China Desk

China News, Summarized 29 Aug 2026, 19:22 UTC 34 stories 22/23 sources

Today's brief

The Pentagon's China list is now being answered in court, not by lobbyists

CXMT takes Section 1260H to a Washington judge

CXMT sued the Pentagon on Friday. ChangXin Memory Technologies — the DRAM maker whose first-half revenue of 150.3bn yuan (~$21bn) made it, by Bloomberg's count, China's most valuable listed business — filed in the US District Court for the District of Columbia to get off the Section 1260H list of firms the Defense Department says support China's military. Its chips are civilian, it says. Reuters has the filing.

The procedural history is the argument. Designated in January 2025, CXMT says it spent over a year submitting information; the Pentagon published a notice in February 2026 saying it would be removed, withdrew that notice the same day, then relisted the company in June without explaining the reversal. The complaint calls the decisions "arbitrary." Defense declines to comment on live litigation.

This is now a standard play. Xiaomi sued and got itself delisted in 2021; YMTC filed in December 2025, WuXi AppTec and Alibaba in June. Chinese champions have decided the US Administrative Procedure Act is a usable instrument — and CXMT chose to file three and a half weeks before Xi is expected in Washington on 24 September.

Then Friday's follow button paid off. I noted that CXMT's new Weibo account followed Xiaomi's phone division first. Saturday, the two confirmed that CXMT's LPDDR6 will debut in the Xiaomi 18 Fold next month alongside Xiaomi's in-house 3nm Xring O3. Domestic memory, domestic silicon, one flagship, one lawsuit.

And memory has passed the processor. A hardware leaker relayed by GeekPark puts a 12+256GB memory package at roughly 2,200 yuan (~$310) to procure against about 500 yuan (~$70) a year ago, while the newest flagship SoC went from ~1,500 to 2,000 yuan (~$282). Trade chatter, but the ordering is the point: memory is now the largest line in a phone's bill of materials.

Go deeper on this section: ClaudeChatGPT

Xi lands in Bishkek before he lands in Washington

Xi meets Putin on Monday. The Kremlin says Power of Siberia 2 will come up on the sidelines of the Shanghai Cooperation Organisation summit in the Kyrgyz capital — their second meeting this year. Foreign-policy aide Yury Ushakov said the issue "will be discussed bilaterally," per SCMP. The 1,600-mile line would move up to 50bn cubic metres a year via Mongolia.

The tell is who keeps announcing it. A memorandum exists; a price never has. Moscow puts the pipeline on the public agenda, Beijing declines to say what it will pay, and Chinese readouts of these meetings lead on multipolarity rather than gas. That pattern survived Putin's Beijing visit in May intact.

Treasury kept firing anyway. The US sanctioned a Hong Kong company Friday for allegedly laundering funds for a designated Iranian exchange house, the latest of several Chinese entities hit in days, with Beijing promising retaliation, SCMP reports. Nobody is protecting the atmosphere before 24 September.

Quieter statecraft: AI is now written into the family-doctor contract. A health-commission notice known as document 273 pulls hospital specialists into primary-care contracting and widens the disease list to include infectious disease by end-2027, then tells providers to "deeply advance" AI in those services — analysed in TMTPost. Follow-up volume rises; grassroots headcount does not.

The vendors are already sized for it. iFlytek's medical assistant covers 828 counties and over 77,000 primary institutions, claims above 80% share where AI triage is deployed, and grew that revenue 47.5% in the half. In Tianjin, WeDoctor reports one health manager covering 2,600 patients, about 3.5 times conventional throughput.

Go deeper on this section: ClaudeChatGPT

Alibaba went sparse, Meitu went to the bank

Qwen3.8-Flash shipped Wednesday night, open-weight as Qwen3.8-Flash-Next: a multimodal mixture-of-experts model with 125B main parameters plus 51B of N-gram embedding, only 6B active per token, at 0.8 yuan (~$0.11) per million input tokens and 2.7 (~$0.38) out.

The attention stack is the interesting part. Three of every four layers use gated delta networks compressing history into fixed-size state; the fourth keeps global attention for exact retrieval, with a compressed indexer selecting micro-blocks rather than re-reading context. Alibaba claims up to 7.6x prefill and 4.9x decode kernel speedups at 1M tokens, and in its Qwen office agent, 75% less token consumption per task — InfoQ has both.

Meitu is the one actually banking AI money. The 18-year-old photo-editing company reported first-half revenue of 2.21bn yuan (~$311m), up 22.1%, adjusted net profit 650m yuan (~$92m), up 39.5%, on 18.44m paying subscribers. Its productivity apps carry ARR of about 620m yuan (~$87m), and compute-credit spending roughly doubled across the half.

Set that against ByteDance's Doubao. A Jiemian commentary puts Doubao's daily revenue under 1m yuan (~$141,000) against 200m-plus daily users, and cites trackers showing 18.65m monthly users lost from May to July after a paid tier arrived. Selling a workflow tax beats selling a platform, for now.

Go deeper on this section: ClaudeChatGPT

A robot made 55-step ice cream in Shanghai

Sharpa put a two-armed robot behind a DQ counter. At the chain's Wujiang Road store in Shanghai it takes a metal holder from a steriliser, pulls a paper cup, draws soft serve, adds three scoops of Oreo, blends, and inverts the cup without it falling — the upside-down test the brand has used since 1940. TMTPost watched the cycle.

