming on espionage suspicions, has had little public information about his case or access to counsel.

More
  • The source says Min Zin was detained after traveling to Yunnan for an academic workshop and has remained in criminal detention for more than 100 days.
  • China's foreign ministry previously said he was suspected of espionage endangering national security; the excerpt supplies no public evidence supporting the allegation.
  • The US government designated the case wrongful detention, and a new statement from rights organizations adds pressure for his release.
  • The case matters to foreign academics and conference organizers because it shows that participation in a China-based academic event can carry serious legal risk.
  • The available account is from advocacy groups and does not include a Chinese judicial filing or response on the evidence.
Open at China Digital Times Ask Claude ChatGPT

Hong Kong's independent bookshops navigate raids, closures and self-censorship

# Initium 5h

After national-security police searched several independent bookstores and detained staff, surviving shops describe a changed daily reality and a shrinking space for cultural activity.

More
  • The report names raids and detentions involving several independent bookstores in 2026, alongside exclusions from the Hong Kong Book Fair.
  • Some shops have closed or announced plans to close; others face rent and relocation pressures after reopening.
  • Booksellers describe adapting ordinary routines around the possibility of police action, an indication of how enforcement reaches beyond formal prosecutions.
  • The story links state pressure to the conditions for selling books and holding events, not only to the fate of individual titles.
  • The long-term cultural effect will depend on whether shops can sustain programming and circulation under this heightened risk.
Open at Initium Ask Claude ChatGPT

HIL-UMI collects robot correction data without a robot body

# Leiphone 1d

A Peking University and Qiyuan team proposes a human-in-the-loop method that spots where a model's predicted actions diverge from a person's demonstration, targeting a costly bottleneck in robot post-training.

More
  • A person uses a handheld gripper while cameras record the task; a model predicts actions in the background but does not control a physical robot.
  • The system flags high disagreement as an out-of-distribution case, so demonstrations are concentrated on failures rather than collected indiscriminately.
  • An advantage estimator distinguishes task-progressing actions from unproductive motion, adding a quality signal to the corrective data.
  • The approach could reduce hardware wear and reset labor compared with real-robot DAgger-style correction, though the excerpt does not provide complete cost comparisons.
  • Its value depends on whether disagreement reliably identifies deployment failures and whether gains transfer across robots and tasks.
Open at Leiphone Ask Claude ChatGPT

Taiwan confirms funding for retired US officers to observe war games

# SCMP 2h

Taiwan's defense ministry says it funded US Indo-Pacific Command personnel, including a retired general, to observe its largest annual exercises, shedding light on deepening security ties.

More
  • The disclosure is notable because Taipei and Washington have generally kept military exchanges out of public view.
  • It puts a more explicit budgetary record behind unofficially close defense coordination as Beijing pressures Taiwan.
  • The source says the funding covered a retired general and other personnel but the excerpt does not give amounts or detailed arrangements.
  • Observers should distinguish this observer role from evidence of US active-duty forces directing the exercise.
  • The significance depends on whether future disclosures reveal a sustained institutional channel rather than a one-off visit.
Open at SCMP Ask Claude ChatGPT

Yuewen opens its IP library to AI-assisted comic-drama creators

# 36Kr 22h

Yuewen is turning its DramaBuddy tool into a creator platform linking licensed web-fiction IP, AI production and distribution, while drawing a line against AI-written novels.

More
  • The company says short-drama and comic-drama revenue reached 430 million yuan in the first half, more than triple the year-earlier period.
  • DramaBuddy is intended to match creators with licensed IP and production opportunities rather than monetize token consumption alone.
  • The platform says it has served more than 5,000 creators and offers subsidies, but access to valuable IP remains conditional on demonstrated production capability.
  • The article reports that AI has reduced production labor and that one studio cut staff from a peak of 1,500 to several hundred; this is a sector example, not an industry-wide count.
  • Yuewen says it will use AI for adaptation but not original literature, and is adding controls against AI-generated filler and leaderboard gaming.
  • The strategy makes rights clearance, creator compensation and editorial standards central to scaling cheap AI-generated content.
Open at 36Kr Ask Claude ChatGPT

MiniMax's M3.1 Flash combines agent benchmarks with low token prices

# AI QbitAI 14h

A second Chinese report says MiniMax's Flash model performs strongly on coding, computer use and tool benchmarks while targeting much lower inference costs.

