China Desk

China News, Summarized 13 Nov 2026 34 stories archived day

Today's brief

Huang's fall shifts the chip fight to who rents Nvidia capacity abroad, while Chinese tech media is still covering the boardroom

Nvidia's board acted in a fortnight

Jensen Huang is out. On Thursday Nvidia's board announced his resignation as chief executive, president and director, effective immediately. CFO Colette Kress is interim CEO. Harvey Jones, a former Synopsys chief executive, becomes the first board chair in the company's 33 years. The stock closed down 4.34% after falling as much as 14%.

"Resigned" is the polite verb. GeekPark, a Chinese tech outlet, notes that Bloomberg's headline said removed. It also notes that only 14 days separated the first New York Times report from the announcement. Solidot, a Chinese tech-news site, reports an internal petition to reinstate him. Staff are backing him on LinkedIn and none publicly urge removal. That is sentiment, not process.

Go deeper on this section: ClaudeChatGPT

Beijing's tech press counts guarantees; Washington counts leaks

TMTPost asked who signs the next backstop. The Chinese tech outlet's column cites a September SemiAnalysis estimate. It puts Nvidia's off-balance-sheet commitments near $530 billion, against $91 billion of liabilities on the books. Guarantees alone rose from $3.5 billion to $108.5 billion.

Its argument is about leverage. An interim CEO and a board with its first chair are unlikely to approve another multibillion-dollar guarantee, the column says. New cloud companies and independent labs would feel that first. It calls the trigger a private scandal and never raises the intelligence question.

Washington's leaks went after the evidence. CNN reports US officials disagree on whether Alice Huang, the businesswoman tied to Megaspeed, was recruited by China's Ministry of State Security. Attorney General Todd Blanche says no conclusions have been reached. CNN also reports White House questions over why Commerce did not put Megaspeed on its entity list.

The Times adds the politics. It reports that Trump's and Commerce Secretary Howard Lutnick's closeness to Huang made officials wary of aggressive steps against Nvidia's chip sales.

THE TELL: each side's lead is an American institution failing, Nvidia's board in one and Commerce in the other. Beijing's alleged role is a footnote in both.

Go deeper on this section: ClaudeChatGPT

The real exposure is rented chips

Reuters counted $3.2 billion. Its review of more than 150 Chinese tenders, contracts and disclosures found 23 billion yuan (~$3.2 billion) of computing projects built around restricted H100, H200 and H800 chips. The total includes buildings and power. It doesn't show the chips arrived, and the projects aren't linked to Megaspeed.

Hong Kong priced something else. Bloomberg reported that Alibaba rents Nvidia chips through Megaspeed. Its Hong Kong shares fell as much as 7.3%, and model developers Z.AI and MiniMax each slid at least 8%. Reuters says Chinese AI firms are approaching overseas data-centre operators to reserve capacity.

The fear is a bill. They worry about the Remote Access Security Act, which would give Commerce authority over cloud access to restricted chips. The House passed it 369–22 in January. The Senate hasn't acted.

Onshore, the trade ran the other way. In Thursday's A-share session, 36Kr's digest says, domestic chip designer Enflame rose more than 10%. MetaX, Moore Threads and Cambricon rose more than 8%. That session followed the Wall Street Journal's report that the board was preparing to oust Huang.

My read: Chinese clouds' offshore rentals are the weak seam. None of the Chinese outlet pieces in today's feed mention them.

WHAT WE'RE WATCHING: Alibaba's results on 19 November, the first chance to say how much of its compute runs through Megaspeed, per BigGo Finance.

Go deeper on this section: ClaudeChatGPT

Alibaba's chip unit gets a price, not a divorce

Spin-offs are out-valuing their parents. TMTPost's analysis says Baidu's chip unit Kunlunxin carries a reported $50 billion target, against Baidu's market value of about $38.5 billion at the time. Kuaishou's video unit Kling is valued at $18 billion, against Kuaishou's roughly $16.3 billion. Alibaba's chip unit T-Head filed confidentially on 11 September for a Hong Kong listing.