The specifics make it a product, not a demo. 55 steps, about 90% success, touch feedback in 98% of them, and zero modification to a store that imports every fixture. Chief executive Li Yifan, also a co-founder of lidar maker Hesai, calls purpose-built robot environments "phone booths" and says doing 54 of 55 steps creates a job rather than filling one. Sharpa has raised over 4.5bn yuan (~$634m) at above 20bn yuan (~$2.8bn).

Now read the Beijing robot games two ways. Solidot records Tiangong, from the Beijing Humanoid Robot Innovation Centre, at 8.64 seconds over 100 metres against Bolt's 9.58 — then spends the piece on why it isn't sprinting: nearly a second to react against Bolt's 0.146, more than 50 steps against his 41, no deceleration, and a stop achieved by hitting a crash mat.

AgiBot's version is a medal table. New Wisdom relays 18 golds and 46 medals, including seven of eight dexterous-hand events and six of twelve scenario events — library shelving, hotel rooms, fire response — on production units. Western wires led on the record and used the crash mats as colour; CBS mentioned the stretchers. Correction: I gave Tiangong 8.86 on 25 August. The scenario events are the real scoreboard.

Go deeper on this section: ClaudeChatGPT

A programmer died in a company toilet

Shenzhen refused to call it a workplace injury. A 39-year-old programmer parked in his office building's second basement at 8:56am on 23 April, asked a guard where the second-floor toilet was at 9:00, went in a minute later, and was found unconscious there at 11:28. He was declared dead at 12:50. Solidot has the file.

The stated reason is the story. The city's human-resources bureau ruled on 3 July that his desk was on the 11th floor, that he had clocked in and gone to the second-floor toilet without going to it, and that he was doing nothing work-related. Recognition is worth 1,130,040 yuan (~$159,000) plus funeral and dependants' payments. The family is appealing. Two and a half hours passed before anyone looked.

A second argument is running between two cities. In July, Hong Kong police searched two independent bookstores on suspicion of sedition and arrested five people; one clerk was led away in a T-shirt reading "I am a bookstore clerk." Initium's podcast traces the slogan to Taiwanese bookstores, concert stages and a post by Taiwan's president — then to a fight over whether that was solidarity or consumption.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • MiniMax: the ARR figure I quoted twice is the softest number in the report. TMTPost's read of the interim: recognised first-half revenue was $116.6m against the $800m annualised claim, which Huxiu suspects is one week times 52 and unaudited. Gross margin came in at 17.9% against Morgan Stanley's 29.5%, because enterprise volume was bought with price cuts — M3 doubled API prices on 1 June, the stock fell 15.71% that day, permanent half-price followed within a week. Adjusted net loss $293m, up 111.2%.
  • Gyirong passed 72 hours; Yunnan started flooding. Excavators are reopening the route to the Gyirong border checkpoint with the barrier lake still a breach risk, and CCTV named the officer, Huang Peng, 34, filmed directing people away minutes before the mudslide. Meanwhile Mohan, on the Laos border, took 187.5mm (7.4 inches) of rain in 24 hours; 2,226 people moved, provincial response at level three, 20,500 relief items plus 10,000 from central stores.
  • The discipline machine went back to county work. A day after a legislative committee stripped the army's two most senior officers of their seats, the same apparatus expelled a former vice mayor of Tongliao in Inner Mongolia for, among other things, falsifying shantytown-redevelopment figures — a political-performance charge, not merely a bribery one — plus a former deputy chair of Ulanqab's legislature, with a Chongqing Airport Group executive newly under investigation.
  • Biren's first post-IPO half is 20x revenue and 23% cash. Revenue 1.236bn yuan (~$174m), up 1,997.6% and more than all of 2025; loss narrowed 76% to 380m yuan (~$54m); gross margin 42.7%, down from 53.8% last year as cluster deals drag in third-party hardware. Receivables 1.117bn yuan (~$157m), collected at 23.3%; R&D exceeds gross profit. AIX notes its HK$101.7bn (~$13bn) cap is a third of A-share peers with the same losses.

Go deeper on this section: ClaudeChatGPT

The river · 34

Inference-stack differences can derail long-context agents

# QbitAI 6h

Tests on Qwen3.6-27B found that attention kernels, KV-cache precision and tensor parallelism can change token choices and break tool calls despite identical weights.

More
  • The reported experiment captured more than 100,000 tokens of logits on an RTX PRO 6000 and held hardware, weights and prompt constant while changing inference settings.
  • Different attention backends produced floating-point drift that eventually flipped greedy top-1 tokens; a cited tool call changed a router interface identifier and sent later actions off course.
  • BF16 KV cache stayed stable in the test, while INT4 drift rose sharply with long context; INT8 sometimes recovered after an error.
  • The comparison also found a community W8A16 weight quantization more consistent than the tested official FP8 and NVFP4 paths, partly because it preserved BF16 activations and excluded sensitive projections.
  • The reported NVFP4 path was not native FP4 in that vLLM build, which makes kernel and runtime details part of the result rather than an incidental implementation footnote.
  • The practical lesson is architectural: agent reliability depends on the whole numerical execution stack, not merely model weights or headline quantization format.
  • The findings are from one long workflow and hardware/software configuration, so they need broader reproduction before becoming a general ranking of quantization methods.
Open at QbitAI Ask Claude ChatGPT

A Shenzhen court case tests the boundary of workplace injury

# Solidot 9h

A labor authority rejected a fatal injury claim because a programmer went to a different-floor restroom before reaching his desk; his family is appealing.

