China Desk

China News, Summarized 21 Oct 2026, 08:08 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

Hong Kong independent bookshops weigh survival after national-security searches

# Initium 9h

After searches and detentions involving several independent bookshops, owners describe fewer public events, commercial strain and uncertainty over what can safely be sold or hosted.

More
  • The report names multiple shops searched or whose staff were detained in 2026, and notes that some stores have closed or announced plans to close.
  • One shop owner describes preparing for possible police detention as part of ordinary work, a stark measure of how security enforcement has entered cultural life.
  • The book trade is affected not only through explicit censorship but through event cancellations, publisher choices, staffing and the pressure to self-censor.
  • The article follows shopkeepers' daily decisions, showing how enforcement can reshape a cultural sector without a single public rule listing every prohibited title.
  • The broader consequence is a shrinking space for independent cultural exchange in Hong Kong.
Open at Initium Ask Claude ChatGPT

DeepSeek opens parts of its production inference stack

# InfoQ China 0m

DeepSeek released an inference framework already used for large-scale service, describing split prefill and decode, multi-vendor hardware support and long-lived KV cache; some tightly coupled components remain closed.

Also: Zhidx, Hacker News

More
  • DeepInfer supports distributed inference, speculative decoding and structured output, and is reported to run on Nvidia and Huawei Ascend hardware.
  • DeepSeek says a single cluster has handled 155 trillion tokens per day; that is a company-reported service figure, not an independent benchmark.
  • The architecture separates prefill and decode resources, letting each side be scaled and optimized for different latency and compute profiles.
  • The reported design target keeps KV cache for more than 72 hours to reduce repeat prefill work; the detailed report describes different retention periods by workload.
  • The release is not a complete infrastructure dump: DeepSeek says KV-cache management and load-balancing code remain coupled to internal systems and are not fully open.
  • Cross-vendor support is strategically notable in China's constrained compute environment, though the source does not establish equal performance across accelerators.
Open at InfoQ China Ask Claude ChatGPT

PingCAP confirms internal breach as attacker claims a massive data haul

# InfoQ China 0m

PingCAP confirmed unauthorized access, copied data and source-code theft, while disputing or not yet verifying the attacker's much larger claims; its China unit says customer production databases were not affected.

More
  • The company says stolen service-account credentials enabled access to some internal systems between September and October; the full customer impact remains under investigation.
  • The attacker claimed 10.6 TB, including API keys and customer records, but those figures are unverified and should not be treated as confirmed exposure.
  • PingCAP said exported records included user and organization details, SQL editor text, errors and some commercial information, while China customer production databases were not found affected.
  • This illustrates why a SaaS control plane, query history and support systems can expose sensitive context even without direct access to customer database contents.
  • The company said identified credentials were invalidated and reported vulnerabilities were fixed; the key open question is whether other credentials or data remain exposed.
Open at InfoQ China Ask Claude ChatGPT

Rural workers are being hired to collect data for embodied AI

# TMTPost 15m

A Chinese report follows rural women and small-town contractors earning modest daily wages to record first-person work data for robot training, exposing the human labor behind embodied-AI datasets.

More
  • The report describes workers wearing data-capture equipment while doing routine tasks such as sorting and bagging peanuts; one cited rate is 80 yuan (about $11) for an eight-hour day.
  • Unlike robot teleoperation, wearable capture can gather human demonstrations without supplying an expensive robot body for every recording session.
  • The work is being brokered through layers of contractors and crowdsourcing platforms, with lower-cost rural areas becoming attractive as projects scale.
  • A 500,000-hour commercial demand was discussed by one contractor but had not been signed, so it should not be read as realized volume.
  • This is an early view of how AI data supply chains can reproduce familiar labor subcontracting and margin compression.
Open at TMTPost Ask Claude ChatGPT

Chinese shopping platforms are moving product discovery into AI chat

# TMTPost 22m

For China's Singles' Day shopping season, Douyin and Alibaba are steering shoppers toward conversational product advice and merchant agents, testing whether chat can replace parts of the search-and-recommendation funnel.