Also: InfoQ China, V2EX

More
  • The article cites an output rate above 200 tokens per second and prices of 1 yuan per million input tokens and 4 yuan per million output tokens.
  • It describes sparse attention and agent workflows as part of the product's performance story, but does not provide a reproducible architectural paper.
  • The account frames low cost as enabling longer, multi-agent workflows, where retries and tool calls can otherwise dominate the bill.
  • These figures are vendor-linked reporting and should be tested against comparable workloads and realized task costs.
Open at QbitAI Ask Claude ChatGPT

Could restricting DUV lithography exports slow China's AI-chip progress?

# Policy ChinaTalk 17h

ChinaTalk's discussion argues that immersion lithography is a major point of leverage, while questioning whether Congress can close the export-control gap with allied support.

More
  • The discussion says China imported roughly 90 immersion lithography systems annually in 2024 and 2025, a figure it uses to argue existing controls leave a significant gap.
  • It examines the proposed MATCH Act's potential to restrict sales and servicing, and whether the Netherlands and other allies would cooperate.
  • The strategic question is whether delaying tool access would materially constrain China's advanced-chip output or mainly accelerate domestic alternatives.
  • This is a policy argument and podcast framing, not a settled estimate of China's future manufacturing capacity.
Open at ChinaTalk Ask Claude ChatGPT

MemT puts a learned memory pathway inside the Transformer

# AI QbitAI 15h

Memory Tensor's open MemT 1.0 models encode experience in an internal, updateable state and report modest gains in repeated software-engineering tasks.

More
  • The design adds a memory computation branch alongside attention, with learned mechanisms for writing, updating and retrieving experience.
  • The authors compare this with retrieval-based text memory: internal state may make procedural reuse more direct, while being harder to inspect, delete or restore precisely.
  • On 500 SWE-bench Verified tasks run three times each, the report gives Nano a 69.73 percent average success rate versus 66.07 percent without experience.
  • The reported improvement in all-three-run success was larger than the average score gain, suggesting repeatability may be the more relevant claim.
  • The results are research benchmarks, not proof that internal memory beats external stores in production systems.
Open at QbitAI Ask Claude ChatGPT

TokenFleet targets inference bottlenecks with cache-aware scheduling

# AI QbitAI 20h

Chinese infrastructure vendor Isilicon says its scheduler cut P95 time to first token by 32.4 percent in a GLM 5.3 test and improved goodput through workload-aware routing.

More
  • TokenPilot weighs KV-cache reuse against queue length and predicted latency rather than routing by request count alone.
  • FleetMind coordinates capacity across data centers but only when model versions, permissions and data-location rules allow a request to move.
  • SLOMax separates interactive, long-context and batch workloads so slow jobs are less likely to crowd out latency-sensitive requests.
  • The vendor's reported latency and goodput improvements come from its own test and estimate, not an independent benchmark.
Open at QbitAI Ask Claude ChatGPT

China briefing: Li Qiang, a new data group and Fifth Plenum watch

# Policy Sinocism 5h

Sinocism's roundup flags Premier Li's regional diplomacy, the launch of a National Data Group, Fifth Plenum expectations and a rare-earth dispute with Japan.

More
  • The excerpt is a newsletter table of contents, not full reporting on any one item.
  • It signals that Chinese governance, data institutions and economic leverage are on the day's policy agenda.
  • The roundup also notes Li Qiang's meetings at the Shanghai Cooperation Organization gathering and travel to Russia.
Open at Sinocism Ask Claude ChatGPT

Chinese buyers race to secure Nvidia's RTX PRO 5500 workstation GPU

# Analysis Leiphone 17h

A market report says demand for the 84GB workstation card is surging in China, but orders, expected deliveries and actual performance remain uncertain.