More
  • The supplied case record says the 39-year-old arrived at the office building, asked where the second-floor restroom was and was later found unconscious there; doctors pronounced him dead after emergency treatment.
  • The authority reportedly reasoned that he had not reached his workstation or begun work-related activity, while a lawyer argues that a workplace restroom is a reasonable extension of the job site.
  • The family says a successful claim could provide about 1.13 million yuan, roughly $159,000, plus funeral and survivor benefits.
  • The dispute exposes how formally defined work locations can lag behind the practical boundaries of employment, especially in large office complexes.
  • The result could matter beyond this case if courts clarify whether clock-in status, physical premises and work activity should be weighted separately.
  • The case is an individual dispute, not evidence that all Chinese workplace-injury claims are handled this way; the appeal remains unresolved.
Open at Solidot Ask Claude ChatGPT

Unitree's IPO turns China's robot boom into a public-market test

# TMTPost 14h

Unitree's market value fell from a brief 4.45 trillion yuan peak to about 2.44 trillion yuan, roughly $344 billion, exposing the gap between robot excitement and industrial demand.

More
  • The company opened at 1,100 yuan per share after a 150.80-yuan offering price, then lost about 45% from its intraday peak within four trading days.
  • The listing gave China's humanoid-robot sector a public valuation benchmark: Unitree reached a reported 219 times earnings and 35.9 times sales at the offering-era peak.
  • Unitree reported 1.699 billion yuan, roughly $239 million, in 2025 revenue and 278 million yuan, about $39 million, in net profit, but 73.6% of its 2025 first-three-quarter humanoid revenue came from research and education.
  • Industrial applications supplied only 9.01% in that period, so the central commercial question is whether a profitable education-hardware business can become a general-purpose industrial and household robotics platform.
  • The company says 2026 first-half revenue grew 48.5% while adjusted profit fell 19.3%, suggesting that scaling production is already pressuring margins.
  • The listing may mark a shift in Chinese robotics financing from prototype theater to shipment, yield, cost and repeat-order scrutiny.
  • Unitree now has to turn a highly visible public valuation into evidence of reliable deployments, not merely faster demonstrations.
Open at TMTPost Ask Claude ChatGPT

Biren's revenue jumps nearly 20-fold as Chinese GPU demand materializes

# TMTPost 11h

Chinese AI-chip maker Biren reported 1.236 billion yuan in first-half revenue, roughly $174 million, up 1,998%, but research spending still exceeds gross profit.

More
  • Smart-computing solutions supplied 1.168 billion yuan, about $165 million, or 95% of revenue; the package includes chips, boards, servers, software and cluster integration.
  • The company says it has passed supplier certification at major internet firms and delivered a several-thousand-GPU cluster in the Beijing-Tianjin-Hebei region.
  • Biren's software platform supports Chinese models including DeepSeek, Qwen and GLM, showing that domestic accelerator competition is increasingly about the full stack rather than silicon alone.
  • First-half gross profit was 527 million yuan, roughly $74 million, while research spending reached 804 million yuan, about $113 million, so growth has not removed the funding burden.
  • The comparison base was only 59 million yuan in first-half 2025, making the percentage increase spectacular but economically easier to achieve than a similar jump from a mature base.
  • The company must still prove that its next products, advanced packaging and large-scale interconnect plans can support sustained margins against Huawei, Cambricon and other domestic contenders.
  • Its results are an important market signal: Chinese customers are beginning to buy local compute systems at scale, but profitability and repeatability remain unproven.
Open at TMTPost Ask Claude ChatGPT

China's family-doctor expansion makes AI follow-up a policy priority

# TMTPost 14h

A new primary-care policy expands specialist and infectious-disease follow-up, pushing Chinese providers to use AI for routine patient management and coordination.

More
  • The policy described in the article shifts family-doctor work from generalist check-ins toward deeper collaboration with specialists and broader disease coverage.
  • The source links this expansion to a payment reform covering 158 primary-care conditions, intended to make treatment in local clinics financially comparable with larger hospitals.
  • Companies are taking different positions: some sell AI-assisted diagnosis and follow-up tools to government buyers, while others operate integrated doctor, hospital and payment networks.
  • One cited provider says its AI assistant covers more than 77,000 primary-care institutions and has generated over 1.2 billion diagnostic suggestions, though these are company-reported figures.
  • The deepest moat may be institutional integration rather than model quality: linking doctors, insurers, hospitals and patient records is harder to copy than adding a chatbot.
  • The policy explicitly calls for safe and controlled AI use, so deployment will depend on clinical accountability, privacy and reimbursement rules as much as on automation quality.
  • This is a strong example of Chinese industrial policy turning a previously optional AI feature into an expected layer of public-service capacity.
Open at TMTPost Ask Claude ChatGPT

MiniMax grows fast while its AI spending accelerates

# TMTPost 11h

MiniMax reported $116.6 million in first-half revenue, up 283%, but adjusted net loss widened to $293 million as enterprise growth was bought with lower prices and heavier training spend.