More
  • The report describes Douyin putting shopping prompts and coupons into Doubao and other apps, while Alibaba routes Taobao users to its Qianwen assistant.
  • The consumer examples suggest chat is most useful when shoppers are undecided or comparing options; repeat purchases still bypass AI advice.
  • Alibaba's assistant can draw on shopping history, while Douyin links conversational discovery more directly to its commerce ecosystem.
  • The strategic contest is over the interface and transaction data: a general assistant may become a new gateway, but closed platforms retain an advantage in inventory, personalization and checkout.
  • The article notes that an overseas attempt by OpenAI at in-chat checkout was discontinued, underscoring that recommendation does not automatically solve fulfillment and merchant economics.
Open at TMTPost Ask Claude ChatGPT

Li Auto unveils its in-house battery system and lifetime fire coverage

# Analysis QbitAI 7h

Li Auto presented an integrated battery, drive, silicon-carbide and charging system after six years of development, pairing technical claims with a lifetime guarantee for qualifying fire losses.

More
  • The company says its new 5C lithium-iron-phosphate cell has higher energy density than its 2C cell, while winter peak power improves by more than 40%.
  • Li Auto attributes improved state-of-charge estimates to an adaptive algorithm that uses charging history and voltage correction, with claimed error below 3%.
  • The package combines cell chemistry, pack structure, thermal management, power electronics and charging infrastructure, illustrating the integration burden behind vehicle performance.
  • Its guarantee covers qualifying battery defects and collision-related fires, a consumer trust strategy as automakers take more responsibility for battery risk.
  • The technical metrics and safety tests are company-reported; long-term fleet data will determine whether the warranty is a durable differentiator.
Open at QbitAI Ask Claude ChatGPT

A proposed DUV export clamp could shape China's future AI compute

# Policy ChinaTalk 21h

A ChinaTalk discussion argues that restricting deep-ultraviolet immersion lithography exports could slow China's chipmaking, while acknowledging allied coordination and Beijing's potential retaliation as key uncertainties.

More
  • The guest argues lithography is a difficult part of the semiconductor stack for China to indigenize and a potential point of leverage for the US and its allies.
  • The discussion says China imported roughly 90 immersion systems annually in 2024-25, nearly two-thirds of ASML's immersion exports; that estimate is central to the argument.
  • The proposed MATCH Act would restrict DUV immersion exports and servicing, but the analysis depends on whether allies adopt comparable controls.
  • A gradual restriction is presented as less likely to provoke a sudden rare-earth response than a sharp cutoff, but that is a strategic hypothesis, not a guarantee.
  • The piece frames export controls as buying time in the AI compute race, while leaving the long-term effects on Chinese substitution and global supply chains open.
Open at ChinaTalk Ask Claude ChatGPT

MemT 1.0 puts learned memory inside the model's computation path

# AI QbitAI 19h

A Chinese research team released memory-native models that encode and update experience as internal state, reporting modest gains in repeated software-agent tasks over baselines using no experience or text memories.

More
  • MemT adds a memory computation path alongside attention in selected Transformer modules, allowing learned state to influence later decisions.
  • The largest model is described as 288 billion total parameters with 17 billion active; a smaller version has 36 billion total and 3 billion active.
  • On 500 SWE-bench Verified tasks run three times each, the smaller model reportedly improved the share of tasks succeeding in all runs from 54% to 62.8%.
  • Internalized memory may make procedural experience easier to reuse, but it is harder to inspect, delete or recover than external text records.
  • The results are early and benchmark-bound; they do not establish that learned memory reliably persists or updates across production workloads.
Open at QbitAI Ask Claude ChatGPT

A CATL factory tour shows why battery manufacturing is hard to replicate

# Analysis QbitAI 3h

A factory visit pushes back on claims that automakers can quickly move away from CATL, showing how cell consistency, process control and production equipment create capabilities beyond cell chemistry.

More
  • CATL says some lines produce a cell per second and apply more than 7,000 quality-control points, including AI vision inspection; the figures are company-provided.
  • The report emphasizes consistency across cells in a pack: variation in capacity or resistance can constrain usable pack performance and worsen with aging.
  • Cell makers also develop manufacturing equipment and process know-how, not just electrochemistry, making a complete supply chain difficult to reproduce.
  • The article frames automaker in-house battery efforts and supplier partnerships as complementary strategies, rather than evidence of a simple supplier exodus.
  • CATL says it has more than 50,000 employees in Ningde and over 230,000 worldwide, illustrating the industrial scale behind the engineering.
Open at QbitAI Ask Claude ChatGPT

MiniMax identifies Space Bunny as its M3.1 Flash model

# AI QbitAI 18h

MiniMax confirmed that the anonymously tested Space Bunny model is M3.1 Flash, positioning it as a low-cost Chinese contender for coding, computer use and complex agent tasks.

