China Desk

China News, Summarized 21 Oct 2026, 12:08 UTC 34 stories 22/23 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

Chinese apps block Doubao’s phone agent, exposing a fight over who controls transactions

# Zhidx 3h

Many Chinese apps have rejected Doubao Phone Assistant’s automated actions, reflecting a clash between agents acting for users and platforms protecting security, rules and their control of traffic and commerce.

More
  • The article says 155 apps opted out of the assistant’s screen automation, including major services in messaging, e-commerce, payments and local services.
  • Platforms rely on user attention, advertising exposure and transaction flows; an agent that jumps directly to a result could bypass those economic layers as well as the apps’ own security controls.
  • The dispute concerns who may authorize an agent to act as a user, what data it can see and which platform terms govern automated access.
  • The article describes similar blocks affecting foreign agents and recounts unresolved litigation over Perplexity’s shopping agent, suggesting the access conflict is not uniquely Chinese.
  • A durable agent ecosystem will likely require explicit identity, permission and liability mechanisms, not just better screen automation.
Open at Zhidx Ask Claude ChatGPT

Hong Kong independent bookstores face searches, arrests and closures

# Initium 13h

Hong Kong bookstores and staff have faced national-security searches, detention and exclusion from the book fair, leaving owners questioning whether they can keep operating and hosting events.

More
  • The report says several independent bookshops were searched in 2026 and multiple owners or staff were detained; some stores have since closed or announced plans to close.
  • Bookshop workers describe preparing for possible arrest while trying to continue ordinary work, illustrating how enforcement pressure extends into cultural spaces and daily routines.
  • The loss of events and the inability to sell or discuss some books threaten more than individual businesses: they weaken venues for public exchange and independent cultural production.
  • Some shops are reopening or relocating, but face practical pressure from leases, staffing and uncertainty about what can be stocked or discussed.
  • The story shows how national-security enforcement can reshape cultural life without a single blanket closure order, through repeated investigations, exclusions and self-censorship.
Open at Initium Ask Claude ChatGPT

China is building embodied-AI training around real industrial workstations

# Leiphone 5h

A national initiative is shifting embodied-AI training from demonstrations toward real production settings, where robots can collect task data, be tested on jobs and return failures to training.

More
  • A government research report counted more than 70 embodied-AI training sites already operating and 46 planned or under construction, but identified weak data throughput, repetitive scenarios and poor reuse as constraints.
  • The central bottleneck is the gap between training and deployment: simulated skills often fail to transfer to a new factory line, shelf layout or operating environment.
  • The initiative launched by industry and state-asset regulators uses real workstations as training and validation settings, tying data collection to tasks defined by users.
  • The article describes a feedback pipeline linking data capture, model training, strategy validation and redeployment, with standardized data and testing intended to reduce the cost each company bears alone.
  • A cited automotive deployment has reportedly completed nearly 80,000 component-handling tasks with 98% operating availability, an example of the shift toward measurable production use rather than a staged demo.
  • The durable test is whether the resulting datasets and evaluation rules transfer across robot bodies and sites, while resolving safety, liability and data-standardization issues.
Open at Leiphone Ask Claude ChatGPT

Taiwan expels mainland exchange student over a social-media post

# Initium 3h

Taiwan revoked a mainland student’s entry permit and imposed a five-year re-entry bar after a TikTok post about President Lai sparked debate over political expression, proportionality and immigration authority.

More
  • The student was studying at National Taiwan University on an exchange from the University of British Columbia and was ordered to leave after posting a video and political slogans about Lai.
  • He removed the post and apologized, saying he had no plan to cause physical harm; Taiwan’s immigration agency said the post included language advocating the elimination of Taiwan’s sovereignty.
  • The case moves from online expression to a severe administrative sanction in five days, prompting questions about the legal threshold and why the maximum five-year bar was used.
  • The controversy shows how cross-strait identity and political speech can quickly become questions of national security and immigration enforcement in Taiwan.
Open at Initium Ask Claude ChatGPT

PingCAP confirms internal breach and source-code theft

# InfoQ China 0m

PingCAP confirms unauthorized access, copied internal data and stolen source code; Chinese coverage stresses that it has not found evidence of impact to China customers’ production databases or the TiDB kernel.