Alibaba keeps the bill. T-Head stays consolidated, so chip design and research costs still hit Alibaba's accounts. The rumoured $2–3 billion raise would be under a third of Alibaba's 67.7 billion yuan (~$9.5 billion) June-quarter capex. TMTPost's point: only the price leaves the group.

Every chip gets two prices. With outside shareholders, TMTPost argues, Alibaba Cloud wants cheap silicon while T-Head's investors want chip-company margins. Internal transfers become disclosed related-party deals. Capacity splits the same way: each chip sold to ByteDance, Tencent or Baidu is one Alibaba Cloud doesn't get. IDC puts T-Head at 6.6% of first-half data-centre accelerator shipments.

My read: with Nvidia's China supply now a politics story, a clean listed domestic supplier is easier for state buyers to sign.

Go deeper on this section: ClaudeChatGPT

A fined ad and a pianist's note

Shanghai fined an ad for glamorising cheating. Putuo district regulators penalised Shanghai Zongzheng Foods, which runs Dove's flagship store on Kuaishou, 800,000 yuan (~$113,000). A video had a woman say her boyfriend wouldn't buy the chocolate box, so a man she was flirting with drove two hours to deliver it. Per 36Kr, regulators cited the Advertising Law and a wrong view of love and marriage. My read: moral tone in ad copy is now an advertising-law matter.

A pianist's note is spreading. SCMP reports that Bai Baoguo, an associate professor at Yan'an University's Luxun Academy of Fine Arts in Shaanxi, was ordered to secure research funding to keep his job. He died on 20 October. A six-page note alleging malpractice at the institution circulated from 3 November. A student posted it with the family's consent. The allegations are unverified. The part that will travel is the grant target for work that isn't grant-shaped.

Go deeper on this section: ClaudeChatGPT

An Olympic cloud stays behind

Alibaba Cloud is leaving Senegal its Olympic cloud. InfoQ China reports that the private cloud built for the Dakar Youth Olympics, which close Friday, will pass to Senegal's government as a sovereign cloud run by local staff. It carried about 1 million tickets plus e-visas. Alibaba calls it an Olympic first.

The Games paid for the platform. Liu Guohua, who runs Alibaba Cloud's private-cloud unit, says in Alibaba's own account that international work is under 10% of its revenue but nearly half its investment. The piece cites Gartner forecasting $80 billion of global sovereign-cloud spending in 2026, but only about $31 million in sub-Saharan Africa. A fixed-budget event covered the first cheque.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • Nvidia: Huang didn't last to earnings. On 12 November I asked whether he'd hold the job for the 17 November results and the 18 November House briefing. He's gone, and a CNBC headline says he has agreed to testify to Congress. Xi hosts Trump in Shenzhen on 18–19 November. China's commerce ministry spokesperson He Yadong said Thursday the two sides are in close contact on reciprocal tariff cuts covering $30 billion each.
  • Tencent: a second denial. Yesterday I wrote that the rumour outran the correction. A source close to Tencent now tells Jiemian the 50% layoff claim is untrue. At the company's 28th-anniversary livestream, management stressed a flatter middle layer. The "army and navy" split described deployment, not cuts.
  • Anthropic: the $1.4 trillion held. Subscriptions are about 6x the shares on offer, per Jiemian via 36Kr. Anthropic and its co-founders aim to raise up to $64.4 billion, plus $26 billion in a private placement, at about $1.4 trillion fully diluted.
  • T-Head: still no company number. Outside estimates in TMTPost's piece run from $20 billion to $62 billion, a threefold spread.

Go deeper on this section: ClaudeChatGPT

The river · 34

The US and China convened the first meeting of a bilateral Board of Trade days before Trump meets Xi in Shenzhen, establishing a channel for trade engagement amid summit diplomacy.

More
  • The US Trade Representative said its office and Treasury met Chinese counterparts, according to the report.
  • The new forum creates a standing mechanism for trade discussions, though the supplied account does not identify specific agreements or deliverables.
  • Its timing suggests both governments want working-level economic channels active ahead of the leaders’ meeting.
  • A formal board can help separate technical trade disputes from high-level political bargaining, but its practical value depends on follow-through.
  • The report confirms a meeting, not a thaw in the broader trade relationship.