Also: Leiphone

More
  • The article reports an Agents' Last Exam score of 29.6% and an Artificial Analysis score of 43, but these are source-reported benchmark results.
  • It says output can exceed 200 tokens per second and lists prices of 1 yuan per million input tokens and 4 yuan per million output tokens.
  • The report highlights different strengths from DeepSeek: coding performance versus multi-step workflow and tool-use results.
  • Its sparse-attention design is presented as a reason for speed, but the article does not disclose enough architecture detail to reproduce the comparison.
  • The practical question is reliability under long agent runs, where tool errors, retries and unsafe actions can matter more than leaderboard rank.
Open at QbitAI Ask Claude ChatGPT

Tencent and Ant back a personal AI assistant startup founded by a former ByteDance executive

# AI Zhidx 4h

Today AI says it has raised $200 million, with Tencent and Ant among investors, to build an assistant that remembers users' work and connects to Chinese apps; the product's long-term value depends on trusted memory and useful actions.

More
  • The company launched its domestic assistant in September and says it runs on iOS, Android, desktop operating systems and Linux.
  • The product can connect to WeChat, WeCom, Tencent Meeting, Baidu Netdisk and other services to organize meetings, documents and personal tasks.
  • The funding syndicate spans major Chinese platforms and venture firms, giving the startup capital but also raising the challenge of cross-platform access and durable user trust.
  • A persistent assistant needs controls over what it remembers, how users inspect or delete memory, and when it acts; the article does not describe those safeguards.
  • The report says the company was founded in 2026 and had raised $200 million cumulatively; independent confirmation is not included.
Open at Zhidx Ask Claude ChatGPT

TokenFleet targets inference delays by routing around queues and cache costs

# AI QbitAI 1d

Chinese inference vendor Shishi launched a distributed scheduler that jointly weighs KV-cache reuse, queueing and cross-datacenter capacity, reporting lower first-token latency in company tests.

More
  • TokenPilot selects instances by balancing cache reuse against expected waiting time, rather than routing only by request count.
  • FleetMind considers whether another datacenter has the right model version, capacity and permission to handle a request before moving it across regions.
  • SLOMax separates interactive, long-context and batch workloads so long prefill jobs do not crowd out latency-sensitive requests.
  • The company reports a 32.4% drop in P95 first-token latency on a GLM 5.3 cluster test; this is a vendor result under specified conditions, not a general cost reduction.
  • Its goodput framing is useful: count successful output that meets service targets, not just GPU utilization or raw token throughput.
Open at QbitAI Ask Claude ChatGPT

Taiwan confirms funding for a retired US general to observe war games

# Policy SCMP 6h

Taiwan's defense ministry disclosed funding for US personnel, including a retired general, to observe its largest annual exercises, offering rare public detail on security cooperation amid pressure from Beijing.

More
  • The ministry's report reportedly confirms a budget allocation for US Indo-Pacific Command personnel to observe the drills.
  • Making the funding public is notable because Taipei and Washington have generally kept military exchanges out of public view.
  • The move signals deeper practical coordination, though the supplied excerpt does not give the amount or describe the observer's role.
Open at SCMP Ask Claude ChatGPT

Chinese team builds a one-gram wearable neural monitor

# Signal SCMP 4h

Researchers from Tsinghua University and a Chinese military medical school describe a wireless one-gram device that can track thousands of neurons in freely moving lab animals.

More
  • The Airscope device is presented as lighter than existing wearable neural monitors and intended to capture activity across a broader cortical area.
  • The research could improve experiments that require natural movement, but the supplied excerpt gives no performance comparison or human application.
Open at SCMP Ask Claude ChatGPT

iQIYI has spent years embedding AI across film and television production

# Culture Leiphone 1d

A Chinese report traces iQIYI's shift from AI experiments to tools for script analysis, production backgrounds and personalized viewing, showing how a media company is reorganizing work around generative systems.