More
  • The article says the accounting loss narrowed partly because an IPO-related preferred-share revaluation charge disappeared; operating losses instead grew to about $327 million.
  • Enterprise and platform revenue rose to $73.93 million, 63.4% of total revenue, while consumer revenue still grew but conversion remained below 1%.
  • Research spending reached $297 million, about 2.5 times revenue, showing that the company is still financing model and product expansion ahead of profitability.
  • Gross margin fell to 17.9% despite the business mix shifting toward enterprise, which the analysis interprets as evidence that volume was purchased through aggressive pricing.
  • MiniMax's M3 model uses sparse attention and a claimed one-million-token context, but its attempted price increase was reversed after market resistance.
  • The company's video model H3 is presented as the stronger commercial asset, while text models face pressure from Kimi, DeepSeek, Alibaba and other Chinese competitors.
  • The strategic question is whether enterprise distribution can turn scale into durable pricing power before compute and customer-acquisition costs overwhelm the growth curve.
Open at TMTPost Ask Claude ChatGPT

A returning researcher becomes a Tsinghua AI class adviser

# AI QbitAI 6h

Tsinghua appointed returning robotics-learning researcher Xu Mengdi as one of four advisers for its elite AI-focused freshman class.

More
  • The 2026 class is reported at about 90 students, including roughly 27 top informatics competitors, making the appointment a signal about how elite Chinese technical training is being staffed.
  • Xu studied vehicle engineering as an undergraduate before moving through robotics and machine learning at Johns Hopkins, Carnegie Mellon and Stanford.
  • Her research emphasizes adaptation in physical environments rather than assuming that scale alone will produce general-purpose robots.
  • Her RoboTool work lets robots choose tools and generate executable plans from task goals, including unconventional uses of objects.
  • That background reflects a broader Chinese research priority: bringing overseas-trained talent back into institutions tied to national elite education and embodied-AI development.
Open at QbitAI Ask Claude ChatGPT

Tencent open-sources a 770B-parameter Hy4 preview

# AI GeekPark 13h

Tencent's Hy4 preview expands active parameters to 49B, context to 1M tokens and cuts prices, while targeting long, tool-using productivity workflows.

More
  • The model grows from Hy3's 295B total and 21B active parameters to 770B total and 49B active, according to Tencent; input costs 6 yuan per million tokens, roughly $0.85, and output costs 18 yuan, roughly $2.54.
  • Tencent says 163 experts blind-tested 203 engineering tasks, where Hy4 averaged 2.99 out of 4 versus 2.94 for Kimi K3 and 2.92 for GLM 5.3.
  • The more important product claim is workflow persistence: document review, coding, testing, error recovery and delivery rather than one-shot generation.
  • Tencent says Hy4 helped optimize its own training methods, data strategy, evaluations and operators, raising inference throughput 31.8% against a baseline.
  • The model is embedded across Tencent's office, coding and consumer products, giving the company a distribution advantage that independent model labs must buy.
  • The results are Tencent's own evaluation and the model is still a preview, so the benchmark gap is too small to establish broad superiority.
Open at GeekPark Ask Claude ChatGPT

OpenClaw's boom fades as agent harnesses become the new layer

# Analysis QbitAI 7h

Chinese coverage argues that OpenClaw has not died but has been absorbed into vendor products, while developers now focus on model-native harnesses and workflows.

More
  • The article says OpenClaw grew from a late-2025 launch to more than 380,000 stars and 80,000 forks, though those figures are not independently validated here.
  • Its original appeal was a local, always-on agent that could connect models to mail, browsers, terminals and messaging platforms; Chinese vendors then released more than 30 related products.
  • The framing shifts the story from a failed viral product to a successful abstraction: permissions, tools and skills became standard components inside larger agent runtimes.
  • The move toward harnesses reflects a deeper engineering concern than the branding cycle: context assembly, execution, sandboxing, persistence and verification determine whether an agent can finish long tasks.
  • Chinese coverage treats the short half-life of AI trends with humor but also as evidence that local vendors rapidly package global developer ideas into consumer products.
Open at QbitAI Ask Claude ChatGPT

China sends nine tunnel specialists to help Nepal's flood rescue

# Policy SCMP 14h

SCMP focuses on nine Chinese tunnel rescuers sent to Nepal, while Chinese coverage presents the deployment as one part of a broader state-led aid and data-sharing effort.

Also: Jiemian

More
  • The report says seven people had died on the Chinese side and 544 remained missing, while Nepal reported at least 675 deaths and nearly 2,500 missing.
  • The specialists were sent after a direct request from Kathmandu as teams searched through the critical early period after the floods.
  • Chinese official accounts add cash, humanitarian supplies, satellite imagery, hydrological data and support for citizens stranded on either side of the border.
  • The contrast is revealing: international coverage foregrounds the rescue clock and specialist personnel, while Chinese messaging integrates the operation into bilateral solidarity and national leadership.
  • The figures are changing rapidly and should be read as an interim operational picture, not a final toll.
Open at SCMP Ask Claude ChatGPT

A Chinese industrial robot firm wins a 100-unit oilfield order

# AI 36Kr 17h

Xuanchuang Robotics says it has delivered the first batch of more than 100 robots to China National Petroleum's Xinjiang oilfield system and is moving from inspection to remote operation.