China’s openEuler unveiled an operating-system stack for agent infrastructure, adding sandboxing, KV-cache support, memory and agent security; the project estimates domestic installations above 22 million.

More
  • The release positions the operating system beneath agent applications, where long-running inference, state, tool access and isolation have to be managed alongside conventional workloads.
  • The announced capabilities include sandboxes, KV-cache functions, a memory system and agent-security features; the report says the stack is already used by leading internet and financial-sector customers.
  • The estimated 22 million installations suggest growing reach for an open-source system rooted in Huawei’s server OS, now governed through the OpenAtom Open Source Foundation.

Nvidia says Huang has resigned as president, CEO, and director, with CFO Colette Kress serving as interim CEO while the board searches for a successor.

Also: InfoQ China, TMTPost, 36Kr, Zhidx, Techmeme, Solidot

More
  • The company named veteran director Harvey Jones its first board chair and said it will consider both internal and external CEO candidates.
  • The report describes an immediate share-price drop followed by a partial recovery, showing how investors are pricing leadership uncertainty at the center of the AI supply chain.
  • The announcement does not give a reason for Huang's departure; this item should not be read as confirmation of the surrounding allegations or their role in the decision.

Unlike Chinese coverage focused on petitions and founder loyalty, this US discussion examines whether Huang’s exit could accelerate export controls and debate the merits of remote access to compute.

More
  • The podcast discussion connects Huang’s resignation to possible US legislation and argues over whether restrictions should target chip exports, equipment, or remote access to computing resources.
  • It frames leasing compute to Chinese firms as both a source of revenue and a potential policy lever, while weighing whether to close Malaysia-based access routes.
  • The discussion also considers Nvidia’s execution and succession risks, and how close Chinese chipmakers may be to competing with US silicon.
  • These are the participants’ analyses and judgments, not confirmed policy changes or company plans.

Chinese robotics company Stardust says its CoMind framework lets robots share compressed, task-relevant visual context and coordinate actions without relying on a central controller.

More
  • In a blocked-vision test, the company reports two robots completed a shared pen-retrieval task with success rising from 0% to 100%; a separate hidden-box task improved from 38.9% to 88.9%.
  • The system extracts task-relevant visual features and operational state rather than sending full camera feeds, which the company says reduced communication volume by 80%.
  • Its architecture combines context extraction with bilateral sharing: each robot can both contribute observations and generate actions using its own view, task state and teammate’s context.
  • The demonstrations span box assembly, folding bedding, restocking a refrigerator and assembling a dumbbell, suggesting coordination under occlusion and with deformable objects.
  • These are company-reported demonstrations; the source does not provide independent replication or evidence of performance in commercial deployments.

China’s openEuler project is adding high-density sandboxes, memory systems, KV-cache acceleration and agent security controls, while reporting production use in e-commerce and finance.

More
  • The team says its sandbox can start in under 50 milliseconds and scale to 100,000 instances in 10 seconds, targeting the short-lived, parallel execution patterns of agents.
  • The proposal includes tiered memory, retrieval and rollback, plus traceability and file recovery intended to constrain agent actions and reduce token overhead.
  • JD.com says the system handled about 30% of its advertising traffic during Singles’ Day, with end-to-end recommendation latency down 15%; these are partner-reported production results.
  • The project’s framing is pragmatic: improve relatively stable infrastructure components such as sandboxing, cache and memory rather than declaring an entirely new “agent OS.”
  • openEuler also reports growing agent use in its developer community, but these participation metrics are self-reported and do not establish software quality.

A Zhejiang University benchmark tests whether vision-language models can predict spatial relations after objects move; the best of 21 models scored 58%, versus 87.2% for people, and specialist models were near chance.

More
  • SpaceCast-Bench evaluates a different capability from recognizing a scene: the model must infer an unseen result after a specified object movement, removal or whole-scene rotation.
  • Its 182 indoor scenes produce 3,862 questions across static perception, local change and global transformation, with relations tested from camera-, object- and world-centered frames.
  • The reported best model reached 58.0% accuracy, while human participants scored 87.2%; six spatial-specialist models scored 25.7% to 29.1%, close to the 25.9% random baseline.
  • The reversal suggests that narrow spatial training does not automatically yield causal state updates or coordinate-frame robustness; broad models performed better in this test.
  • For robotics, the limitation matters because an agent must predict consequences before acting, not merely describe the current camera view.
  • The result is benchmark-specific, and the article excerpt does not establish whether the tasks predict performance on physical robots or other spatial datasets.