More
  • The company says attackers used stolen service-account credentials to access some internal systems and copy data, including source code; it has not confirmed the much larger volume and credential counts claimed by the attackers.
  • The attacker’s allegations include millions of files, customer records, API keys and SQL histories, but those figures remain unverified and should not be treated as the confirmed breach scope.
  • PingCAP says exposed records included some cloud-user identity and project data, SQL editor text, error messages and commercial information; the article reports conflicting claims over whether API keys were usable.
  • The company says its external responders found no evidence of commands being run or persistent access in reviewed production systems, and no evidence that data stored in TiDB was accessed.
  • The incident illustrates how a cloud database provider’s control plane, support systems and query histories can expose sensitive customer context even without a confirmed intrusion into customer databases.
Open at InfoQ China Ask Claude ChatGPT

DeepSeek’s open inference stack targets both Nvidia and Huawei Ascend

# TMTPost 1h

DeepSeek’s new inference code lists Nvidia GPUs and Huawei Ascend as supported targets, potentially making hardware choice more flexible for Chinese inference operators, though comparable Ascend benchmarks are still scarce.

More
  • Supporting both accelerator families in one codebase could reduce the cost of maintaining separate serving stacks when chip availability, price or policy constraints change.
  • Inference is a more immediate target for this kind of substitution than frontier training: model weights are fixed, and operators can compare latency, utilization, reliability and cost per generated token.
  • The article says public performance tables lean on Nvidia hardware, so support for Ascend is not yet proof of comparable throughput under matched workloads.
  • DeepSeek describes the release as offline inference code, not a complete production service; fleet management, cache coordination and load balancing remain separate engineering problems.
  • The strategic test is whether the shared software path runs reliably under peak traffic and makes domestic accelerators operationally interchangeable, rather than merely buildable.
Open at TMTPost Ask Claude ChatGPT

US lawmakers eye DUV lithography controls as leverage over China’s AI compute

# Policy ChinaTalk 1d

A ChinaTalk discussion argues that restricting exports and servicing of immersion lithography could slow China’s chipmaking, while noting that China has already imported large numbers of DUV systems.

More
  • The case for tighter controls is that lithography is difficult to indigenize and is a leverage point in the semiconductor manufacturing stack.
  • The discussion cites estimates of roughly 90 immersion systems imported annually in 2024 and 2025, suggesting current controls may leave a significant gap.
  • The proposed MATCH Act would restrict DUV immersion exports and servicing, but effectiveness depends on coordinating with allies and on what China can repair, substitute or build domestically.
  • The argument also weighs retaliation risk, including whether gradual restrictions are less likely to provoke a sharp response over rare earths.
  • This is a policy argument, not an enacted restriction; its impact would depend on allied participation, China’s installed base and the pace of domestic equipment development.
Open at ChinaTalk Ask Claude ChatGPT

Li Auto unveils battery, powertrain and charging-system upgrades

# Analysis QbitAI 11h

Li Auto presented an integrated electric-technology system spanning cells, silicon-carbide power electronics and charging, alongside lifetime battery-fire guarantees and performance claims that remain company-reported.

More
  • The company says its new fast-charging cells improve energy density while supporting 5C charging, and that software controls can improve battery power and range in cold weather.
  • Its battery-management approach combines cell-level manufacturing records, vehicle and charger telemetry and cloud risk alerts, linking battery safety to quality control across the product lifecycle.
  • Li Auto also reported improvements to its electric drive, including a direct power-module connection design and reduced component count, but the article gives no independent validation.
  • The lifetime guarantee is a commercial signal: an automaker is willing to make battery safety part of the product promise, though the detailed claim conditions matter.
  • The broader strategy is vertical integration across cells, power electronics and charging, where manufacturing yields and real-world durability will determine whether the performance claims become an advantage.
Open at QbitAI Ask Claude ChatGPT

MemT 1.0 adds a trainable memory pathway inside the Transformer

# AI QbitAI 23h

The MemTensor team’s MemT 1.0 models encode experience in an internal, updateable state and retrieve it through a parallel memory branch, rather than relying only on external retrieval.