Also: Solidot

More
  • The article says buyers value large memory and multi-instance GPU support, which can let two isolated workloads share a card.
  • It reports expected prices of 80,000 to 120,000 yuan and a possible delivery timeline extending into 2027.
  • The source cites more than three million Chinese orders against a much smaller quarterly allocation, but these are channel estimates rather than confirmed shipments.
  • The buying rush signals demand for AI-capable hardware below the data-center accelerator tier, while export compliance and supply allocation remain decisive.
Open at Leiphone Ask Claude ChatGPT

Tsinghua team builds a one-gram wearable brain monitor

# AI SCMP 7m

Researchers from Tsinghua and a PLA medical school describe a wireless, one-gram device that tracks thousands of neurons in untethered lab animals.

More
  • The reported Airscope aims to remove the tether that can constrain animal movement during neural recording.
  • The unusually low mass is relevant because it may let experiments sample behavior with less device-imposed burden.
  • The excerpt gives no performance benchmarks or study details, so the claimed world-first status remains the researchers' characterization.
Open at SCMP Ask Claude ChatGPT

Li Auto unveils its in-house battery and drive technology stack

# Analysis QbitAI 3h

Li Auto says its new system combines fast-charging cells, power electronics, software and a charging network, alongside a lifetime battery-fire guarantee under stated conditions.

More
  • The company claims its 5C lithium-iron-phosphate cell has 2.6 percent higher energy density than its 2C cell, an unusual attempt to improve charge rate without sacrificing density.
  • Its battery-management algorithms target state-of-charge error below 3 percent and more than 40 percent higher winter peak power, according to Li Auto.
  • A 24-hour no-fire claim after a single-cell thermal failure and the replacement-car guarantee are company claims, not independent safety findings.
  • The launch illustrates Chinese automakers' push to control more of the battery, inverter and charging stack, but fleet data will matter more than stage demonstrations.
Open at QbitAI Ask Claude ChatGPT

Autonomous-delivery company Neolix raises nearly $400 million

# Analysis Leiphone 23h

Neolix says it will use new funding to expand operations and train its end-to-end driving model as Chinese rules move toward broader unmanned delivery trials.

More
  • The company says it operates the world's largest delivery-robot fleet and has begun deployments across several overseas regions; those are company claims.
  • Its stated architecture combines cloud, edge and vehicle models with operational data, aiming to improve fleet performance through a deployment loop.
  • The report situates growth alongside national plans for unmanned delivery pilots and logistics automation.
  • Commercial utilization, safety performance and local operating permissions will determine whether the funding translates into durable infrastructure.
Open at Leiphone Ask Claude ChatGPT

Tencent and Ant back a former ByteDance executive's personal AI startup

# AI Zhidx 44m

Today AI says it has raised $200 million for an assistant designed to remember work and personal context and connect to Chinese apps.

More
  • The assistant can connect to WeChat, WeCom, Tencent Meeting, Baidu Netdisk and Kingsoft Docs, positioning it as a persistent layer across existing services.
  • The founder has worked across Teambition, Alibaba Cloud Drive, Feishu and Doubao, a career path linking several major Chinese platform ecosystems.
  • The report describes product ambitions and investor backing, but offers no retention, privacy or quality evidence.
Open at Zhidx Ask Claude ChatGPT

Chinese aircraft and ships track Japan-US drills near the Ryukyus

# Policy SCMP 16h

Japan says PLA aircraft, drones and warships monitored exercises focused on a possible Taiwan contingency, including a supply ship that could support Liaoning.

More
  • The reported flights through international airspace triggered a Japanese fighter response.
  • The drills were explicitly framed around a possible Taiwan scenario, making the activity part of wider cross-strait and regional signaling.
  • The excerpt provides Japan's account but no Chinese military explanation or confirmation of intent.
Open at SCMP Ask Claude ChatGPT

iQiyi spent 580 days embedding AI across film and television

# Analysis Leiphone 21h

iQiyi's AI push spans script analysis, staff training, production tools and a viewing feature that lets users skip through selected plot points.