More
  • The company's AEGIS architecture separates high-level task reasoning, motion simulation and a millisecond-scale safety barrier that can reject unsafe model commands.
  • Its products use wheels, tracks, rails, drones and dual-arm systems instead of humanoid bodies, matching each hazardous chemical environment to a specific mobility and manipulation need.
  • The company says it has more than 100 million yuan, roughly $14 million, in orders and that one chemical-site deployment can repay its cost within 18 months.
  • Remote operation is positioned as both an immediate safety product and a data-collection path toward autonomy: workers move from hazardous physical tasks to supervised decision-making.
  • The reported funding is several tens of millions of yuan, roughly $4-14 million, and will support data pipelines, vision-language-action models, world-model experiments and standardized production.
  • The shift from proof-of-concept purchases to batch procurement is the strongest commercial signal in the article, though customer economics and long-term autonomy remain unverified.
  • This is a useful counterexample to humanoid hype: in high-risk factories, certification, deterministic control and accumulated process knowledge can matter more than human form.
Open at 36Kr Ask Claude ChatGPT

China's Middle East strategy faces a test as the Iran war drags on

# Policy SCMP 8h

The war is pushing Gulf states toward self-reliant defense networks, challenging China's preference for regional involvement without a binding security commitment.

More
  • The analysis argues that Gulf governments now doubt both Washington and Beijing can protect them from Iranian attacks, weakening the value of diplomatic ambiguity alone.
  • China's approach has emphasized economic ties, political mediation and presence without a formal alliance or military guarantee; a more militarized Gulf could expose that limit.
  • The strategic question is whether Beijing can preserve influence while avoiding the cost and risk of becoming a security provider.
  • The article's framing is not that China is losing the region, but that a crisis is forcing partners to price security commitments more explicitly.
  • Energy, trade routes and diplomatic relationships give China strong reasons to stay engaged, yet the source offers no evidence of a new Chinese military commitment.
Open at SCMP Ask Claude ChatGPT

A Shanghai ice-cream shop may be a real robot deployment test

# AI TMTPost 11h

A dual-arm robot at a Shanghai Dairy Queen performs a 55-step Blizzard workflow with reported 90% success, using force control and no store redesign.

More
  • The robot handles opening equipment, taking cups, dispensing mix, adding toppings, blending and inverting the finished dessert, rather than demonstrating a single pick-and-place motion.
  • The developer, Sharpa, says it deliberately chose a tightly standardized but unmodified restaurant environment because household tasks share the same object and tool interactions.
  • The reported architecture combines planning, dexterous manipulation and tactile or force control; the key challenge is sustaining a long chain of small actions under real-world variation.
  • The company argues that remote operation is not a dead end: each supervised intervention supplies data for later autonomy while moving workers away from hazardous or repetitive physical tasks.
  • The deployment is strategically different from many exhibition demos that simplify objects, lighting and workstations to suit the robot.
  • The evidence remains company-reported and the source does not provide uptime, labor economics, maintenance cost or repeatability across stores.
  • If the system can maintain quality without custom fixtures and expand beyond one menu item, it would be stronger evidence for transferable embodied skills than another athletic demo.
Open at TMTPost Ask Claude ChatGPT

Xi and Putin may discuss Power of Siberia 2 in Bishkek

# Policy SCMP 6h

The Kremlin says Xi Jinping and Vladimir Putin are expected to discuss the Power of Siberia 2 gas pipeline at their meeting in Kyrgyzstan.

More
  • The report attributes the planned discussion to a Russian presidential aide and does not say that a final agreement has been reached.
  • For Beijing, the project would deepen a Russia energy relationship while potentially improving overland supply diversification; for Moscow, it offers leverage after European demand collapsed.
  • The meeting's setting and timing underline the strategic value of Central Asian transit diplomacy, even if the pipeline's economics and construction schedule remain unresolved.
Open at SCMP Ask Claude ChatGPT

China tests a stealth drone as its Type 076 ship nears trials

# Policy SCMP 15h

China's GJ-21 stealth attack drone has appeared in flight tests while the first Type 076 amphibious assault ship enters final sea trials.

More
  • The GJ-21 is described as a naval version of the GJ-11 stealth attack drone and is intended to operate from the Type 076 platform.
  • The combination points toward a carrier-like amphibious ship using unmanned aircraft for reconnaissance or strike missions, expanding the role of large deck ships beyond troop transport.
  • Public evidence remains limited: the report cites rare sightings and state-broadcaster computer-generated footage rather than a disclosed operational capability.
  • The strategic significance is the integration of ship design, unmanned aviation and long-range sensors, not simply another drone airframe.
Open at SCMP Ask Claude ChatGPT

Viture and Phantom Blade Zero turn XR glasses into game merchandise

# Culture Zhidx 16h

Chinese XR brand Viture partnered with the martial-arts game Phantom Blade Zero on a collector headset built around a 174-inch virtual display and themed software.