More
  • The architecture adds a memory computation path alongside attention, training the model to write, update and read experience as part of its own computation.
  • The reported versions range from a 36-billion-parameter model with 3 billion active parameters to a 288-billion model with 17 billion active parameters.
  • On the reported software-engineering tests, the smaller model modestly improves average success over both a no-memory baseline and a text-memory setup, with a larger gain in repeated-run reliability.
  • The tradeoff is that internal state may be harder to inspect, edit or delete than explicit text memory; the report acknowledges the need for mechanisms to support those operations.
  • The result is an engineering demonstration of memory-native scaling, not proof that internal memory will broadly outperform external retrieval in real deployments.
Open at QbitAI Ask Claude ChatGPT

MiniMax positions M3.1 Flash as a low-cost agent model

# AI QbitAI 22h

MiniMax says its M3.1 Flash model combines fast output with low token prices and strong scores on coding, computer-use and multi-step agent tests.

Also: Leiphone

More
  • The article reports top domestic performance on an agent benchmark and a tie with a leading model on a finance tool-use benchmark, but these are benchmark results rather than broad production evidence.
  • The model is described as using sparse attention and producing output at more than 200 tokens per second, with low listed prices for input, output and cached reads.
  • The coverage frames the model as complementing DeepSeek: DeepSeek is portrayed as stronger on coding value, while MiniMax emphasizes cross-service workflows.
  • The launch shows Chinese model makers competing on both capability and serving economics, a combination that could matter more for agent workloads with many tool calls.
Open at QbitAI Ask Claude ChatGPT

Tencent and Ant back former ByteDance product chief’s personal AI startup

# AI Zhidx 8h

Today AI, founded by former Doubao PC lead Junyuan Qi, says it has raised a cumulative $200 million to build a cross-device assistant that records context and can act proactively.

More
  • The product connects to work and personal services to retain information such as meetings, documents and schedules, positioning itself around persistent memory rather than one-off chat.
  • Investors reportedly include Tencent and Ant, an unusual combination of major ecosystem players backing a startup that could become a personal-agent layer across their services.
  • Its proposition depends on users granting continuous access to private context and on the assistant producing enough useful actions to justify that level of access.
  • The funding figure is company-reported, and the article provides no independent usage or retention data.
Open at Zhidx Ask Claude ChatGPT

Shanghai pilots digital IDs and governance rules for AI agents

# Policy Leiphone 5h

A Shanghai platform is issuing identifiers for AI agents as China develops rules for agent identity, permissions and accountability when software acts across systems.

More
  • The article says an agent identity code is intended to make each agent identifiable and searchable; identity alone does not establish what it may do or who is responsible for its actions.
  • National guidance cited in the report calls for agent registration and capability declarations, while governance plans stress differentiated permissions and human decision rights.
  • The organizational problem is assigning work, reviewing outputs and tracing responsibility when several agents call tools and systems in sequence.
  • The report describes human review and operational logs as necessary controls; without a named accountable owner, a chain of automated actions can leave liability unclear.
Open at Leiphone Ask Claude ChatGPT

Chinese buyers rush to reserve Nvidia RTX PRO 5500 cards

# Signal Leiphone 1d

Industry sources cited by Leiphone say Chinese buyers have ordered millions of RTX PRO 5500 workstation GPUs, attracted by large memory and multi-instance support, but supply and performance remain uncertain.

Also: Solidot

More
  • The report says the card offers 84 GB of memory and can be divided into two isolated compute instances, making it potentially useful for inference and other workloads that benefit from large memory.
  • Interviewees estimate Chinese orders exceed three million units, while the report says Nvidia’s quarterly China supply may be far lower; order numbers and delivery timing are market-source claims.
  • Buyers are reportedly targeting the card as a less expensive path to high-memory compute, with expected delivered prices around 80,000 to 120,000 yuan (roughly $11,000 to $17,000).
  • The article notes that theoretical compute advantages over an earlier China-compliant card do not prove real-world performance; workloads and software tuning determine the actual gap.
  • The rush exposes a strategic vulnerability in China’s compute buildout: even a card that can be purchased is not useful capacity until it is delivered, tested, integrated and supported.
Open at Leiphone Ask Claude ChatGPT

China-watchers track proposed export controls and the coming Fifth Plenum

# Policy Sinocism 3h

A Sinocism episode surveys proposed Chinese export controls on AI, solar and machinery, rare-earth tensions with Japan, economic questions and expectations for the Fifth Plenum.

