China Desk

China News, Summarized 13 Aug 2026, 17:06 UTC 34 stories 18/19 sources

Today's brief

DeepSeek spent two years being the floor under everyone's prices. On Monday it stops.

The model landed first, quietly. At around 3am Beijing time the API docs flipped: the fingerprint behind deepseek-v4-pro became the 0813 build, 111 days after the preview, with no blog post and no changelog. Architecture untouched — 1.6 trillion total parameters, 49 billion activated, 1M context, 384K max output. Everything that moved, moved in post-training.

And it moved a lot. DeepSWE, DeepSeek's own software-engineering benchmark, went from 12.8 to 62.7. Cybergym 52.7 to 83.3. Terminal Bench 72.1 to 87.9. Those are all long-horizon code, terminal and tool-calling tasks — exactly where the preview fell over. On pure knowledge reasoning it still trails: HLE 42.7 without tools against Claude Opus 4.8's 49.8 and Fable 5's 53.3.

Then came the invoice. DeepSeek is introducing peak and off-peak billing, off-peak at half the peak rate, new prices live at 16:00 UTC on 16 August. Reuters read the statement as rates ranging from 50% to 1,100% above current prices depending on model, token type and hour. A developer's tabulation on Hacker News put V4-Flash peak output at $1.32 per million against $0.28 today.

Look at when peak is. 01:00–04:00 and 06:00–10:00 UTC — the Chinese working day, morning and afternoon. Cache reads take the worst multiple, and cache reads are what agent loops burn. That is not margin expansion; that is load-shedding by price, from a lab that started hiring electrical and HVAC engineers for its own data centres two weeks ago.

Why now is in the leaked minutes. ChinaTalk has been working through notes from a four-hour meeting between Liang Wenfeng — the quant fund manager who founded DeepSeek and is now, on paper, richer than Dario Amodei or Sam Altman — and his investors. Enterprise revenue this year in the hundreds of millions of dollars. Net profit soon. Then an IPO.

A company priced for a listing cannot subsidise the planet's inference. The February 2025 disclosure that R1's API earned $562,027 a day at a 545% margin was a flex; it is now a floor to defend. Developers on V2EX spent the day joking that the rheostat needs a new setting and complaining that models were shipped and pulled, a mess. The cheap era was a customer-acquisition budget, and it has been spent.

What Beijing is signalling

Washington published 25 pages instead. The White House Office of Trade and Manufacturing Policy released "The Great Transshipment Scam", alleging Chinese exporters have laundered origin through more than 40 third countries since 2018 via relabelling, re-invoicing and minor assembly. It estimates the annual value at $40bn to $303bn and, assuming $75bn, claims 450,000 displaced American jobs.

The targets are the middlemen. Korea, Japan, Taiwan, the EU, India, Mexico and Canada are filed as "diversified scale leaders"; Vietnam, Malaysia, Thailand, Turkey, Brazil and Indonesia as deeply China-integrated ones. Nothing in today's mainland feed touches it — the Chinese press is on Zhu Rongji and V4-Pro. Xi arrives in Washington next month with this on the table and the investment board still unbuilt.

Nuclear is the AI answer nobody can sanction. CCTV Finance puts 57 reactors under construction and 135 GW operating-plus-approved, the largest build programme in the world. Eight units were approved in late July — the first batch of the 15th Five-Year Plan. Nuclear is under 2% of installed capacity but nearly 5% of generation, and above 15% in Guangdong, Fujian, Liaoning and Hainan.

Set that against yesterday's number: 800 billion kWh of compute electricity demand by 2030. Reactors approved now come online in the 2030s. Beijing is answering a question about tokens with twenty-year concrete.

And state TV is now bragging about exploit agents. A Fudan University team's Whitzard agent scored 91.2% on CyberGym, second globally and first among universities, ahead of Anthropic's security-focused frontier model and teams from Microsoft and Google, across 1,507 real vulnerabilities in 188 large open-source projects. One base model, path reachability recast as reasoning constraints, total inference cost under 5,000 yuan (~$700). Framework and benchmark open-sourced, and CCTV News carried it.

The AI race

Tencent's free cash flow went negative, on purpose. Q2 revenue 204.79bn yuan (~$28.8bn), up 11%; capex 52.78bn yuan (~$7.4bn), up 176% year on year and 65% on the quarter; R&D up 35%. The market hated it — shares closed down 4.46% the next morning.

The interesting claim is what the money buys. 36Kr's read is that Tencent is hoarding high-spec memory as much as accelerators, with Samsung, SK Hynix and Micron effectively sold out and expansion running one to three years behind demand. Hunyuan's next generation is meant to be bigger. Racing the clock on DRAM is a stranger bet than racing it on GPUs, and a more honest one.

DeepSeek shipped Harness the same day. A developer preview, one npm install, everything a plugin, built on an existing open-source plugin framework. What US developers on Hacker News fixed on was the raw reasoning trace being visible where American labs obfuscate theirs.