More
  • The platform's 'skip-watch' feature uses viewing behavior and models to identify plot points, offering a shorter route through long-form shows.
  • Production teams used generative images and video to create virtual backgrounds, while researchers and editors labeled scripts to teach models narrative structure.
  • The company trained staff broadly and encouraged experimentation, treating workforce adoption as part of the technical transition rather than a separate rollout.
  • The report says management rejected easy consumer AI wrappers when they lacked a defensible capability, preferring deeper integration in its content pipeline.
  • The broader industry question is whether AI lowers production costs without flattening creative roles or making viewing more fragmented.
Open at Leiphone Ask Claude ChatGPT

Chinese buyers reportedly chase Nvidia's high-memory workstation GPU

# Analysis Leiphone 21h

Chinese buyers are said to have ordered millions of RTX PRO 5500 cards before deliveries begin, driven by large memory, expected compliance and demand for AI inference and embodied-AI workloads.

Also: Solidot

More
  • The report says the card offers 84 GB of memory and supports MIG partitioning, potentially allowing separate inference jobs to share one GPU.
  • Expected China pricing is reported at 80,000-120,000 yuan (about $11,000-$17,000) per card, with delivery timing still uncertain.
  • A source claims domestic orders exceed three million while Nvidia's China supply ceiling is 500,000 per quarter; these market estimates are not confirmed company disclosures.
  • The rush illustrates how hardware access and memory capacity can matter as much as peak compute for deployment-oriented AI workloads.
  • The card's real performance and supply eligibility remain unproven, so orders should not be confused with installed compute.
Open at Leiphone Ask Claude ChatGPT

Chinese startup pitches AI creation tools instead of test-prep for children

# Culture Leiphone 5h

CapaLeap is positioning its children's AI products around guided exploration and making projects rather than answer delivery, a counterpoint to China's crowded test-prep device market and growing worries about dependence on chatbots.

More
  • The company's founder says its agent should help children think and create while leaving choices to the child, rather than acting as a tutor or emotional companion.
  • Its proposed architecture combines a shared agent platform, memory across devices and age-specific hardware, from cameras to a learning and creation device.
  • The report cites surveys indicating widespread student AI use and parental concern about reduced independent thinking, but the figures come from separate surveys and are not evaluated here.
  • The commercial challenge is whether open-ended creative use can support repeat engagement and learning outcomes without turning the device into another screen-time product.
  • The article is a startup profile, not evidence that the product's educational approach works.
Open at Leiphone Ask Claude ChatGPT

Chinese aging-care advocates argue technology cannot replace human contact

# Culture 36Kr 3h

A Beijing forum on older adults and AI highlighted both growing interest in new tools and the persistence of the digital divide, with speakers arguing that devices can support but not replace care relationships.

More
  • An organizer said older residents in community programs are learning AI tools, while sensors are being used to monitor some people living alone.
  • Participants raised concerns about misinformation, dependence and older adults being excluded by digital interfaces, alongside potential gains in safety and convenience.
  • A sociologist argued that many older people want not only services but also opportunities to contribute and be recognized.
  • The report includes program figures from a real-estate company's charity arm, so they should be read as organizer-reported reach rather than independent impact evaluation.
  • The wider issue is China's demographic transition: technology can extend monitoring, but staffing, patience and social connection remain scarce inputs.
Open at 36Kr Ask Claude ChatGPT

QCraft says its urban driver-assistance system is reaching mid-priced cars

# Analysis QbitAI 1h

QCraft touts nine vehicle launches for its single-chip urban navigation-assistance system, an example of Chinese suppliers pushing advanced driver assistance into the 100,000-200,000 yuan segment.

More
  • The vendor says its system runs on Horizon Journey 6M hardware rated at 128 TOPS and uses software-hardware tuning to target stronger performance.
  • QCraft describes defensive driving behavior at unlit intersections and narrow mixed-traffic roads; the claims are from a company-promotional article.
  • The report cites more than one million vehicles equipped and over 4.5 billion kilometers of assisted-driving use, but offers no independent safety comparison.
  • The strategic question is whether a lower-cost compute footprint can deliver reliable urban performance outside carefully selected demonstrations.
Open at QbitAI Ask Claude ChatGPT

Chinese policy adviser urges safety rules for US-China AI talks

# Policy Pekingnology 19h

A Chinese globalization think-tank founder argues that upcoming US-China AI talks should address safety in both technology and business, while warning about investment risks and autonomous-agent misuse.