Chinese company Wujie Frontier has open-sourced Fx-Bio-1031, a locally deployable model for bioinformatics workflows, and reports high scores on difficult scientific benchmarks.

More
  • The model is designed to work in a Linux tool environment, writing code, installing software and querying biological databases rather than answering only from retrieved text.
  • The company reports 70.59% Pass@5 on a difficult BioMysteryBench subset and 89.50% Pass@3 on BiomniBench, comparing favorably with named general models in its evaluation.
  • The use case is technically meaningful because scientific workflows involve large datasets, missing labels and multiple tool calls, while local deployment can address data-governance requirements.
  • The benchmark claims come from the company’s evaluation framework and need independent replication; high pass@k scores do not establish reliability in regulated drug or clinical work.

Chinese coverage presents DeepSeek’s search, file, web-reading, and speech interfaces as building blocks that let developers reproduce much of its consumer app.

More
  • The model can decide when to search, retrieve pages, and incorporate them into an answer, reducing the orchestration developers previously had to implement.
  • Search, fetch, OCR, speech recognition, and speech synthesis are also exposed as separate tools, with tool calls billed apart from model tokens.
  • DeepSeek’s demo and sample code lower the cost of building an app around its services, but also make the model provider more central to the application stack.
  • The announcement advises developers serving external users to enable safety checks and meet applicable content and AI-labeling obligations.

A Chinese analysis argues that wearable recorders have a temporary opening before phone and watch makers absorb always-on transcription and personal-memory features.

More
  • The products increasingly discard audio after transcription, keeping text and summaries instead; that design shifts the value from recording hardware to searchable context.
  • The article says third-party agents cannot freely access phone apps and microphones, creating an opening for dedicated devices as a workaround to platform boundaries.
  • If phone makers expose similar capabilities through controlled interfaces, recorder startups could lose their hardware advantage and compete on memory retrieval and agent integration.
  • The trend raises privacy and consent questions because a device that records continuously can capture other people’s conversations as well as the wearer’s.

A Zhejiang University researcher favors a general multimodal model for planning combined with specialized tools for spatial memory and fast control.

More
  • The proposed split assigns high-level reasoning and route choice to a multimodal agent, while a separate navigation controller handles movement.
  • The researcher argues that remote 30-billion- or 100-billion-parameter models may reason well but face network latency, while physical interaction needs much faster control loops.
  • For spatial memory, the approach uses structured mapping such as SLAM, with a vision-language model interpreting the resulting representation.
  • The design reflects a broader robotics tradeoff: preserve general model capabilities for planning while offloading high-frequency actions to specialized systems.

A student published a late professor's account alleging that a Chinese university tied teaching employment to research funding, raising questions about academic labor and grant incentives.

More
  • Bai Baoguo, an associate professor at Yan'an University's art academy, died on October 20; a six-page note attributed to him circulated online after his death.
  • The note reportedly describes alleged institutional practices, but the supplied article excerpt does not include the full allegations or the university's response.
  • The case touches a broader governance tension in Chinese universities: research funding and measurable output can become employment criteria even for faculty whose work is not grant-intensive.
  • Because the account was published posthumously and has not been independently adjudicated in the supplied report, its claims should remain attributed.
  • The episode has drawn attention because it links workplace pressure, academic evaluation and the public use of a deceased scholar's testimony.

Anthropic reports that its Anvil automated red-team system found severe safety bypasses across biology, cybersecurity and information-security tests, in some cases using far fewer requests than human testers.