More
  • The company trained models on structured human ratings of story beats in scripts, illustrating the labor needed to translate editorial judgment into machine-readable labels.
  • A reported 'skip-watch' feature uses viewing behavior and model analysis to identify plot highlights, changing how long-form dramas can be consumed.
  • The account describes company-wide AI training and internal creative competitions, suggesting adoption depended on organizational change as much as model capability.
  • Generative backgrounds and script tools are entering production, but the article does not establish their effect on budgets or creative employment.
Open at Leiphone Ask Claude ChatGPT

Aito plans European expansion with Bosch supplying overseas driver-assistance systems

# Analysis 36Kr 18h

Aito is reportedly preparing European models with Bosch's driving, cockpit and chassis systems, a notable shift from Huawei-centered domestic integration.

More
  • The report says Bosch has begun development and that high-speed navigation-on-autopilot is targeted for production around September 2027.
  • It describes the overseas arrangement as similar to Huawei Inside, with Bosch providing a complete intelligent-vehicle package.
  • Aito's move reflects the localization and regulatory burden Chinese premium EV brands face in Europe, and the value of established suppliers' compliance experience.
  • Seres disputed some organizational claims in the article, so the partnership details and timetable remain reported rather than confirmed.
Open at 36Kr Ask Claude ChatGPT

Chinese virtual power-plant startup raises several hundred million yuan

# Analysis 36Kr 21h

Hefei-based Yuanli Energy says its software combines distributed power assets, market trading and AI forecasts, with a pilot shifting a storage project from negative to positive returns.

More
  • The company says its platform connects solar, storage, charging and industrial loads, then participates in power trading and grid services.
  • A cited 2.5 MW/5 MWh Hubei project reportedly moved to about a 4 percent return after local pricing shifted away from fixed peak-off-peak rates.
  • The key business distinction is between assets merely connected to a platform and resources actually eligible to trade in local markets.
  • Provincial rules differ, making market access and settlement integration as important as forecasting models.
Open at 36Kr Ask Claude ChatGPT

Chinese AI infrastructure firms say interconnects, not GPUs, are the bottleneck

# AI TMTPost 34m

Industry sources say network communication can consume 30 to 50 percent of large distributed training time, shifting attention toward interconnect chips and congestion control.

More
  • The article's core argument is that large clusters can leave expensive accelerators idle while waiting on data movement.
  • It maps a layered stack from die-to-die links to server fabrics and scale-out networks, with different protocols competing at each level.
  • Embedding congestion control, multipath selection and retransmission in hardware can reduce the response time of software-only handling.
  • A domestic vendor says it combined a PCIe switch with a network adapter and has customized an 800G AI NIC for a large customer; these are company claims.
Open at TMTPost Ask Claude ChatGPT

Chinese researchers find dendrites may help neurons combine rules and signals

# AI SCMP 15h

A mouse study reports evidence that dendrites can integrate learned rules with sensory inputs, offering a possible mechanism for flexible decision-making.

More
  • The work focuses on dendrites, branching structures attached to neurons, rather than treating each neuron as a simple on-off unit.
  • The proposed mechanism could help explain how brain circuits adapt responses to context.
  • The excerpt does not name the lab, journal or experimental scale, so the result needs the original paper for evaluation.
Open at SCMP Ask Claude ChatGPT

Feishu and Doubao contest prizes students who can solve real business problems

# Culture QbitAI 1h

A Beijing final paired student teams with company problems and job offers, while organizers published a talent report drawing on more than 4,300 teams and 108 companies.

Also: New Wisdom

More
  • The contest frames AI talent as the ability to understand a business problem, use tools appropriately and deliver a verifiable result, not simply generate slides.
  • Examples included a production-scheduling assistant and an AI content workflow; the performance claims come from the event organizers and participants.
  • The offer-selection round turns a marketing event into a direct recruiting channel and an experiment in assessing work samples instead of relying only on resumes.
  • The report is sponsored by the companies behind the tools, so its claims should be read as employer positioning.
Open at QbitAI Ask Claude ChatGPT

A Xi-Trump-Putin APEC meeting depends on agreeing an agenda

# Policy SCMP 13h

A possible three-way meeting remains uncertain, with analysts saying the leaders first need to agree what it would cover.