The leaked shortlist is the real strategy. Of 17 accepted ecosystem projects, about 70% are near-zero-star personal repos — sandboxes with credential injection at the gateway, cost-aware model routing, persistent "failure notebooks", multi-agent schedulers. Famous high-star projects were rejected; medical, legal and financial verticals are under 4%. DeepSeek is buying the execution layer and leaving applications to everyone else.

Meanwhile, the least glamorous work of the week. The FlagOS community got Alibaba's 2.4-trillion-parameter Qwen3.8 running day-zero on nine different accelerator families — Nvidia plus Ascend, Moore Threads, Kunlunxin, Hygon, MetaX, Enflame and others — largely by adding INT8 quantisation paths, because more than 80% of China's installed AI capacity cannot do FP8. A frontier model reaching hardware bought two generations ago is worth more than a leaderboard position.

The vibe

The Unitree lottery is the story of the week in China, and it isn't about robots. The humanoid maker's STAR Market allotment rate finished at 0.01809759%, a record low, with more than 9.78 million accounts applying for 19,414 winning numbers of 500 shares each. One Chengdu winner told reporters he expects about 200,000 yuan (~$28,000) and will pay cash for a Xiaomi car.

The detail that lands is what he won't do. He won't post it to WeChat Moments — afraid colleagues will resent him and relatives will ask for loans. Also in the book: Liang Wenfeng's two funds took the largest private-fund allocation by account count, 198 products for about 38.94m yuan (~$5.5m).

Someone in Hangzhou built a flying skateboard. Kufei launched a standing single-person eVTOL — no seat, no cabin, a flattened black platform with caged rotors, about 11 minutes aloft with a 70kg pilot — plus a seated model doing 100 km/h for 40 minutes, on a dual-redundant in-house flight controller with three aviation-grade IMUs per channel. It flew live at the Spring Festival Gala's Harbin stage at minus 30.

The route to market is American. Roughly 100 paid deposits overseas, aimed at FAA Part 103 ultralights, which need no airworthiness certificate and no pilot licence — plus a package from PICC and CPIC covering aircraft, occupant and third parties. A Chinese consumer aircraft finds its first legal home in a 1982 US rule.

Read differently

One release, two entirely different stories. Chinese tech press led on benchmarks and explicitly reported the price as unchanged at 3 yuan in, 6 yuan out per million (~$0.42/$0.85) — QbitAI called it a rehearsal of the 2025 Spring Festival. English-language developer forums skipped the benchmarks and went straight to the hike, doing substitution arithmetic against OpenAI's Luna price cut and noting DeepSeek is still roughly twice as slow.

Only the Chinese write-ups tested it properly. GeekPark found thinking mode consuming 13,394 of 16,384 output tokens on a pathfinding task and truncating the HTML mid-file; with thinking off, 715 clean lines in 54 seconds. The chain of thought itself has degenerated into clipped telegraphese to save tokens. Great for agents, probably fatal for prose.

Threads we are pulling

  • Called it, and fast. Yesterday's note on a registered Harness account and data-centre job ads; the developer preview shipped within the day.
  • Still missing: Alibaba's promised 27B dense companion to Qwen3.8. The multi-chip port covers only the 2.4T text checkpoint.
  • Lin Junyang's new lab is confirmed as an angel round at a $2bn valuation, co-led by Gaorong and HSG with Tencent and a Shanghai future-industries fund.
  • Zhu Rongji: tributes, including a personal one from Malaysia's Anwar Ibrahim, but no funeral grade yet. Who eulogises is still the thing to watch.
  • Insta360 outgrew its own category: revenue approaching 10bn yuan against a global panoramic market Leiphone puts under 7bn. Which explains the pivot.

The river · 34

DeepSeek Harness is an open, inspectable runtime for long-running agents

# QbitAI 1h

DeepSeek Harness is now open source; Chinese coverage stresses its plug-in architecture and trace-level observability, while US discussion focuses on the accompanying price increase.

Also: Leiphone, Techmeme, Hacker News, V2EX

More
  • The runtime makes models, tools, policies, storage, context handling and presets interchangeable, with more than 100 built-in plug-ins reported in the article.
  • Its trajectory view exposes event-level execution rather than only a polished chat transcript, including tool failures, token use and cache-hit behavior.
  • The architecture targets the gap between a model and a production agent: sandboxing, orchestration, memory and debugging become modifiable system components.
  • The article reports that long tasks often run far longer under Harness than under Codex, suggesting that orchestration can materially change completion quality even with the same model.
  • Chinese developer coverage frames Harness as an Android-like base layer for agents, emphasizing community modification rather than a finished consumer product.
  • The open-source bet matters if third parties build durable plug-ins and deployment patterns; without that ecosystem, it remains an unusually transparent coding shell.
Open at QbitAI Ask Claude ChatGPT

China's quantum sector becomes a policy-built industrial flywheel

# ChinaTalk 1d

A ChinaTalk deep dive argues that the new five-year plan and Party directives are turning quantum technology from a research field into a deliberately scaffolded industry.