More
  • The headset keeps the game device as the processor and controller while replacing its small screen with a wearable display, a practical rather than fully immersive XR model.
  • The reported hardware offers a 58-degree field of view, variable transparency and a side-view mode that keeps the real environment visible.
  • The collaboration shows Chinese game and hardware brands using cultural style, packaging and collectible props to sell XR as an identity object, not only a display accessory.
  • The game's reported preorders and sales give the headset an existing fan base, but the source does not establish meaningful standalone demand for XR gaming.
Open at Zhidx Ask Claude ChatGPT

Intel pitches local AI video production on its B70 GPU

# AI QbitAI 7h

At a Guangzhou event, Intel and Chinese partners demonstrated a local video-generation workflow using the 32GB Arc Pro B70 and an open model from MiniMax.

More
  • The proposed workflow combines local inference, asset retrieval, super-resolution, editing and multi-GPU scheduling rather than treating the GPU as a standalone accelerator.
  • The article claims the card can run a 27B quantized language model at up to 52 tokens per second with a 256K context, but these are vendor-supplied results.
  • The China-specific signal is ecosystem alignment: Intel is using a Chinese open model and local software partners to build an alternative to cloud-only creative production.
  • The article is sponsored content, so claims about cost advantages, workflow compression and enterprise readiness should be treated as product positioning rather than independent validation.
Open at QbitAI Ask Claude ChatGPT

China maps dialect diversity through 1,800 field sites

# Culture SCMP 10h

A decade-long Chinese language project has collected dialect, minority-language and oral-culture recordings, now presented through interactive museum exhibits.

More
  • The SCMP report describes a museum booth that uses a visitor's speech to identify the likely home region, turning a national language archive into a public experience.
  • The collection spans more than 1,800 field sites, making it both a cultural-preservation project and a large speech-data resource.
  • The effort matters technologically because speech recognition systems trained on dominant Mandarin varieties can miss regional pronunciation, vocabulary and minority languages.
  • It also shows how Chinese cultural digitization combines state-scale collection, public education and machine-learning applications.
  • The source does not say how the recordings are licensed, whether communities control access or how the data will be used outside exhibits.
Open at SCMP Ask Claude ChatGPT

Chinese scholars warn AI translation can make policy sound harsher

# Policy SCMP 13h

A state-affiliated Chinese academic publication urges human review of AI translations so Beijing's political language is not rendered as more threatening abroad.

More
  • The recommendation reflects concern that literal or context-poor machine translation can alter the perceived intent of official policy language.
  • The proposed remedy is not to abandon AI translation but to retain human translators for political messaging, where tone and ambiguity carry strategic weight.
  • This is a useful window into how Chinese institutions view generative AI: valuable for scale, but risky when it mediates state communication to foreign audiences.
  • The article does not identify a particular mistranslation or establish how widely AI translation is already used in official communications.
Open at SCMP Ask Claude ChatGPT

A Chinese war game puts the chance of coastguard weapons use at 0.3%

# Policy SCMP 5h

A Chinese coastguard simulation reportedly models how patrols can avoid escalation in contested waters, with Chinese coverage stressing controlled decision-making.

More
  • The SCMP account describes researchers running scenarios in which fishing boats or small craft probe a coastguard ship and force split-second judgments.
  • The reported 0.3% figure is a result of a war game, not a forecast of real-world behavior or a measured probability.
  • The technical interest lies in escalation management: the simulation treats restraint, identification and maneuvering as operational variables before weapons use.
  • The story suggests Chinese maritime researchers are studying gray-zone encounters as a decision problem rather than only a force-balance problem.
Open at SCMP Ask Claude ChatGPT

China's vertical apps are losing users by trying to become malls

# Culture TMTPost 14h

Keep's monthly active users fell to 18.58 million as the fitness app shifted toward products, subscriptions and commerce, reflecting a wider Chinese platform fatigue.

More
  • Keep reported first-half revenue of 825 million yuan, roughly $116 million, nearly flat year over year, while online membership and paid-content revenue fell 26.9%.
  • Average monthly active users dropped from 36.4 million at the 2022 peak to 18.58 million in the first half of 2026; paying members fell to 2.169 million.
  • The article argues that many users, including committed exercisers, dislike feeds, stores and gated content displacing the simple tracking and workout functions that created loyalty.
  • The remaining users may be more valuable, but the lower membership penetration in the latest period suggests even the core audience is not automatically willing to pay.
  • The broader signal is about Chinese consumer software: vertical services can lose their original utility when monetization layers make the product feel like a generic marketplace.
  • Demographics also matter: older users are increasingly active in physical gyms, while younger users are less willing to sustain the discipline that home fitness requires.
Open at TMTPost Ask Claude ChatGPT

China says its AI-compute capacity has reached 2.45 million PFLOPS

# Signal IT Home 5h

China's data authority reports 2.45 million FP16 PFLOPS nationwide, with more than 85% in eight national hubs and three power-compute zones.

Also: Jiemian

More
  • The reported total equals roughly 2.45 exaFLOPS of FP16 throughput, though aggregate PFLOPS is not the same as usable model-training capacity.
  • Only 1.45 million PFLOPS is said to be connected to the national monitoring and scheduling platform, highlighting a gap between installed capacity and centrally visible capacity.
  • More than 85% of the capacity in the named hubs suggests Beijing is consolidating AI infrastructure around a small number of national compute corridors.
  • The policy agenda explicitly links compute supply with power coordination, monitoring and industrial ecosystem development.
  • The same briefing says agent-related patent grants exceeded 3,400 in 2025, but the text gives no utilization, revenue or model-performance data to show how efficiently the capacity is being used.
Open at IT Home Ask Claude ChatGPT

Huawei adds semantic image search and super-resolution APIs

# AI InfoQ China 6h

HarmonyOS 7's Core Vision Kit adds beta APIs for natural-language image retrieval and on-device super-resolution, aimed at real app workflows.