More
  • The article says a trilateral meeting would carry symbolic weight, particularly for Russia, but gives no confirmation that it will occur.
  • The agenda itself is presented as leverage over whether the meeting happens.
  • The story centers on China's leader and its relations with the US and Russia.
Open at SCMP Ask Claude ChatGPT

Chinese policy advocate calls for safer US-China AI talks

# Policy Pekingnology 15h

Henry Huiyao Wang argues that bilateral AI talks should address safety and business risks, contrasting more positive Chinese attitudes toward AI with US skepticism.

More
  • The column cites a survey in which 85 percent of Chinese respondents saw more benefits than drawbacks from AI, versus 38 percent of Americans; the figures are the author's cited source.
  • It argues that agent behavior and AI-enabled cyberattacks make technical safety a shared interest.
  • This is an advocacy column, not a joint government agreement or an official statement of Beijing's position.
Open at Pekingnology Ask Claude ChatGPT

BYD showcases batteries, fast charging and driver assistance in Paris

# Analysis QbitAI 18h

BYD presented European-market vehicles and technology at the Paris auto show, part of a wider effort to build local products and charging infrastructure abroad.

More
  • The company says its second-generation Blade battery and fast-charging system will anchor its European product push.
  • BYD says it plans 3,000 fast-charging stations in Europe by the end of 2027, tying vehicle sales to infrastructure rollout.
  • Its 2027 European launch of advanced city-driving assistance will test whether Chinese software can satisfy local regulatory and road conditions.
  • The item is company-sponsored coverage, so product and market claims should be treated cautiously.
Open at QbitAI Ask Claude ChatGPT

Is 55 percent of China's AI compute really going to short dramas?

# Analysis Initium 5h

A Chinese feature questions a widely circulated estimate that short dramas consume most AI compute; a separate ByteDance response cited in another report disputes the figure.

More
  • The supplied excerpt contains page code and the headline, not the article's underlying calculation or sources.
  • The claim matters because video generation may be becoming a large commercial workload, but the percentage is not established here.
  • The debate itself reflects public interest in what China's rapidly expanding AI compute is being used for.
Open at Initium Ask Claude ChatGPT

China's shopping platforms are putting AI inside their existing businesses

# Analysis TMTPost 45m

Chinese e-commerce platforms are shifting from talk of AI replacing shopping apps to using it for product discovery, logistics, content conversion and merchant operations.

More
  • The article's central observation is that platforms are embedding AI within their existing supply chains and user habits rather than building a standalone shopping chatbot.
  • Taobao's assistant targets purchase decisions, JD.com's use cases focus on inventory and delivery coordination, and Douyin is trying to connect video discovery to transactions.
  • Pinduoduo is presented as applying transaction data to factory demand and production decisions, while Xiaohongshu faces the harder problem of preserving trust in user-generated recommendations.
  • The piece is an industry analysis, not a measured account of adoption or commercial return.
Open at TMTPost Ask Claude ChatGPT

Shenzhen loosens eligibility for talent housing

# Policy Jiemian 20m

Shenzhen's latest housing allocation lowers the minimum social-insurance record to one year and shortens the lookback for property transactions to three years.

More
  • The change widens access for newer residents and some households that previously owned property but have since been without a home.
  • Housing allocation is one local tool for competing for workers amid high urban housing costs.
  • The notice changes eligibility, but the excerpt does not specify how many additional households will qualify.
Open at Jiemian Ask Claude ChatGPT

China's hotel rebranding boom leaves guests facing mixed signals

# Culture TMTPost 29m

A hotel near Chengde station reportedly carried four brands at once, illustrating a wider wave of franchise changes and weak control over physical brand remnants.