More
  • The roundup presents the controls as part of a wider effort to manage strategic technologies and supply-chain leverage, rather than as a single narrow AI measure.
  • It also notes questions about growth targets and stock-market weakness despite strong reported third-quarter GDP, and flags speculation about the next Party plenum.
  • Because the supplied text is show notes, it lists topics but does not establish the proposals’ precise scope or final policy status.
Open at Sinocism Ask Claude ChatGPT

China’s token economy links AI usage, compute and model efficiency

# Signal Leiphone 5h

A Chinese industry analysis says domestic model calls exceeded 500 trillion tokens per day by June, while arguing that token counts are useful for billing but do not capture task quality or total cost.

More
  • The article describes token usage as a common accounting interface for model services, but warns that tokenization, caching, modality and output quality differ across providers.
  • Chinese examples include time-of-day pricing by DeepSeek and a local carrier’s prepaid model credits, showing how providers are testing consumer and enterprise billing formats.
  • The piece argues that compute buyers should compare useful throughput, latency, availability, energy and task cost, not just hardware counts or price per million tokens.
  • It describes hardware-software co-optimization as a way to shorten domestic accelerator adaptation from months to days in some cases, though the figures are attributed to companies.
  • For embodied systems, token volume is only one part of economics: sensing, inference time, device power, action success and equipment cost determine the cost of completing a physical task.
Open at Leiphone Ask Claude ChatGPT

Chery lays out an unusually broad plan for chips, models and physical AI

# AI GeekPark 2h

Chery unveiled plans spanning automotive chips, an industrial and vehicle model stack, autonomous driving, robotaxis and electric farm machinery, though most performance and deployment figures are company claims.

More
  • The proposed chip family covers central compute, silicon-carbide power devices and control chips; Chery says its planned C100 central processor could deliver 1,200 TOPS per chip and target deployment in 2027.
  • Its CheryMind model is presented as an internal enterprise system as well as a base for vehicle, aviation and agricultural applications, with the company claiming use across hundreds of internal systems.
  • For autonomous driving, Chery describes a world-model stack that generates simulated scenes and separates scene construction, perception and decision-making; the article does not provide independent validation of those capabilities.
  • The company also presented electric tractors and other products beyond passenger cars, suggesting a bid to reuse its software, power-electronics and autonomy capabilities across adjacent machinery.
  • The test is execution: chip production, safety validation, real customer use and recurring economics matter more than the breadth of the announced roadmap.
Open at GeekPark Ask Claude ChatGPT

Beijing’s Taiwan affairs chief reaffirms red lines in rare meeting with US scholars

# Policy SCMP 1h

Chinese coverage presents a rare meeting between Taiwan Affairs Office director Song Tao and a US foreign-policy delegation as both an opening for peaceful dialogue and a reaffirmation of Beijing’s position on Taiwan.

More
  • The meeting took place in Beijing with a delegation from the National Committee on American Foreign Policy led by its president, Susan Elliott.
  • Song situated the exchange in the context of recent high-level US-China diplomacy, while reiterating Beijing’s political red lines on cross-strait relations.
  • The combination of dialogue and red-line messaging reflects Beijing’s effort to distinguish engagement with US interlocutors from any retreat on its Taiwan position.
Open at SCMP Ask Claude ChatGPT

Taiwan says it funded a retired US general to observe major war games

# Policy SCMP 10h

Taiwan’s defense ministry confirmed funding for US Indo-Pacific Command personnel, including a retired general, to observe its largest annual military exercises, offering rare public detail on security ties.