More
  • The opinion piece says both presidents agreed to establish a formal AI dialogue and hotline, with another meeting expected in November; these are the author's account of recent diplomacy.
  • It contrasts public attitudes, citing a survey in which more Chinese than US respondents saw net benefits from AI, but the excerpt does not give survey methodology.
  • The author uses examples of agent and cyber incidents to argue that AI safety is not just about model behavior but also deployment and corporate exposure.
  • This is a Chinese policy voice advocating engagement, not an official government position.
Open at Pekingnology Ask Claude ChatGPT

Ether Tech raises more than 100 million yuan for a spatial-intelligence engine

# AI 36Kr 5h

Ether Tech raised more than 100 million yuan (about $14 million) to build infrastructure that converts changing physical spaces into reusable, machine-readable representations for robots and other systems.

More
  • The proposed stack aligns visual, location, inertial and force-touch data, then represents objects, physical constraints and interactions rather than just static geometry.
  • The company describes five layers from sensor input and state estimation to reusable spatial assets, platform APIs and industry applications.
  • It is prioritizing human hand interactions, arguing that grasping, assembly and tool use encode actionable spatial knowledge for embodied AI.
  • A shared asset standard could reduce the cost of reusing spatial data across robots and industrial settings, but the company has not demonstrated a widely adopted standard.
  • Investors include Huawei's Hubble fund and other industrial and financial firms; the financing is a bet on infrastructure, not proof of product-market fit.
Open at 36Kr Ask Claude ChatGPT

Aito plans a European launch with Bosch supplying driver-assistance systems

# Analysis 36Kr 22h

Aito's European models are reportedly being developed with Bosch's full intelligent-vehicle stack, a sign that Chinese premium brands may use established foreign suppliers to meet local rules and expectations.

More
  • The reported package spans advanced driver assistance, cockpit software and chassis systems, with highway navigation assistance targeted for production around September 2027.
  • The article says Huawei has left overseas-model leadership to automakers, while Bosch offers regulatory experience and brand recognition.
  • That makes the arrangement a different export model from relying on Huawei-branded systems throughout the vehicle.
  • The report says Aito plans to enter Europe in autumn 2027, though the details are attributed to industry sources and the automaker disputed some organizational claims.
  • Success will depend on compliance, service networks and brand trust, not just exporting a domestic vehicle platform.
Open at 36Kr Ask Claude ChatGPT

China policy watch: National Data Group, Fifth Plenum and rare-earth pressure

# Policy Sinocism 9h

Sinocism's China roundup flags the launch of a National Data Group, preparations for a Fifth Plenum and rare-earth pressure on Japan, but the supplied excerpt gives only a brief contents list.

More
  • The newsletter says Premier Li Qiang attended a Shanghai Cooperation Organization meeting before traveling to Russia for a regular prime-ministerial meeting.
  • The National Data Group and Fifth Plenum items could matter for state data governance and policy priorities, but the excerpt does not explain either development.
  • Because this is a China-focused roundup rather than a single reported event, the underlying stories need separate verification and context.
Open at Sinocism Ask Claude ChatGPT

China plans a cautious expansion of connected-car road trials

# Policy Jiemian 26m

China's industry ministry says it will expand connected-car demonstrations while proceeding cautiously on vehicle approvals and road access, pairing deployment goals with stronger safety and data rules.

More
  • The ministry cited cross-sector coordination, vehicle safety and system reliability as conditions for scaling intelligent connected vehicles.
  • Plans include continued vehicle-road-cloud pilots and upgrades to connected infrastructure, linking cars with transport, energy and city systems.
  • The notice also points to standards for automated parking, operational monitoring and data security, signaling a governance layer that will shape deployment.
  • The policy message is pro-scale but not a blanket opening: approvals and public-road trials remain explicitly cautious.
Open at Jiemian Ask Claude ChatGPT

China tech roundup mixes a DeepSeek ban with consumer and AI controversies

# Culture Leiphone 21h

A Chinese roundup pairs a reported US federal ban on DeepSeek with viral debates over a crowdsourced spending app, AI-made supermarket rumors and claims about short-drama compute use.