More
  • The article describes properties changing franchise affiliations while leaving old signs, room fittings or shared lobbies in place.
  • It cites 1,887 midrange and higher-end hotels changing brands in 2025, though another industry count in the same report gives a different total.
  • The reported shift toward de-branding suggests competition is increasingly about converting existing properties, not only adding new ones.
  • For guests, mismatched booking listings and on-site branding undermine trust in what a hotel affiliation guarantees.
Open at TMTPost Ask Claude ChatGPT

China's office-agent race is becoming a fight for enterprise budgets

# Analysis TMTPost 29m

Market trackers put Tencent's WorkBuddy ahead of rivals, but the article notes that activity rankings do not establish enterprise customer wins.

More
  • The cited trackers measure different things: one combines monthly active users across products and devices, while another counts desktop visits.
  • Tencent's WeCom integration gives WorkBuddy access to work context, while its connectors for rival collaboration tools weaken any simple platform lock-in story.
  • Baidu approaches tasks through both enterprise workflows and employee files; Alibaba is targeting e-commerce merchants with specialized workflows.
  • The commercial test is repeatable team-level work and paid deployments, not consumer usage or raw visits.
Open at TMTPost Ask Claude ChatGPT

How a Beijing mall turned food, fandom and pop-ups into a draw

# Culture TMTPost 16m

Chaoyang Hopson One thrives without luxury anchors by packing food stalls, viral brands, anime merchandise and hundreds of pop-ups into a transit-linked mall.

More
  • The article says the mall topped a Beijing popularity ranking for 19 months and drew more than 41 million visits in 2025; these are reported figures, not independently checked here.
  • Its basement food market and rotating pop-ups turn social-media attention into a high-throughput retail format, rather than relying on prestige labels.
  • The mix of snacks, collectible merchandise and celebrity events points to younger consumers' use of malls as entertainment and social destinations.
  • The account also highlights a sharp contrast between the crowded interior and the less developed surroundings.
Open at TMTPost Ask Claude ChatGPT

TSMC's results point to a long 3nm runway as 2nm ramps

# Analysis TMTPost 41m

Chinese coverage reads TSMC's results as both an aggressive 2nm ramp and continued 3nm expansion, while warning that new capacity is pressuring margins.

More
  • The account says 2nm reached 8 percent of revenue after two quarters, while 3nm still contributed about a third of wafer revenue.
  • It reports a 66.4 percent gross margin, down 1.3 points sequentially as early 2nm costs weigh on results.
  • TSMC reportedly plans additional 3nm capacity even years after launch, reflecting sustained demand and slower progress at competing nodes.
  • The article's characterization of AI investment as more rational is an analyst interpretation, not a conclusion established by the earnings figures.
Open at TMTPost Ask Claude ChatGPT

Beijing office rents may not bottom out until 2028

# Signal 36Kr 15h

A property consultancy says Beijing office rents remain under pressure, with the west stabilizing earlier than the east as new supply and cautious corporate demand reshape the market.

More
  • The cited report puts third-quarter vacancy near 17 percent and average prime rent at 198 yuan per square meter per month.
  • More than one million square meters of new high-quality office space is expected from 2026 to 2030, adding pressure to leasing.
  • The east-west split suggests technology and other demand clusters are unevenly distributed rather than moving with one citywide cycle.
  • These are consultancy forecasts, not a guaranteed turning point.
Open at 36Kr Ask Claude ChatGPT
Source ledger
ChinAI 403 0/0 HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk 200 1/20
Pekingnology 200 1/20
Ginger River Review 200 0/20
Sinocism 200 1/20
QbitAI 200 10/10
36Kr 200 8/30
Solidot 200 12/20
Leiphone 200 13/20
GeekPark 200 2/30
IT Home ERR 0/0 ChunkedEncodingError: Response ended prematurely
SCMP 200 10/50 +5 over cap
Initium 200 3/15
Jiemian 200 8/30 +11 over cap
China Digital Times 200 1/15
The Wire China 200 0/50
Techmeme 200 0/15
Hacker News 200 1/20
V2EX 200 8/50 +38 over cap
Zhidx 200 10/20
InfoQ China 200 8/20 +10 over cap
New Wisdom 200 8/15 +4 over cap
TMTPost 200 8/16 +8 over cap