More
  • The ministry’s disclosure sheds light on exchanges that Taipei and Washington have generally kept out of public view.
  • The report places the arrangement amid growing security cooperation and pressure from Beijing, making it a politically sensitive signal as well as a budget disclosure.
  • The supplied excerpt does not specify the amount or the observer’s operational role, so it cannot establish how much the arrangement changes military cooperation.
Open at SCMP Ask Claude ChatGPT

Sinocism tracks Li Qiang’s diplomacy, a new data group and the Fifth Plenum

# Policy Sinocism 13h

A China-focused newsletter roundup covers Premier Li Qiang’s regional diplomacy, the National Data Group, Fifth Plenum expectations and rare-earth tensions with Japan.

More
  • The newsletter’s listed subjects span government, economic policy and foreign relations, but the supplied excerpt gives few details beyond the agenda.
  • The National Data Group’s inauguration and the approaching Fifth Plenum are potentially important governance signals, but the item does not explain their mandates or policy implications.
  • The roundup also flags rare-earth pressure on Japan, an area where trade and strategic leverage intersect.
Open at Sinocism Ask Claude ChatGPT

Chinese researchers build a one-gram wearable monitor for brain activity

# AI SCMP 8h

Researchers from Tsinghua University and a military medical school report a wireless one-gram device that can track thousands of neurons across a broad area of the cortex in untethered laboratory animals.

More
  • The device, called Airscope, is described as enabling wider neural monitoring without tethering the animal, potentially allowing researchers to observe activity during more natural behavior.
  • The report calls it the smallest wearable brain monitor of its kind, but the supplied excerpt does not provide recording resolution, battery life or independent replication.
  • If validated, the engineering advance could make longitudinal neuroscience experiments less constrained by bulky or wired equipment.
Open at SCMP Ask Claude ChatGPT

China refits an older destroyer with domestic air-defense systems

# Policy SCMP 2h

Photos circulating online appear to show the Type 051C destroyer Shenyang at sea trials after a refit that may have replaced Russian hardware with domestic systems; details remain unconfirmed.

More
  • The report places the refit amid a broader PLA Navy transition as newer ships enter service and older vessels approach retirement.
  • If the apparent replacement is confirmed, it would show an effort to extend the utility of legacy hulls while reducing reliance on imported combat systems.
  • The available evidence is social-media imagery, so the ship’s precise equipment and operational status are not independently established.
Open at SCMP Ask Claude ChatGPT

Chinese companies recruit students who can use AI to deliver business results

# Culture QbitAI 9h

A Feishu-Doubao contest put university teams on real company problems, with employers judging whether students could use AI to produce verifiable work rather than polished presentations.

Also: New Wisdom

More
  • The organizers say 4,300 teams and 108 companies contributed to a report on AI-era talent, but the event was also a branded recruiting exercise.
  • Examples included production scheduling, supply-chain analysis and radar incident workflows, emphasizing domain understanding and human oversight alongside tool use.
  • The contest’s job-offer stage let students choose among employers, presenting recruitment as a two-way evaluation rather than résumé screening alone.
Open at QbitAI Ask Claude ChatGPT

OpenAI extends API prompt-cache eligibility to 12 hours

# AI Zhidx 5h

OpenAI is adding a 12-hour prompt-cache option for some API models, while Chinese coverage compares it with DeepSeek’s longer retention target and notes that cache eligibility does not guarantee a cache hit.

More
  • The article says the longer API option costs more for cache writes, while reads keep their existing price; it does not establish that the stored key-value state will always remain available.
  • Cache billing eligibility, actual cache reuse and request latency are separate properties, so a longer window alone does not prove lower latency or higher hit rates.
  • The comparison with DeepSeek highlights how inference providers compete on both infrastructure behavior and the price of repeated context.
Open at Zhidx Ask Claude ChatGPT

Chinese roundup captures consumer frugality and disputes over AI-generated viral claims

# Culture Leiphone 1d

A Chinese news roundup covers a viral app that asks strangers to approve purchases, a disputed claim about AI compute use in short dramas and allegations of AI-fabricated videos targeting Sam’s Club.