More
  • The analysis says quantum technologies were placed first among future industries in the 2026 five-year plan, followed by directives from top Party and government bodies.
  • It identifies more than 40 first-half investment rounds, nearly 30 neutral-atom quantum-computing hardware companies and new quantum funds as signs of rapid commercialization.
  • China is using familiar industrial-policy tools such as scenario cultivation, pilot centers, challenge programs and patient capital to translate broad mandates into company formation and procurement.
  • The important distinction is between scientific breakthrough and state-created market structure: Beijing is trying to manufacture demand, validation sites and financing at the same time.
  • This policy machinery can accelerate a sector by coordinating capital and users, but it can also produce redundant firms and headline capacity before technical economics are proven.
  • For companies outside China, the signal is that quantum competition will be shaped not only by laboratory performance but also by coordinated deployment ecosystems and government-backed patience.
Open at ChinaTalk Ask Claude ChatGPT

DeepSeek V4 Pro closes its agent gap without changing its MoE architecture

# GeekPark 8h

DeepSeek quietly shipped V4 Pro 0813 with the same 1.6 trillion-parameter MoE design but much stronger agent benchmarks; Chinese tests stress tool use, while US discussion centers on price and cache economics.

Also: Leiphone

More
  • The model reportedly keeps 49 billion active parameters, a one-million-token context and a maximum output of 384,000 tokens, while post-training drives large gains on software and cyber benchmarks.
  • Reported scores rose from 12.8 to 62.7 on DeepSWE, from 52.7 to 83.3 on CyberGym and from 72.1 to 87.9 on Terminal Bench.
  • The practical lesson is that post-training and execution scaffolding can transform an existing architecture; the release is not evidence that parameter count alone produced the gains.
  • The article's real-task tests found that reasoning tokens can consume most of an output budget, causing long code generation to truncate unless the mode or token ceiling is changed.
  • At 3 yuan per million input tokens (roughly $0.42) and 6 yuan output (roughly $0.85), the model's strategic advantage is capability per dollar rather than absolute benchmark leadership.
  • The result becomes more important if Harness makes these agent gains reproducible across tools and environments; without that layer, API benchmarks may overstate production reliability.
Open at GeekPark Ask Claude ChatGPT

Remembering Zhu Rongji, China's reform-era premier

# The Wire China 2h

A Wire China essay reflects on Zhu Rongji after his death, using his reputation for blunt reform to examine a formative era in China's economic integration.

Also: SCMP, Initium

More
  • Zhu's importance lies in the combination of domestic restructuring and external opening, rather than in a single policy announcement.
  • The retrospective framing treats his personality and administrative style as part of the explanation for how reform was implemented.
  • His legacy remains contested because the reforms that improved efficiency also imposed large social costs, especially on state-sector workers.
  • The story is useful as political context for today's China: it contrasts a more market-oriented reform vocabulary with the current emphasis on state direction.
Open at The Wire China Ask Claude ChatGPT

Hong Kong's July cases show how national-security enforcement reaches cultural life

# Initium 10h

Hong Kong coverage details bookstore raids, a sedition conviction, judicial discipline and student policy litigation, showing how national-security governance now touches books, courts and education.

More
  • National-security police searched two independent bookstores and arrested five people over allegedly seditious publications, extending enforcement into the retail circulation of political ideas.
  • An 18-year-old with autism received a non-custodial order after conviction for sedition over graffiti, illustrating both the reach of the law and the court's use of supervised sentencing.
  • A judge was required to retire early after repeated findings of judicial copying, an institutional accountability case separate from political prosecutions but important to public trust.
  • The roundup also records the release of a former student-union leader after a riot sentence and a private prosecution involving a Wall Street Journal employer over union rights.
  • A court rejected a challenge by mainland Chinese students over subsidized university places, explicitly prioritizing applicants with closer local ties when public resources are limited.
  • Taken together, the cases show governance operating through policing, courts, education rules and labor law rather than through one single censorship mechanism.
Open at Initium Ask Claude ChatGPT

Open FlagOS adapts Alibaba's 2.4 trillion-parameter model across nine chip platforms

# Leiphone 8h

FlagOS says it made Alibaba's 2.4 trillion-parameter Qwen model available on nine Chinese and foreign accelerator platforms on release day, using quantization and a common software stack.