Also: GeekPark

More
  • The DeepSeek item claims US civil agencies were told to remove its products and avoid future procurement, but the roundup gives little legal or implementation detail.
  • A small Chinese app asking strangers to approve purchases reportedly went viral, turning inflation-conscious restraint into a participatory online joke.
  • The roundup says ByteDance disputed a claim that over half of domestic AI compute goes to short dramas, saying about 10% of its own compute serves video generation.
  • Sam's Club said viral videos alleging product removals were fabricated with AI-generated material, illustrating how generative media can contaminate consumer trust.
  • The bundle is useful as a snapshot of online mood and policy anxiety, but its items require separate verification.
Open at Leiphone Ask Claude ChatGPT

Virtual power-plant startup raises several hundred million yuan

# Analysis 36Kr 1d

Hefei-based Yuanli Energy raised several hundred million yuan (roughly $30-60 million) to expand its energy software and trading business, as Chinese power-market reforms make flexible loads and storage more valuable.

More
  • The company says its software aggregates solar, storage, charging and industrial loads, then supports trading, demand response and asset operations.
  • It cites a 2.5 MW/5 MWh customer storage project whose reported return improved from negative to about 4% after its charging strategy was adjusted to spot prices.
  • Yuanli says it has aggregated more than 10 GW of resources, but only about 2 GW actively participates in market transactions, a more relevant measure of monetization.
  • Provincial rules and market-access requirements vary, limiting how easily one operating model can be copied nationwide.
  • The company is not yet profitable and expects to reach profitability in 2027; the financing is a bet on electricity-market reform, not proof of sustainable returns.
Open at 36Kr Ask Claude ChatGPT

GoGoX agrees to sell its mainland China business for a nominal price

# Analysis TMTPost 16m

Hong Kong-listed GoGoX plans to sell its loss-making mainland operating unit for one yuan (about 14 cents) and focus on Hong Kong and overseas markets, a stark retreat for a once-prominent local freight platform.

More
  • The unit generated 440.16 million yuan (about $62 million) in revenue and lost 492.59 million yuan (about $69 million) in the first half of 2026, according to the article.
  • Its net liabilities were reported at about 2.48 million yuan (roughly $350,000); the token price is presented as a legal consideration rather than an estimate of business value.
  • The company grew from the merger of mainland 58 Express and Hong Kong-based GOGOVAN, then lost ground as competition and subsidy-led customer acquisition took a toll.
  • The divestiture highlights how platform scale and brand recognition did not translate into durable economics in China's intensely competitive local-delivery market.
  • Completion still depends on shareholder approval and other conditions.
Open at TMTPost Ask Claude ChatGPT

ManageEngine brings more than 30 AI agents to its China products

# AI Zhidx 3h

ManageEngine launched locally hosted IT-operations agents for Chinese customers, with integrations for domestic databases and operating systems; its claims of zero token cost and productivity gains are vendor-reported.

More
  • The company says agents are embedded in its existing IT-management products and run on its own Zia model, rather than charging customers separately per token.
  • The China offering supports Kingbase, Kylin, openEuler and Arm environments and can also connect to models such as DeepSeek.
  • ManageEngine says it plans to invest 425 million yuan (about $60 million) in China over five years, including data centers, AI research and support.
  • The pitch addresses a local enterprise concern: using AI without sending sensitive operational data to an external model provider.
  • Claims about reducing service-desk workload and incident-resolution time are not independently validated in the article.
Open at Zhidx Ask Claude ChatGPT

Chinese companies judge students on AI work, not just tool fluency

# Culture QbitAI 5h

At a Feishu-Doubao contest, students used AI tools on company problems and presented deliverables to employers, reflecting a push to assess applied judgment and execution rather than prompt-writing alone.

Also: New Wisdom

More
  • The event brought 12 teams to a final round and paired them with business cases in areas such as manufacturing, content production and supply-chain operations.
  • Organizers said the competition drew more than 4,300 teams and included talent views from 108 companies; those are organizer-reported figures.
  • Several projects combined AI with existing workflows, such as routing radar detections into human decisions or turning operational data into production schedules.
  • The format doubles as a recruiting channel: companies offered jobs, while students could choose among employers.
  • The event was sponsored by Feishu and Doubao, so its claims about AI productivity and talent demand are not independent labor-market evidence.
Open at QbitAI Ask Claude ChatGPT

Australian Open expands its China play beyond broadcast rights

# Culture 36Kr 1m

The Australian Open says China brings a large, young audience and is adding local qualifying for its fan tournament, turning tennis fandom into a broader consumer and lifestyle proposition.