Also: Techmeme

More
  • The purchase-approval mini-program turns consumption anxiety into a social game, with users asking strangers whether small discretionary purchases are worth making.
  • A ByteDance representative disputes a widely circulated claim that more than half of domestic AI compute goes to short dramas, saying the company uses about 10% of its compute for video-model training and inference.
  • Sam’s Club said viral complaint videos about product removals were fabricated or misleading, illustrating how synthetic media can become part of commercial reputation fights.
  • The roundup also repeats the US DeepSeek restriction story; its Chinese coverage emphasizes the federal procurement ban and Beijing’s rejection of the allegations.
Open at Leiphone Ask Claude ChatGPT

openEuler promotes China’s Linux distribution at a European open-source summit

# Analysis Solidot 54m

China’s openEuler community drew conspicuous attention at Open Source Summit Europe, showcasing a distribution that supports long-term releases and several domestic processor architectures.

More
  • The project originated from Huawei’s EulerOS but is now hosted by the OpenAtom Foundation, whose founding members include major Chinese technology firms.
  • The distribution supports RISC-V and LoongArch, with older releases also supporting the Shenwei architecture; this ties the OS effort to China’s broader push for domestic computing stacks.
  • The summit presence suggests Chinese developers are treating international open-source events as a venue to build recognition and adoption, not just domestic software substitution.
Open at Solidot Ask Claude ChatGPT

Chinese startup livestreams the training process for a physical-AI model

# AI Leiphone 2h

Shengshu Technology has begun a two-week public livestream of Motubrain 2.0 training, aiming to show how an embodied model is trained, tested and iteratively improved before open-sourcing.

Also: New Wisdom

More
  • The planned stages include pretraining, task-specific post-training, agent task execution and experiments with recursive self-improvement based on feedback.
  • The team says the project is intended to expose failures and iteration, not just polished robot demonstrations; actual capabilities still require independent evaluation.
  • Public training visibility could help researchers compare methods in a field where datasets, deployment conditions and failure modes are often difficult to inspect.
Open at Leiphone Ask Claude ChatGPT

China’s delivery-robot race is shifting from replacing riders to line-haul routes

# Analysis TMTPost 2h

A report says Meituan has struggled to make autonomous vehicles worthwhile in complex last-mile delivery, while New Stone’s nearly $400 million funding round shows investors still backing the sector.

More
  • The article describes a mismatch between autonomous vehicles and the final few hundred meters: building access, elevators, customer coordination and handoffs remain difficult to automate.
  • Meituan reportedly found that human-and-robot delivery could be slower than human-only service in dense, complicated neighborhoods, prompting a focus on more standardized route segments.
  • The stronger use case may be moving goods between depots and delivery zones, leaving riders to handle the flexible doorstep work where local knowledge matters.
  • That division of labor points to augmentation rather than simple rider replacement, with unit economics depending on route density, vehicle utilization and the cost of human handoffs.
Open at TMTPost Ask Claude ChatGPT

Chinese policy advocate urges safety-focused US-China AI talks

# Analysis Pekingnology 23h

Henry Huiyao Wang argues that the planned US-China AI dialogue should address both technical safety and commercial risks, including agent behavior and AI-enabled cyberattacks.

More
  • The column supports formal channels for cooperation, citing risks from poorly controlled agents and the spread of cyber capabilities to less specialized actors.
  • Its claims about US and Chinese public attitudes and specific cyber incidents are presented as argument in an opinion column, not independent reporting.
  • The piece reflects a Chinese policy advocate’s case for dialogue rather than an official Chinese government position.
Open at Pekingnology Ask Claude ChatGPT

Starbucks China requires customers to buy before taking a seat

# Culture TMTPost 2h

Starbucks China is ending general free seating, a shift that lands amid competition from cheaper local chains and the transfer of majority control to Boyu Capital.