More
  • The model activates 95 billion parameters and supports a native 262,144-token context that can extend to about one million tokens, according to the release.
  • FlagOS uses BF16, FP8 and INT8 paths, with W8A8 quantization intended to reduce memory pressure on installed accelerators that lack native FP8 support.
  • The architectural value is portability: compilers, kernels, plug-ins and validation are shared so a model release does not require a separate engineering project for every chip.
  • The claim that adaptation can reach usable service within 24 hours is vendor-reported, but it captures a central Chinese infrastructure goal: extend the life of existing accelerator fleets.
  • A common open stack could weaken dependence on a single software ecosystem if precision parity and performance remain close enough in production.
Open at Leiphone Ask Claude ChatGPT

Chinese university agent ranks second in real-world cyberattack benchmark

# Jiemian 1h

A Fudan University agent reportedly achieved 91.2 percent success across 1,507 real vulnerabilities, ranking second globally and first among universities.

More
  • The evaluation covered 188 open-source projects and required vulnerability discovery, exploit construction and verification inside complete codebases.
  • The team used a modular framework with one base model and converted program reachability into reasoning constraints.
  • The report says the run cost under 5,000 yuan (roughly $700) and solved 1,374 vulnerabilities, though the figures are source-reported.
  • The result matters because it tests autonomous security work rather than static code generation, but independent reproduction and benchmark governance remain important.
  • Chinese coverage presents the result as evidence that domestic university teams can compete with major foreign labs on applied agent evaluation.
Open at Jiemian Ask Claude ChatGPT

China's nuclear fleet reaches 135 million kilowatts under construction or operation

# IT Home 2h

China reports 57 reactors operating or approved for construction and 135 million kilowatts of combined capacity, reinforcing nuclear power as coastal industrial infrastructure.

More
  • The reported fleet supplies nearly 5 percent of national generation while representing under 2 percent of installed capacity.
  • Four coastal provinces reportedly obtain more than 15 percent of their electricity from nuclear power, making the buildout regionally consequential.
  • The report says eight additional units were approved in late July, the first approvals of the current five-year planning period.
  • China's construction scale creates a long-lived demand base for engineering, heavy equipment, fuel services and grid integration.
  • The safety claims come from domestic reporting and operator metrics; they should be read as official performance reporting rather than an independent risk assessment.
Open at IT Home Ask Claude ChatGPT

Lenovo's infrastructure business drives a record quarter

# Signal QbitAI 10h

Lenovo reported record quarterly revenue of 183.4 billion yuan (roughly $25.8B), with infrastructure revenue nearly doubling as AI servers become a core growth engine.

More
  • Adjusted net profit exceeded $1 billion, up 176 percent year over year, while infrastructure operating profit reached about 5.3 billion yuan (roughly $750M).
  • Lenovo says AI-related revenue rose 60 percent to 63.4 billion yuan (roughly $8.9B), about 35 percent of group revenue, with AI services up 141 percent.
  • The result reflects a hybrid model: global sourcing and local delivery let Lenovo sell infrastructure without needing to own every layer of the accelerator stack.
  • The company is using AI PC, enterprise infrastructure and services together, seeking recurring deployment and support revenue rather than relying only on hardware cycles.
  • The claims are company reported, but the mix shift is strategically important: a Chinese PC incumbent is becoming a global AI-infrastructure integrator.
Open at QbitAI Ask Claude ChatGPT

The DeepSeek thesis: open models, learning and China's political economy

# Analysis ChinaTalk 6h

A ChinaTalk analysis uses leaked investor minutes to argue that DeepSeek is pursuing a long-horizon theory of learning and general intelligence rather than a conventional product strategy.

More
  • The analysis portrays Liang Wenfeng's thesis as a belief that automated learning, not simply larger pretraining runs, is the path toward general intelligence.
  • It argues that open weights can still support a business through enterprise services, giving DeepSeek a different commercial logic from closed model vendors.
  • The political-economy angle is important: the lab's ambitions are shaped by Chinese hardware constraints but framed against a much longer technological horizon.
  • The source is analysis based partly on leaked minutes, so its account of Liang's motives should not be treated as an official company position.
  • DeepSeek's unusual management and capital structure help explain why it can prioritize capability and openness over near-term product monetization.
Open at ChinaTalk Ask Claude ChatGPT

Hangzhou startup unveils a standing personal aircraft with redundant flight controls

# Analysis QbitAI 6h

Chinese startup Coolfly launched standing and seated single-person electric aircraft, pairing a low-flying consumer pitch with dual-redundant control architecture.

More
  • The standing Urban model reportedly uses multiple protected rotors and an expected 11-minute flight time for a 70-kilogram pilot.
  • Its NA80 controller uses two independent flight-control units and three aviation-grade inertial sensors in each unit, with cross-checking and failover.
  • The engineering challenge is human-machine coupling: a standing pilot changes the vehicle's center of gravity continuously, unlike a seated pilot constrained by a harness.
  • The company says it has secured nearly 100 paid overseas deposits and insurance partnerships, but regulatory treatment and real-world safety evidence remain unresolved.
  • The product shows how China's low-altitude-economy push is moving from drones toward consumer aviation, where certification, training and liability become the bottleneck.
Open at QbitAI Ask Claude ChatGPT

Didi's China ride orders rise for a fourteenth straight quarter

# Signal Leiphone 6h

Didi reports 3.648 billion China ride orders in the quarter, while US coverage emphasizes its return to profit and Chinese coverage stresses service quality and platform expansion.