More
  • The semantic search API targets cross-modal retrieval over indexed images, while super-resolution improves low-resolution assets after the user finds them.
  • The developer account argues that the two capabilities should complement filenames, tags and OCR rather than replace them.
  • For a small team, a system-provided vision kit avoids building its own embedding, vector index, model-serving and device-compatibility stack.
  • The APIs remain beta, and the article says support, compilation and runtime quality still require device-level validation.
  • Huawei's privacy documentation says images are not retained by the cloud service, but applications remain responsible for explaining data processing and obtaining consent.
Open at InfoQ China Ask Claude ChatGPT

A Chinese humanoid robot outruns humans but crashes into the barrier

# Signal Solidot 8h

The Chinese humanoid Tiangong reportedly ran 100 meters in 8.64 seconds, exposing a gap between raw actuator speed and perception, braking and control.

More
  • The reported time is faster than Usain Bolt's 9.58-second record, but the robot took nearly a second to start, used more than 50 steps and did not slow before the finish.
  • Experts cited in the report say the gait reflects motor capability rather than human biomechanics, with short steps and high cadence reducing fall risk.
  • The robot reportedly stopped by hitting a padded barrier, underscoring that speed benchmarks can hide weak stopping, turning and decision-making behavior.
  • For real deployment, the next meaningful milestone is not a faster straight line but safe autonomous maneuvering under changing terrain and constraints.
Open at Solidot Ask Claude ChatGPT

Alibaba's Qoder moves coding agents beyond the IDE

# AI QbitAI 7h

Alibaba's Qoder desktop agent can turn natural-language requests into projects, edit existing repositories and verify results through browser and computer use.

More
  • The product article says Qoder has served 6 million users and 100,000 companies, but those figures are company claims in a promotional account.
  • Its architecture combines repository context, goal and plan decomposition, a tool harness, code execution and browser-based verification.
  • The useful technical distinction is not that it generates code, but that it can inspect an existing project, modify multiple files, run the result and revise failures.
  • The desktop companion and pet-like interface are a distribution experiment: Alibaba is trying to make coding-agent capabilities usable by non-programmers without removing the underlying execution complexity.
Open at QbitAI Ask Claude ChatGPT

Japan's fighter deployment to India is aimed partly at China

# Policy SCMP 7h

Japan plans its first fighter deployment to India for a joint exercise, signaling tighter security cooperation among states concerned about China's military rise.

More
  • The exercise is a Japan-India operation, but China is a substantive part of the strategic rationale rather than a passing mention.
  • The deployment indicates Tokyo is using training access and interoperability with partners as part of its broader response to China's military power.
  • For Beijing, the important variable is the gradual normalization of security cooperation across the Indo-Pacific, not the small exercise itself.
Open at SCMP Ask Claude ChatGPT

Chinese hard-tech brands move into premium shopping malls

# Culture TMTPost 11h

3D-printer and robotics companies are taking prime retail space in China, using malls for product education even as foot traffic rises faster than sales.

More
  • Bambu Lab's reported stores can draw more than 10,000 visitors a day, while Unitree's smaller robot showroom attracts hundreds on ordinary days and more on weekends.
  • The mall strategy addresses a consumer-education problem: expensive printers, robots and AI glasses are easier to understand through hands-on demonstrations than through online specifications.
  • The business case is less settled; the source cites shopping-center vacancy of 9.8% and falling average spend despite modestly higher traffic.
  • Chinese mall audiences skew toward young women, while early buyers of robotics and maker hardware skew more male, creating a possible mismatch between footfall and conversion.
  • The stores currently look more like commercialization laboratories and brand theaters than proven mass-retail channels.
  • If products gain recurring household use and accessories, these spaces could become durable retail; if not, malls may be subsidizing demonstrations with little repeat purchase.
Open at TMTPost Ask Claude ChatGPT

Xiaomi says its next foldable will use CXMT LPDDR6 memory

# Signal Techmeme 13h

Xiaomi and Chinese memory maker CXMT say a September Xiaomi 18 Fold will pair CXMT LPDDR6 DRAM with Xiaomi's in-house 3nm Xring O3 chip.

More
  • The claimed configuration would put Chinese DRAM and a domestically designed application processor together in a flagship foldable, making the phone a visible supply-chain demonstration.
  • CXMT supplying a premium device would matter more than a low-end substitution because it tests performance, yield and customer confidence under demanding thermal and power constraints.
  • The information comes through a Reuters item surfaced by Techmeme, and the companies' claim is not independently validated in the supplied text.
  • If shipment volumes are material, the event would show how Chinese handset makers can use vertical integration and domestic components to reduce exposure to foreign memory suppliers.
Open at Techmeme Ask Claude ChatGPT

CXMT sues the Pentagon over its Chinese military-company listing

# Policy Techmeme 17h

Chinese memory maker CXMT is challenging its Pentagon designation, saying it has no military affiliation and that its chips are not intended for military use.