More
  • The company says Anvil found more severe jailbreaks than human testers in 23 of 24 biology test combinations and broke through 24 cybersecurity test combinations in 0.5 to 3.5 hours.
  • In one information-security test, Anthropic says Anvil found a reasoning-information leak after 435 requests, compared with roughly 102,200 requests by human experts.
  • The system runs independent Claude tasks in parallel and adjusts prompts in response to the target model, using multi-step reasoning rather than fixed test cases alone.
  • Anthropic also says Anvil’s attack success depends on the underlying model’s capability; its own safeguards blocked most test sessions when enabled.
  • Automated red-teaming could compress pre-release testing, but it also increases the volume and severity of findings that security teams must triage.

Sun Yat-sen University researchers report a 3.6-fold speedup by training a vision-language model to generate useful spatial features internally instead of calling a 3D geometry encoder at inference time.

More
  • The first GeoThinker version selectively retrieved relevant geometry from an external VGGT encoder, improving a spatial benchmark but requiring the encoder on every query.
  • GeoThinker-F uses VGGT features as training supervision for lightweight latent-frame prediction modules, then removes the external encoder at inference.
  • The team reports that spatial-reasoning time fell from 1.45 seconds to 0.40 seconds with performance largely maintained.
  • The system trains the model to predict geometry and temporal features from internal representations while leaving the main backbone frozen, reducing serving-time dependencies.
  • The work matters for robotics and driving applications where external geometry models add latency, though the source does not provide deployment evidence in real vehicles or robots.

The ID.ERA 5X is a test of Volkswagen’s move toward Chinese-led vehicle platforms and electronics architecture, with local development aimed at lower cost and faster iteration.

More
  • The report says the CMP electric platform was developed in China and the CEA electronic architecture was co-developed by Volkswagen China, CARIAD China and Xpeng.
  • CEA reportedly cuts electronic control units by about 30%; the source says its development took 18 months, a sign that the local team is taking on a larger share of vehicle-system definition.
  • The car’s driver-assistance stack uses a Horizon Robotics chip and model through a joint venture, with a “white-box” arrangement that the article says gives the team access to underlying technology for iteration.
  • The broader strategic shift is from China as a market receiving global products to China as an engineering base whose platforms may be exported to other regions.
  • The report is highly promotional and the claimed price, safety and cost advantages require validation in production and on-road performance.

At its display conference, TCL CSOT presented printed-OLED panels, high-refresh gaming displays and AI tools for display design and manufacturing.

More
  • The company says a design agent spans requirements, design, simulation and validation, and reports shorter customer-response and development cycles; these are company-reported figures.
  • A 27-inch printed-OLED gaming panel combines an oxide backplane with a compensation technique intended to support 500 Hz refresh and sub-millisecond response.
  • TCL also described low-refresh-rate power management for laptops and glass-based through-via packaging aimed at high-performance computing workloads.
  • The showcase links display manufacturing to AI in two directions: AI-assisted design and production, and panels designed to conserve power or support AI devices.
  • The technical claims were presented by the company and are not independent evaluations of yield, reliability or commercial deployment.

Westlake University’s 4-billion-parameter Awomo-0.5 aims to represent reusable physical concepts rather than generate future video frame by frame, and reports strong results on robot-manipulation benchmarks.

More
  • The model’s proposed concept-world approach represents object properties, spatial relations, materials and contact states in a composable latent space, then uses them to predict what happens next.
  • The article distinguishes this from video-generation approaches and from JEPA-style latent prediction, describing the concepts as linked to both language and physical phenomena.
  • Westlake reports 80.6% success on a perturbed two-arm manipulation benchmark and 73.4% on a zero-shot benchmark with unfamiliar scenes and objects.
  • Those results provide early evidence for the representation, but remain concentrated in simulated robot tasks and do not establish general physical intelligence.
  • The university’s research center signals that Chinese institutions are organizing long-horizon world-model work alongside startup and municipal initiatives.

A Chinese medical-AI company says its MedGPT system outscored clinicians and general models on complex lung-cancer cases, but the study design and scoring method need careful scrutiny.