Also: Techmeme

More
  • China ride orders rose 8.1 percent and gross transaction value rose 9.5 percent to 904 billion yuan (roughly $127B), according to the company report.
  • Didi says its matching assistant uses more than 90 service tags and a multi-agent architecture, with satisfaction in core scenarios up nearly 10 percentage points.
  • The company is broadening from ride hailing into delivery, travel, charging and driver services, a familiar Chinese platform strategy built around an existing high-frequency entry point.
  • The key question is whether higher service segmentation and AI matching improve unit economics or mainly add operational complexity.
Open at Leiphone Ask Claude ChatGPT

Unitree's IPO may be the opening act for China's humanoid-stock boom

# Analysis Initium 18h

Chinese analysis treats Unitree's IPO demand as a possible precursor to a broader humanoid-robot investment cycle, not merely a single listing event.

More
  • The framing centers on the tiny retail allocation and the possibility that public markets will give robotics firms a new source of capital and public legitimacy.
  • The article is analysis rather than a prospectus or valuation assessment, so expected returns and market size should not be treated as facts.
  • Unitree's appeal reflects a rare convergence of consumer visibility, robotics demonstrations and national industrial policy, which can amplify both capital inflows and speculative risk.
  • The key test after listing will be whether deployments, margins and manufacturing reliability catch up with the narrative surrounding humanoid robots.
Open at Initium Ask Claude ChatGPT

AdvNav shows how black-box vision attacks can derail navigation agents

# AI Leiphone 8h

A Hong Kong research team introduced a black-box attack that uses behavior feedback to mislead vision-language navigation agents, exposing a practical safety gap in embodied AI.

More
  • AdvNav does not need model parameters or gradients; it perturbs the agent's visual input, observes its path and optimizes the perturbation from trajectory and action feedback.
  • The framework combines path-level measures such as success weighted by path length with action-level drift, then uses adaptive perturbations and genetic search under limited queries.
  • The reported attack success rates reached 49.70, 65.96 and 87.30 percent on tested systems, but the benchmark setting is not equivalent to a deployed robot in an uncontrolled environment.
  • The result matters because navigation agents compound small perception errors over time, making a modest visual disturbance capable of producing a large route failure.
  • Chinese research coverage frames this as a missing evaluation layer: systems that work in clean environments may still lack a meaningful adversarial safety case.
Open at Leiphone Ask Claude ChatGPT

Pony.ai prepares fourth-generation autonomous trucks for fleet deployment

# Analysis Leiphone 13h

Pony.ai unveiled a fourth-generation autonomous heavy truck and targets 500 to 1,000 operating vehicles within two to three years, shifting the debate from demos to fleet economics.

More
  • The company says more than 80 percent of the software and over 90 percent of the hardware can be reused across its truck platform, while adapting to long-haul infrastructure.
  • Its priority cases are long-distance freight, remote bulk logistics and ports, where driver shortages, shift coverage and round-the-clock utilization create measurable economic pressure.
  • Port operations remain difficult because automated and human-driven vehicles share many facilities, so the system must handle negotiation rather than only follow a mapped route.
  • Pony.ai is testing both technology-as-a-service and transport-as-a-service, indicating that autonomy vendors still need to absorb operational responsibility before the market is ready for pure licensing.
  • The three-year target is a commercial proof point: success will depend on cost per ton-kilometer, uptime and regulatory permission, not just perception accuracy.
Open at Leiphone Ask Claude ChatGPT

Didi is using Brazil's ride-hailing network to move Chinese EVs into fleets

# Analysis Leiphone 7h

Didi's Brazilian platform and Movida plan discounted EV leases for 50,000 drivers, showing how Chinese companies export operating ecosystems as well as vehicles.

More
  • The partnership is expected to reduce rental costs by about 25 percent and addresses the financing barrier that prevents drivers from adopting lower-cost EVs.
  • Didi's 99 platform already combines rides, delivery, payments and local financial services, giving Chinese automakers an operating channel beyond dealership sales.
  • The model mirrors China's own EV adoption path, where high-mileage ride-hailing fleets helped turn new vehicles into everyday infrastructure.
  • The strategic test is whether local fleet economics, charging access and residual values can make Chinese EVs productive assets in Brazil rather than imported novelties.
Open at Leiphone Ask Claude ChatGPT

Chinese robotics teams argue that the body, not the model, is the bottleneck

# Analysis Leiphone 11h

A Chinese robotics analysis says reliable hardware, synchronized data and open control interfaces are being undervalued in embodied AI, despite the industry's focus on models and datasets.