More
  • The article traces Starbucks’ open-door seating policy to a 2018 US discrimination crisis, when the company made access for non-buyers part of its public image.
  • In China, the company’s premium cafe model faces pressure from low-priced domestic rivals and the cost of maintaining large, comfortable stores as customers become more price-conscious.
  • Boyu’s majority stake may sharpen the focus on store economics, but the article notes that Starbucks had already reversed free-seating rules in North America before the China change.
  • The change illustrates how a branded social space can be treated first as a corporate value and later as a monetizable operating asset when market conditions shift.
Open at TMTPost Ask Claude ChatGPT

Chinese mathematician Sun Zhiwei offers more than 1,000 problems to AI researchers

# Culture SCMP 8m

Nanjing University mathematician Sun Zhiwei has built a collection of more than 1,000 number-theory problems, arguing that asking worthwhile questions matters as much as producing answers in the AI era.

More
  • The collection is the product of a long research career and is intended to give mathematicians worldwide problems to investigate, rather than simply another benchmark score.
  • The story casts mathematical curiosity and problem formulation as enduring human contributions as AI systems become better at generating solutions.
Open at SCMP Ask Claude ChatGPT

China’s AI-actor business tests how much audiences will trust synthetic stars

# Culture TMTPost 2h

Chinese production companies and platforms are building AI performers and dramas, but early ad deals and audience reactions expose limits around likeness rights and fabricated personal experience.

More
  • Some production teams buy low-cost likeness rights to generate AI performers, while established actors have disputed claims that they authorized inclusion in a platform’s AI roster.
  • The article describes one AI character gaining followers through everyday social posts and attracting high ad rates, but a contact-lens endorsement drew criticism because the synthetic persona cannot have the claimed physical experience.
  • Platforms and studios are also trying to own character design, distribution and commercialization rather than depend on human performers’ schedules or revenue splits.
  • The business model depends on audience willingness to suspend disbelief; deceptive endorsements or unclear consent could quickly undermine that trust.
Open at TMTPost Ask Claude ChatGPT

Chinese aging-care advocates argue AI cannot replace human attention

# Culture 36Kr 7h

A Beijing discussion on aging and AI paired calls for assistive technology with warnings that older adults still need human companionship, digital support and control over their care.

More
  • Participants described older adults learning AI tools and using home sensors, while pointing to scams, confusion and the digital divide among people with less support.
  • One speaker argued that technology can assist with monitoring and daily tasks but cannot substitute for the time and relationships that define care quality.
  • The discussion also emphasized older people’s wish to contribute and be recognized, not only to receive services, a point often missed by product-centered aging-tech plans.
  • The event included project figures from the organizers, but its central value is the social argument about dignity and human support rather than a new device.
Open at 36Kr Ask Claude ChatGPT

Fingerprints suggest women and children helped build the Terracotta Army

# Culture SCMP 5h

A new study of fingerprints on Terracotta Army figures suggests that some components were made largely by adult women and minors, challenging assumptions that the labor force was exclusively male.

More
  • The research adds evidence about who performed large-scale production work in ancient China, extending archaeological interpretation beyond the finished artifacts.
  • The supplied account does not identify the study’s methods or the size of the fingerprint sample, so the extent of the conclusion is unclear.
Open at SCMP Ask Claude ChatGPT

Alibaba Cloud holds 45% of China’s automotive public-cloud market, IDC says

# Signal Leiphone 2h

IDC puts China’s automotive public-cloud infrastructure market at 8.34 billion yuan in the first half of 2026 and gives Alibaba Cloud a 45.1% share.

More
  • The report says Alibaba Cloud’s automotive-cloud revenue grew 111% year over year and that it has led this market for 11 consecutive IDC tracking periods.
  • The figures indicate concentrated cloud infrastructure supply to Chinese automakers, though the article’s explanation of the company’s advantages is promotional.
Open at Leiphone 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 4/10
36Kr 200 6/30
Solidot 200 10/20
Leiphone 200 14/20
GeekPark 200 3/30
IT Home 200 12/60 +38 over cap
SCMP 200 10/50 +8 over cap
Initium 200 4/15
Jiemian 200 8/30 +9 over cap
China Digital Times 200 0/15
The Wire China 200 0/50
Techmeme 200 3/15
Hacker News 200 0/20
V2EX 200 8/50 +38 over cap
Zhidx 200 10/20 +2 over cap
InfoQ China 200 8/20 +12 over cap
New Wisdom 200 8/15 +4 over cap
TMTPost 200 8/14 +5 over cap