Also: QbitAI

More
  • The reported evaluation used 40 unpublished cases and 250 doctors from tertiary hospitals; senior clinicians created 240 scoring rules, including 40 safety red lines.
  • The company reports that MedGPT beat unaided physician submissions in 88% of comparisons and that doctors’ scores improved 24.1% with assistance.
  • Its system reportedly combines structured clinical evidence, multiple specialist agents, alternative treatment paths and review against safety criteria rather than generating a single unqualified recommendation.
  • The comparison was scored by three AI judges, and the source does not provide independent clinical-outcome evidence or enough detail to establish that the evaluation predicts better patient care.
  • These are company-associated results in an important clinical domain, not proof that the model is ready to make treatment decisions without physicians.

A Chinese variety show’s reunion of two former Super Girl contestants has become a viral pairing, bringing 2000s queer fan memories into mainstream online conversation.

More
  • The show reunites Xu Fei and Li Na, contestants from the 2006 season of Super Girl, after years of estrangement and a public confrontation about their relationship.
  • Online audiences have turned their interactions into a widely shared romantic narrative, with the phrase “middle-aged lesbian melodrama” capturing the mix of humor, longing, and queer reading.
  • The writer, a longtime fan, connects today’s viral discussion to an earlier generation’s fan fiction and private identification with women performers.
  • The attention has moved beyond fandom: brands and shopping centers have joined the public campaign urging one contestant to forgive the other, according to the article.
  • The moment shows how Chinese viewers use reality television and online “shipping” to make space for queer interpretations, even when the program itself is presented as a friendship show.

Chinese coverage centers on public anger over damage to a rare Han-era inscription in Mongolia and questions about how a newly rediscovered heritage site was protected.

More
  • Two Chinese nationals were reportedly convicted in Mongolia after making unauthorized impressions of the cliff inscription.
  • The site has faced illegal tampering since it became widely known nearly a decade ago, according to the report and Chinese online discussion.
  • The controversy shows how heritage visibility can attract both public interest and physical damage when visitor access and protection lag behind publicity.
  • The inscription is a Chinese historical artifact located outside China, adding a cross-border dimension to heritage stewardship.
  • The supplied account does not detail the site's current security arrangements or the specific penalties imposed.

The US Treasury secretary says Commerce delayed blacklisting Megaspeed, a customer suspected of diverting advanced AI chips to China, sharpening an interagency export-control dispute.

Also: Techmeme

More
  • The timing, immediately before a Trump-Xi meeting, makes enforcement of chip controls a live diplomatic issue rather than an abstract rulemaking question.
  • The report names Megaspeed as the suspected diversion risk, but does not establish that the company violated export law.
  • The dispute exposes a recurring policy tradeoff: restricting access to protect security while managing the commercial interests of US chip suppliers.
  • Chinese coverage foregrounds the agency blame, whereas the related US item focuses on Nvidia’s pressure campaign in Washington.

Alibaba Cloud’s dedicated infrastructure for the Dakar Youth Olympics is set to remain in Senegal as a locally operated sovereign cloud rather than being dismantled after the event.

More
  • The platform supports ticketing and e-visas across Dakar and Saly and is planned for transfer to the Senegalese government after the Games.
  • The project treats a time-limited sports event as the initial demand and funding anchor for longer-lived digital infrastructure, rather than as a temporary cloud deployment.
  • The account describes two availability zones and months of failover drills, reflecting the reliability challenge of operating through unstable power and network conditions.
  • Alibaba says it will train local operators and support them with remote experts; the long-term test is whether Senegalese teams can sustain the service and attract users beyond the event.
  • The deal illustrates a shift in China’s cloud exports from selling compute toward transferring an operating platform and skills, while leaving questions about long-run cost and vendor dependence.

Trump is expected to press Southeast Asian governments to join Pax Silica, a US effort to reorient AI and chip supply chains away from China.

More
  • The report says AI and critical minerals are expected to feature prominently in Trump's agenda at the ASEAN summit in Manila.
  • Pax Silica frames supply-chain alignment as a regional choice, potentially drawing Southeast Asian economies into tighter US-led technology networks.
  • The article relies on people familiar with the plans; it does not say which governments will join or what commitments membership would entail.
  • For China, the initiative tests whether its trade and manufacturing ties can withstand efforts to build a parallel technology-supply bloc.