More
  • The cited TA2 robot reportedly won three industrial tasks, including 213 sorting operations in 15 minutes and millimeter-scale assembly work.
  • Its control interface is shared between teleoperation, data collection and model inference, reducing the need to remap learned actions onto a different robot.
  • The article identifies latency, drift, sensor timestamp alignment and developer tooling as practical constraints that flashy demos conceal.
  • An open, high-performance body can function as public infrastructure for model teams by lowering the cost of experimentation and producing deployable data.
  • The strategic implication is that the winning robotics stack may be vertically coordinated across mechanics, control, data capture and software rather than dominated by a foundation model alone.
Open at Leiphone Ask Claude ChatGPT

US report accuses Chinese exporters of routing goods through third countries

# Policy SCMP 6h

A US report alleges Chinese exporters use limited assembly and relabeling in more than 40 countries to evade tariffs, raising the cost of future supply-chain scrutiny.

More
  • The accusation concerns transshipment practices developed during the US China trade conflict and is presented as an administration report, not a court finding.
  • If enforcement expands, Chinese manufacturers may face greater origin documentation, audits and exposure at intermediate production sites.
  • The story illustrates how tariff policy is shifting from border rates toward supply-chain provenance and beneficial ownership.
  • The report's rhetoric is part of the policy signal: Washington is treating third-country routing as organized industrial strategy rather than isolated customs fraud.
Open at SCMP Ask Claude ChatGPT

Tencent's AI spending surge is putting pressure on cash flow and memory supply

# Signal 36Kr 7h

Tencent's capital expenditure rose 176 percent year over year to 52.78 billion yuan (roughly $7.4B), as Chinese analysis links AI infrastructure spending to scarce memory.

More
  • Tencent reported 204.785 billion yuan (roughly $28.8B) in quarterly revenue and 67.276 billion yuan (roughly $9.5B) in operating profit, while free cash flow turned negative after investment.
  • The analysis argues that AI infrastructure requires not only accelerators but high-bandwidth memory and storage, making memory a strategic bottleneck rather than a peripheral component.
  • Tencent is positioning its current businesses as cash generators for larger models, inference capacity and products such as WorkBuddy and CodeBuddy.
  • The investment thesis is an opinion from a market-oriented article, not a company forecast; the important signal is Tencent accepting near-term cash pressure to secure compute supply.
  • For technical executives, the lesson is architectural: larger models and longer agent contexts shift procurement pressure toward memory bandwidth, capacity and data movement.
Open at 36Kr Ask Claude ChatGPT

Chinese researchers turn yeast into a thousand-fold cancer-drug factory

# Signal SCMP 10h

Zhejiang University researchers reportedly engineered yeast to produce vinblastine precursors at far higher yield, targeting a major supply constraint in an expensive cancer drug.

More
  • Vinblastine has traditionally depended on Madagascar periwinkle, with the report saying up to two tonnes of dried leaves yield one gram.
  • The work matters as a platform-engineering result: a biological host can potentially replace slow, land-intensive plant extraction with a controlled fermentation process.
  • The reported thousand-fold increase concerns production in the engineered pathway, not proof that a commercial drug supply chain is ready.
  • Scale-up, purification, regulatory validation and cost per clinical-grade dose will determine whether the laboratory result changes oncology manufacturing.
Open at SCMP Ask Claude ChatGPT

US and China compete for Ethiopia's $12.5 billion airport project

# Policy SCMP 12h

Ethiopia's new mega-airport has become a contest between US and Chinese contractors, making infrastructure procurement a direct arena for strategic influence.

More
  • The Bishoftu airport project is intended to serve as a major African trade and travel hub near Addis Ababa.
  • The report says Chinese state-owned firms are competing with US-linked and other international contractors, though the final award is not established.
  • Such projects bundle construction, financing, standards, operations and long-term political relationships, so the contract matters beyond the airport itself.
  • The episode shows how competition with China is increasingly fought through third-country infrastructure offers rather than only through export controls or military alliances.
Open at SCMP Ask Claude ChatGPT

Chinese startup raises several million yuan for embodied-AI data infrastructure

# Analysis QbitAI 5h

SCALEFORCE raised several tens of millions of yuan (roughly $4-14M) to build a physical-AI data network and its MatrixOS platform.

More
  • The company argues that embodied systems need synchronized visual, tactile, joint and physical-interaction data that cannot simply be scraped from the web.
  • It claims a data gap above 99 percent between required and currently usable real-world interaction hours, though the estimate is an industry assertion.
  • MatrixOS is positioned as an operating layer linking collection, processing, distribution and deployment rather than as a narrow labeling vendor.
  • The financing and early customer claims indicate that Chinese robotics companies are beginning to buy data infrastructure before general-purpose embodied models are mature.
Open at QbitAI Ask Claude ChatGPT

Chinese maker-tool brands chase a broader consumer market

# Analysis Leiphone 9h

Chinese industry coverage says 3D printers, laser engravers, desktop CNC and UV printers are moving from hobbyist niches toward mass-market tools, but cost and usability remain barriers.