More
  • The lawsuit places a major Chinese DRAM supplier directly inside the legal and geopolitical machinery of US technology restrictions.
  • The company's argument distinguishes commercial memory manufacturing from military affiliation, but the supplied report does not give the Pentagon's evidence or expected legal timetable.
  • Even without a final ruling, the designation can raise compliance, procurement and investor-risk costs for foreign firms dealing with CXMT.
  • The case matters alongside CXMT's reported move into premium smartphone memory: domestic technical progress is occurring while geopolitical classification becomes a parallel constraint on market access.
Open at Techmeme Ask Claude ChatGPT

Lijiang revokes licenses after tourists were pressured to pay

# Policy Jiemian 5h

Authorities in Lijiang revoked a travel agency and guide licenses and detained the guide after videos showed tourists being threatened over paid activities.

More
  • The tourism bureau said both videos concerned the same tour and found the guide had deceived and coerced travelers into spending money.
  • The response combined administrative punishment, license revocation and police detention, illustrating how viral consumer complaints can trigger a rapid local enforcement campaign.
  • The bureau's language promises a sustained crackdown, but the supplied report gives no evidence about wider industry conditions or recurrence.
Open at Jiemian Ask Claude ChatGPT

Taiwan's debt pressure offers a warning to its remaining allies

# Analysis SCMP 17h

A SCMP analysis argues that Taiwan has used debt leverage in ways rarely discussed alongside Western warnings about Chinese debt diplomacy.

More
  • The article contrasts Taiwan's lending and debt disputes with the much more familiar Western debate over Chinese bilateral loans.
  • Its subject is Taiwan's external financing behavior, but the comparison is relevant to Beijing because it challenges a narrative usually applied only to China.
  • The supplied excerpt offers a framing argument rather than a new Chinese policy decision or verified financial dataset.
  • The strategic value is rhetorical and diplomatic: competing development-finance narratives can shape which side small states see as coercive or dependable.
Open at SCMP Ask Claude ChatGPT

Huawei Cloud pitches an AI-ready data infrastructure stack

# Policy Leiphone 12h

Huawei Cloud presented a hybrid-cloud architecture linking governed data, trusted data spaces, model training and agent execution at China's data expo.

More
  • The proposed stack separates reliable data management, rapid data supply, intelligent use and trusted exchange across cloud, factory and office systems.
  • Huawei says its platform supports cross-zone and cross-region disaster recovery, large-scale metadata management and multimodal data lakes for training and inference.
  • The architecture includes privacy computing, data APIs, sandboxes and blockchain to control data ownership and prove operations when data crosses organizations.
  • Examples from power, manufacturing and logistics suggest the commercial target is regulated industrial and public-sector workflows rather than general consumer AI.
  • The claims are largely vendor and conference claims; the supplied text does not independently measure agent accuracy, deployment cost or customer return on investment.
Open at Leiphone Ask Claude ChatGPT

AI video tools are displacing performers in China's short-drama boom

# Culture Techmeme 8h

A Techmeme roundup says AI production made up 89 of the 100 most popular animated dramas on Douyin in May, intensifying pressure on online performers.

More
  • The item is an aggregator signal about China: it points to a labor and media shift rather than merely another video-generation product launch.
  • The cited DataEye figure suggests AI content is already concentrated in a high-volume, platform-native entertainment category.
  • The strategic issue is who captures the value: platforms and tool vendors gain scalable output, while actors, influencers and production workers lose bargaining power.
  • Because the source is a roundup of Financial Times reporting, the figure and labor impact should be treated as attributed reporting rather than independently verified fact.
Open at Techmeme Ask Claude ChatGPT

ChaPanda sells 17 million lychee milk drinks in 50 days

# Culture 36Kr 9h

Chinese tea chain ChaPanda says first-half revenue reached 2.655 billion yuan, about $374 million, as product innovation and supply-chain density offset slower store expansion.

More
  • The company reported 345 million yuan, roughly $49 million, in first-half profit and said its 8,863 stores were supported by 27 distribution centers.
  • The lychee drink's reported 17 million-unit run shows how Chinese beverage chains use short-lived flavors to create demand while testing repeatable product families.
  • Fresh-produce distribution reached 91% coverage, with about 95% of stores receiving at least two deliveries per week, making cold-chain execution part of the product strategy.
  • The expansion into breakfast, light meals, hydration and coffee suggests chains are trying to increase daily usage rather than rely only on afternoon tea traffic.
  • These figures come from company disclosures and an industry report; they do not establish that the new categories are profitable.
Open at 36Kr Ask Claude ChatGPT
Source ledger
ChinAI 403 0/0 HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk 200 0/20
Pekingnology 200 0/20
Ginger River Review 200 0/20
Sinocism 200 0/20
QbitAI 200 7/10
36Kr 200 3/30
Solidot 200 5/20
Leiphone 200 2/20
GeekPark 200 2/30
IT Home 200 12/60 +39 over cap
SCMP 200 10/50 +3 over cap
Initium 200 1/15
Jiemian 200 8/30 +22 over cap
China Digital Times 200 0/15
The Wire China 200 0/50
Techmeme 200 6/15
Hacker News 200 0/20
V2EX 200 8/50 +8 over cap
Zhidx 200 1/20
InfoQ China 200 3/20
New Wisdom 200 8/15 +1 over cap
TMTPost 200 8/19 +11 over cap