SCMP reports that U.S. lawmakers are seeking testimony from Jensen Huang amid allegations involving a suspected Chinese intelligence operation and questions about White House oversight of chip exports.

More
  • The report says several committee leaders requested testimony as concern grows over national-security risks and advanced-chip exports to China.
  • The alleged intelligence link is not established by the article; the concrete development is lawmakers seeking answers and connecting the claim to export-control oversight.
  • The episode shows how personal allegations can become entangled with the policy debate over Nvidia’s access to China and Washington’s management of strategic technology.
  • The excerpt does not give the full letters or the company’s response, so the scope of the inquiry remains uncertain.

A Chinese-language reader-response column gathers debate on dog control in Hong Kong, celebrity politics, a Huawei vehicle test and labor exploitation of Chinese workers in Japan.

More
  • Comments on Hong Kong's dog incidents debate sterilization and adoption policy after officials seized dogs following fatal attacks; the column shows public concern about both safety and animal welfare.
  • The roundup describes mainland online backlash against a Taiwanese celebrity's appearance at a political event, followed by a Chinese automaker removing related promotional content.
  • Other responses challenge the disappearance of a video alleging a Huawei-linked vehicle brake failure and discuss exploitative recruitment of Chinese workers in Japan, emphasizing how trust and accountability recur across unrelated controversies.
  • Because this is a curated collection of reader comments, it is evidence of arguments being made, not independent verification of each underlying claim.

Analysts say Washington and Beijing may agree on preventing AI misuse but disagree on the threshold for a serious incident, complicating efforts to establish a bilateral channel.

More
  • The report describes a proposed mechanism for communicating about AI risks, not a settled agreement or formal incident-response protocol.
  • The central obstacle is definitional: governments may classify the same model behavior or failure differently, limiting what information they would share.
  • The discussion comes as the two sides are trying to maintain narrow cooperation while acknowledging deep strategic differences.
  • A functioning channel would require agreed incident categories, reporting triggers, and protections for sensitive technical information; the report says those questions remain open.

A new national standard specifies safety, accessibility and service requirements for neighborhood facilities, turning elder-friendly design into a more measurable part of China’s community-care system.

More
  • The standard covers facility location, circulation and lighting, service layouts, safety monitoring, and recurring inspections rather than treating accessibility as a single construction feature.
  • It sets concrete requirements such as elevators for services above the ground floor, limits on hazardous stairs, and minimum areas for care and community-service facilities.
  • Self-service terminals must use larger text, offer voice prompts and provide a route to human assistance, an acknowledgment that digital access can itself exclude older residents.
  • The rules support the government’s 15-minute home-based eldercare goal; their effect will depend on local budgets, retrofits and enforcement, not just new projects.

Chinese coverage emphasizes the risk to China’s access to Nvidia computing, while US discussion centers more on the leadership scandal; analysts cited here expect a successor to have less political influence in Washington.

Also: IT Home

More
  • The article argues that Huang had been a prominent advocate for allowing US chip exports to China and had relationships with senior US officials that observers believed could affect policy.
  • It predicts that a successor may have less influence and that controls on advanced chips and overseas compute access could be enforced more tightly; this is an analyst forecast, not a policy announcement.
  • The report says Chinese companies still have unmet demand for advanced computing despite rapid domestic-chip development, and that overseas cloud routes are also exposed to possible licensing restrictions.
  • It describes how Nvidia’s leadership uncertainty can ripple through Chinese firms, suppliers and investors tied to AI infrastructure, but does not quantify a new restriction or supply disruption.
  • The underlying strategic uncertainty is whether Washington’s export policy depends partly on corporate advocacy, or whether the security consensus will drive controls regardless of who leads Nvidia.

A national plan sets a 2030 target of 300 GW of new energy storage and backs several technologies, including sodium-ion batteries, as China shifts the sector from rapid buildout toward market-based operation.

More
  • The plan targets roughly three hours of average storage duration, putting storage capacity and duration together as measures of the system’s intended role.
  • It calls for further lithium-ion improvements and wider sodium-ion deployment, while moving compressed-air, flow-battery and flywheel systems from pilots toward commercial use.
  • Grid-forming storage is a priority: the plan calls for deployment and validation of its reliability and ability to support the grid, with technical standards still to be developed.
  • The document also warns against disorderly expansion by battery makers and emphasizes recycling, quality testing and supply security, pairing industrial growth with capacity discipline.
  • The key test is whether new storage can earn revenue in a market-based system, rather than relying on targets and local industrial investment alone.