More
  • The article cites a global maker-tool market of $18.9 billion in 2025, projected at $36.2 billion by 2033, though the figures come from cited industry research rather than a primary dataset.
  • The strongest Chinese category is 3D printing, where better calibration and faster machines have moved products from expert setup toward appliance-like use.
  • Desktop CNC has a harder path because dust, noise, fixturing, tool wear and computer-aided manufacturing create a longer workflow than additive printing.
  • The common bottleneck is component cost: imported laser sources, print heads and other parts keep machines above the price point needed for mass adoption.
  • The strategic opportunity is not just cheaper hardware but software that absorbs setup, material and maintenance complexity for ordinary households and small businesses.
Open at Leiphone Ask Claude ChatGPT

Satellite images suggest China is enlarging Type 054B missile capacity

# Policy SCMP 4h

Satellite imagery reportedly shows a revised launch-system shape on a Type 054B frigate under construction, suggesting more missile capacity and stronger air defense.

More
  • The report attributes the imagery to Vantor and says the newer bow launch system differs from the first two commissioned ships.
  • The change could increase vertical-launch reserves, but imagery alone does not establish the final weapon load or combat performance.
  • The story illustrates how open commercial satellite data is becoming a routine source for tracking Chinese naval production.
  • A larger launch system would matter most if it reflects a repeatable class-wide design change rather than a one-off construction variation.
Open at SCMP Ask Claude ChatGPT

Insta360's X6 turns a panoramic camera into an invisible director

# Analysis Leiphone 12h

Insta360's X6 combines panoramic, gimbal and first-person capture with automated editing, showing a Chinese camera maker shifting from hardware specifications to reduced user attention.

More
  • The flagship uses a larger sensor, dedicated imaging processors and 8K capture, but the more consequential change is the attempt to automate framing and post-production.
  • The company's original insight was that recording first and choosing the view later solved the attention problem in action sports; the X6 extends that logic to everyday memory capture.
  • The product's direction resembles a platform strategy: camera hardware, computational imaging, editing software and cloud workflows become one experience.
  • The challenge is whether users value an invisible camera enough to accept a more closed software workflow and recurring processing costs.
Open at Leiphone Ask Claude ChatGPT

Chinese robotics supplier raises nearly $14 million for miniature roller screws

# Signal 36Kr 15h

NOUSBOT raised nearly 100 million yuan (roughly $14M) to expand production of miniature planetary roller screws, a component bottleneck in humanoid robots and precision machinery.

More
  • The company reports a 1.5-millimeter screw diameter, 5.5-millimeter nut diameter, C5 accuracy and plus or minus 0.01-millimeter repeatability.
  • Its integrated actuator can reportedly generate 15 kilograms of force in a USB-drive-sized package, while newer versions add force, position and temperature feedback.
  • Planetary roller screws offer self-locking and intermittent motor operation, potentially reducing heat and extending life in compact linear joints.
  • The supplier says its factory currently produces three million units annually and targets ten million after the financing, but the demand forecast is management's own estimate.
  • The strategic signal is that humanoid-robot competition is pulling investment into small, difficult-to-manufacture transmission parts rather than only into motors, batteries and models.
Open at 36Kr Ask Claude ChatGPT

Chinese travelers criticize error-filled English signs in public spaces

# Culture SCMP 9h

A mistranslated sign in Gansu reignited online complaints about public English signage, revealing how tourism, local government quality and national image collide in China.

More
  • The reported error turned a hot-water-room sign into the phrase open water rooms, prompting an apology from the highway operator.
  • Similar airport and public-facility mistakes have repeatedly become social-media stories, suggesting that translation quality is treated as a visible proxy for administrative competence.
  • The issue is not merely linguistic: tourist-facing infrastructure is being built for domestic and international mobility, while local procurement often lacks robust review.
  • Chinese online reaction tends to focus on embarrassment and accountability rather than on language policy alone.
Open at SCMP Ask Claude ChatGPT

A Chinese startup builds a knee-focused hiking exoskeleton

# Analysis Leiphone 14h

WudaX is developing a hiking exoskeleton that supports the knee on descents as well as ascents, challenging the market's dominant hip-assist design.

More
  • The proposed knee architecture targets eccentric loading and downhill shock, where hip-focused products offer little support.
  • The founder says downhill motor operation could recover energy, analogous to regenerative braking, although the claim remains a product-level assertion.
  • The control philosophy follows user force directly instead of first classifying gait modes, potentially reducing mode-switch lag but making intent sensing harder.
  • The market is still exploring hip, knee and combined designs; the unresolved question is whether less visible assistance can be demonstrated and sold as effectively as obvious lifting power.
Open at Leiphone Ask Claude ChatGPT

Chinese missile researchers claim AI can classify stealth-fighter heat signatures

# Analysis SCMP 3h

A partial report says Chinese researchers used AI to distinguish heat signatures resembling advanced stealth fighters with over 90 percent accuracy, but technical details are missing.