More
  • The tournament representative cited 6.5 million Chinese social followers and 126.5 million viewing hours for the 2026 event; these are figures supplied in an interview, not audited here.
  • Its one-point tournament will hold qualifying events in five Chinese cities, with a China winner advancing to the Melbourne final event.
  • The strategy reflects a shift from selling match broadcasts toward platform-specific content, merchandise, celebrity participation and live fan experiences.
  • The article presents China as the only international market outside Australia hosting this tournament format, a sign of the market's perceived commercial importance.
Open at 36Kr Ask Claude ChatGPT

China's Huiyuan juice group breaks with the brand it helped build

# Culture TMTPost 12m

A long-running ownership fight has split Huiyuan's brand rights from parts of its production chain, with the former parent announcing a new label as the trademark holder disputes its account.

More
  • The group says it will stop making juice under the Huiyuan name and use a new brand; Beijing Huiyuan says it alone owns the trademarks.
  • The dispute follows a restructuring that separated brand ownership from upstream orchards and production assets, leaving each side with leverage over a different part of the business.
  • The article reports a first-instance trademark ruling ordering affiliated manufacturers to stop unauthorized use and pay more than 97 million yuan (roughly $14 million); the group may appeal.
  • The market share figure cited for the first nine months of 2026 was 4.57%, illustrating the commercial cost of prolonged control disputes.
  • The case shows how restructuring can fragment a consumer brand's operating system across IP, manufacturing and distribution.
Open at TMTPost Ask Claude ChatGPT

Fingerprints suggest women and children helped build the Terracotta Army

# Culture SCMP 1h

A new study of fingerprints on Terracotta Army figures challenges the assumption that the workforce was exclusively adult men, offering a more granular view of ancient Chinese labor.

More
  • The report says fingerprint evidence points to adult women and minors among the craftspeople who made some figures.
  • The finding adds evidence about labor organization in the Qin era, though the excerpt does not give the study's sample size or methods.
Open at SCMP Ask Claude ChatGPT

iFlytek's AI bet is still consuming cash despite better collections

# Analysis TMTPost 14m

iFlytek's third-quarter revenue and reported profit rose, but it remained loss-making on an adjusted basis as heavy model investment continued; cash collection improved while returns on AI spending remain unproven.

More
  • The company reported third-quarter revenue of 6.564 billion yuan (about $924 million) and net profit of 226 million yuan (about $32 million), while adjusted net loss was 126 million yuan (about $18 million).
  • The article says government grants and investment gains accounted for much of reported profit; research spending exceeded 1.7 billion yuan (about $240 million) for the quarter.
  • Management is pulling back from customized government projects and prioritizing education and healthcare, which the outlet frames as a shift toward better-quality revenue.
  • Operating cash flow was 264 million yuan (about $37 million) for the first nine months, with collections improving; receivables remain large.
  • The strategic test is whether the domestic-compute-trained Spark model can translate high R&D spending into recurring API, enterprise and device revenue.
Open at TMTPost Ask Claude ChatGPT

European panel calls 100% technological sovereignty unrealistic

# Analysis InfoQ China 0m

A European open-source panel argued that sovereignty means retaining credible alternatives and the skills to switch, not excluding global vendors; it contrasted Europe's software focus with China's attention to chips and optics.

More
  • The panel framed portability and operational capability as more useful tests of sovereignty than the location of a supplier or access to source code alone.
  • Its brief comparison says Chinese sovereignty debates put more weight on access to chips and optical systems, highlighting hardware dependencies that open software cannot remove.
Open at InfoQ China 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 9/10
36Kr 200 9/30
Solidot 200 12/20
Leiphone 200 10/20
GeekPark 200 2/30
IT Home ERR 0/0 ChunkedEncodingError: Response ended prematurely
SCMP 200 10/50 +6 over cap
Initium 200 3/15
Jiemian 200 8/30 +19 over cap
China Digital Times 200 0/15
The Wire China 200 0/50
Techmeme 200 2/15
Hacker News 200 1/20
V2EX 200 8/50 +37 over cap
Zhidx 200 10/20
InfoQ China 200 8/20 +12 over cap
New Wisdom 200 8/15 +3 over cap
TMTPost 200 8/18 +10 over cap