Huawei-backed openJiuwen describes an agent runtime where agents can package new capabilities as versioned plugins and install them without changing the host application’s code.

More
  • The extension points span the client, gateway, session server, multi-agent orchestration and individual agent tools, with event handlers registered through a runtime event bus.
  • The design supports version selection and rollback, letting a generated plugin be tested in an application and reverted if it breaks behavior.
  • The architecture is a proposal and product demonstration; the source does not discuss security review of agent-generated extensions or production adoption.

A Chinese analysis says Singles’ Day sales growth slowed to 4.8% from 14.2% the prior year as year-round discounts, promotion fatigue and merchants’ focus on margins eroded the event’s old promise of a rare low-price window.

More
  • The cited third-party estimate puts this year’s festival sales growth at 4.8%, down from 14.2% last year; the article does not provide the underlying methodology.
  • The piece argues that year-round promotions have made it harder for shoppers to believe Singles’ Day offers the lowest price, while elaborate discount rules add friction.
  • Merchants are said to be prioritizing profitability, customer retention and price discipline over headline order volume, with some avoiding costly livestream sales channels.
  • The shift suggests a more cautious consumer mood and a mature e-commerce market where platforms compete for existing spending rather than creating a new buying occasion.
  • Singles’ Day remains useful for product launches and inventory clearance, but the article argues it no longer reliably functions as an industry-wide growth engine.

Kuaishou says its commerce agents balance answer quality, latency and cost using dynamic knowledge stores, user simulation and training methods for sparse conversational rewards.

More
  • Kuaishou's example involves rapidly changing livestream promotions, so the knowledge layer must ingest updates quickly and reconcile potentially conflicting seller pages, spoken content and customer-service records.
  • The company says it uses user simulators and self-evolution strategies to limit model degradation as new situations arise, while keeping inference within a one-to-two-second response target.
  • Because sales or issue resolution is observed only at the end of a conversation, the team describes assigning intermediate reward signals to improve credit assignment across turns.
  • The presentation argues for a purpose-built model trained on business data, citing cost, stability and data sensitivity as reasons not to rely directly on external APIs or generic open models.
  • These are practices described by Kuaishou's own engineering team; the excerpt does not provide independent results or quantified conversion gains.

The companies say Baidu will supply cloud and compute infrastructure while Pudu trains its own embodied foundation model using its deployed service-robot fleet and real-world data.

More
  • Pudu says its robots operate in 85 countries, serve more than 50,000 customers and have shipped over 130,000 units, giving it a potential source of deployment data.
  • The partnership divides responsibilities: Baidu provides its Baige AI compute platform, while Pudu develops PuduFM, its embodied model, and upgrades the PuduAgent platform.
  • The stated objective is to address data scarcity, training cost and transfer across robot bodies, with a “one brain, many forms” architecture.
  • The companies want to move commercial service robots from isolated machine-level intelligence toward a reusable platform layer.
  • The announcement does not disclose data-sharing terms, model results or whether performance will transfer across different robot hardware.

Alibaba plans to list its T-Head semiconductor unit while retaining control, a move that could give the chip business its own market price without removing its costs from the group.

More
  • The article says T-Head would remain a consolidated subsidiary, so research and tape-out spending would continue to flow through Alibaba's accounts.
  • It argues that the main change is a separately quoted valuation, not a transfer of control or a material near-term funding solution; the reported fundraising estimates are speculative.
  • The rationale includes attracting outside customers and talent, especially because Alibaba Cloud competes with other potential buyers of domestic AI accelerators.
  • China's procurement environment may also favor a visibly independent local supplier, while Alibaba Cloud benefits from being able to offer competing chips without appearing to privilege its own unit.
  • The strategic trade-off is between making T-Head easier to finance and staff, and weakening the market's ability to value Alibaba as an integrated AI-cloud stack.