More
  • The reported target signal is infrared radiation from exhaust and air friction, which differs from flare signatures.
  • The source text ends before describing the model, data, test design or operational conditions, so the accuracy claim cannot be evaluated.
Open at SCMP Ask Claude ChatGPT

Iflytek unveils enterprise agents across seven business domains

# AI QbitAI 7h

Iflytek announced a broad enterprise-agent portfolio spanning hiring, procurement, sales, IT, manufacturing, analysis and overseas operations, but the evidence is mostly corporate positioning.

More
  • The common architecture is a unified agent layer connected to domain workflows, documents, data and enterprise controls.
  • The procurement product describes a chain from rule interpretation through evidence retrieval, judgment and human review, illustrating where enterprise agents need auditability.
  • The company reports speech recognition across more than 100 languages and service nodes in five overseas locations, but does not provide independent adoption data.
  • The launch reflects China's enterprise-AI market moving from chat features toward workflow ownership and measurable operational outcomes.
Open at QbitAI Ask Claude ChatGPT

Chinese game platforms say AI has moved from creation to distribution

# Culture 36Kr 14h

At ChinaJoy, developers said AI can now generate playable game prototypes; the harder problems are discoverability, moderation and quality grading for the resulting flood of content.

More
  • TapTap says users created more than 4,000 AI-assisted games in the first half of the year, shifting the bottleneck from coding ability to finding players.
  • The platform and Alibaba Cloud frame the stack as model service plus agentic cloud: code generation needs sandboxes, tools, identity, evaluation and governance.
  • The Chinese market's concentrated distribution economics make exposure especially difficult for amateur creators, even when production costs collapse.
  • This is a useful counterweight to AI coding hype: when generation becomes cheap, platform ranking, compliance and audience access become the scarce resources.
Open at 36Kr Ask Claude ChatGPT

Chinese small exporters move into AI and robotics supply chains

# Analysis 36Kr 13h

XTransfer says Chinese small exporters are finding overseas growth in robotics, AI-server components and medical supplies, while AI agents increasingly automate payment compliance.

More
  • The company reports that small firms are entering new industries as component suppliers rather than trying to build complete robots or foundation models.
  • Its TradePilot system is used for onboarding, transaction verification and anti-money-laundering review, with 72 agents reportedly embedded in those workflows.
  • The pattern shows how industrial upgrading can reach small firms through narrow, exportable components such as frames, cooling parts and packaging.
  • The constraint is operational rather than purely technological: local collection, currency conversion and compliance differ across nearly every target market.
Open at 36Kr Ask Claude ChatGPT

Chinese AI labs seek harder reasoning problems at IJCAI 2026

# Analysis Leiphone 7h

Chinese coverage of IJCAI 2026 says domestic institutions remain prominent even after the conference's local ranking fell, with emphasis shifting from model scale to reasoning and robustness.

More
  • The article reports 6,500 submissions and 990 accepted papers, with reasoning, planning, constraint solving and search forming a large share of the main track.
  • It frames the ranking downgrade as a filter against paper-count optimization, though the article's statistics and interpretation are not independently checked here.
  • The industrial message is that Chinese companies are redirecting research toward explainability, optimization, planning and safety for high-risk applications.
  • The story is also a useful window into how Chinese industry evaluates academic prestige: less as branding and more as a pipeline for deployable technical methods.
Open at Leiphone Ask Claude ChatGPT

US developers weigh DeepSeek's price rise against cache performance

# AI Hacker News 4h

US commenters focus on DeepSeek's higher cost and latency, while Chinese coverage emphasizes Harness observability and unusually high cache-hit rates.

Also: V2EX

More
  • The Hacker News thread contains unverified user calculations about cache memory, provider pricing and comparative speed; they are useful as practitioner reaction, not market facts.
  • The comparison shows a different failure boundary: US users are recalculating model choice, while Chinese developers are asking whether orchestration and caching preserve the economic advantage.
  • If dynamic pricing persists, agent workloads will need explicit cache, routing and latency policies rather than a single static model choice.
Open at Hacker News Ask Claude ChatGPT
Source ledger
ChinAI 403 0/0 HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk 200 2/20
Pekingnology 200 0/20
Ginger River Review 200 0/20
Sinocism 200 0/20
QbitAI 200 10/10
36Kr 200 5/30
Solidot 200 9/20
Leiphone 200 19/20
GeekPark 200 2/30
IT Home 200 12/60 +42 over cap
SCMP 200 10/50 +9 over cap
Initium 200 4/15
Jiemian 200 8/30 +10 over cap
China Digital Times 200 0/15
The Wire China 200 1/50
Techmeme 200 3/15
Hacker News 200 2/20
V2EX 200 8/50 +40 over cap