China Desk — 14 Aug 2026, 05:43 UTC

The DeepSeek thesis: open models, AGI, and China's constraints

# ChinaTalk 19h

ChinaTalk's reading of leaked DeepSeek meeting notes presents Liang Wenfeng as pursuing automated learning and generalized intelligence above near-term commercialization, within China's hardware limits.

The essay says DeepSeek's open-model economics rely on business customers and strategic influence rather than simply charging for weights, but the supplied extract does not provide the full revenue picture.
Liang's reported thesis is that automated learning naturally leads toward generalized intelligence and that advanced machine intelligence is the central problem worth solving, a more ideological posture than the usual product roadmap.
China's hardware constraints are framed as an engineering condition rather than a reason to narrow ambition: the lab is said to be optimizing for a longer horizon while making its work broadly available.
The analysis contrasts DeepSeek's management and capital position with more conventional Chinese technology companies, suggesting unusual founder control and patience.
Leaked minutes and interpretive commentary are not equivalent to an authenticated strategy document, so the account should be read as a hypothesis about motive.
The strategic question is whether open release can create a business and talent ecosystem strong enough to compensate for weaker access to frontier hardware.

Open at ChinaTalk Ask Claude ChatGPT

Zhu Rongji dies at 97, leaving a contested reform legacy

# Initium 11h

Former Chinese premier Zhu Rongji died at 97, prompting renewed debate over the tax reforms, state centralization, and worker dislocation that shaped China's global integration.

Also: The Wire China, SCMP

Zhu's Shanghai tenure was associated with reform and Pudong development; at the center he pursued fiscal centralization and other measures that strengthened the central government's economic control.
The legacy is deliberately double-sided: admirers remember anti-corruption rhetoric and administrative toughness, while critics link the reforms to painful industrial restructuring and lost security for workers.
The fiscal reforms helped the central state capture more revenue and coordinate national policy, but also shifted burdens and incentives between central and local governments.
His political style was unusually candid by elite standards, which helped create a durable public image even as later governance became less personally expressive.
The renewed attention reveals how Chinese society uses retired leaders as vessels for arguments about today's political economy, not merely as historical figures.
Zhu's career sits at the boundary between reform-era openness and the stronger centralization that followed, making his death a useful lens on China's institutional evolution.

A tribute recalls Zhu Rongji's global and political legacy The Wire China
Malaysia's Anwar remembers Zhu Rongji as more than a statesman SCMP

Open at Initium Ask Claude ChatGPT

China tightens taxation of offshore trusts as fiscal pressure grows

# Pekingnology 3h

A Chinese economist frames new offshore-trust tax rules as a fairness move against wealthy tax deferral, while arguing that China's 45% labor-income ceiling is itself inequitable.

The Finance Ministry and State Taxation Administration rules tax key stages of offshore trusts, including transfers, income, distributions, and termination, narrowing a route for Chinese tax residents to defer or avoid tax.
The policy matters beyond wealthy families because it signals stronger scrutiny of offshore structures and a state willing to pursue revenue across legal and financial boundaries.
Professor David Daokui Li says ordinary people are largely unaffected because they do not hold offshore trusts, but uses the debate to question why labor income can face a 45% top rate while capital income tops out at 20%.
That framing is notable: official-adjacent public discussion can endorse enforcement against elites while still arguing for a lower and more uniform tax ceiling.
The source is a translated personal video and commentary, not a legal interpretation; the exact scope, enforcement mechanics, and affected asset base require the underlying rules.
The broader signal is fiscal centralization through tax administration rather than a headline wealth levy, which may change how Chinese entrepreneurs structure cross-border assets.

Open at Pekingnology Ask Claude ChatGPT

FlagOS makes Alibaba's 2.4T-parameter model usable across nine chips

# Leiphone 20h

Alibaba's Qwen3.8-2.4T-A95B was adapted for nine Chinese and foreign AI-chip platforms on day zero through the open FlagOS stack, using quantization to keep older hardware viable.

The MoE model has 2.4 trillion total parameters but activates 95 billion, with a native 262,144-token context extendable to about 1 million tokens; the scale creates a systems problem beyond model architecture.
FlagOS used BF16, FP8, and INT8 paths, compiler and operator changes, and cross-chip plugins to align accuracy across Nvidia, Huawei Ascend, Moore Threads, Biren, Hygon, and other platforms.
The reported adaptation window is under 24 hours, which could let smaller Chinese cloud providers offer new models without waiting for each chip vendor to build a bespoke stack.
The deeper policy and industrial signal is software portability: a common open layer can extend the useful life of installed accelerators and reduce dependence on one hardware ecosystem.
Alibaba's reported coding scores improved from 74.5 to 86.6 on Terminal Bench 2.1 and from 60.6 to 67.7 on SWE-bench Pro, but these are vendor-provided comparisons.
Quantization can reduce memory pressure, yet multi-chip performance depends on bandwidth, collective communication, kernel maturity, and model-parallel topology; a one-day launch does not prove production efficiency.
If FlagOS becomes widely adopted, the strategic bottleneck in China's AI stack shifts from merely obtaining chips toward maintaining a credible, shared software compatibility layer.

Open at Leiphone Ask Claude ChatGPT

DeepSeek V4 Pro closes the agent gap without changing its architecture

# GeekPark 20h

DeepSeek's V4 Pro production release keeps the 1.6T-parameter MoE architecture but sharply improves long-horizon coding and tool use through post-training, while retaining a low API price.

The model has 1.6 trillion total parameters, 49 billion active parameters, a 1-million-token context window, and a 384,000-token maximum output, unchanged from the preview.
Reported agent 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, indicating that post-training targeted environment interaction and code repair.
A hands-on test produced working Boids and productivity-panel applications from natural-language prompts, but a complex pathfinding task failed when thinking tokens consumed most of the output budget.
That failure is architecturally revealing: reasoning tokens and deliverable tokens compete for a fixed generation budget, so developers may need to tune thinking mode and maximum output rather than assume more reasoning is always better.
The reported API price is 3 yuan per million input tokens (about $0.42) and 6 yuan output (about $0.85), roughly one-tenth of the article's cited top closed-model pricing.
The model is not universally strongest; it reportedly trails leading closed models on some knowledge and repository tasks, but its combination of useful agent behavior and low cost is more commercially disruptive than a narrow benchmark lead.
The release also validates DeepSeek's Harness strategy: model quality and execution scaffolding are being developed as a coupled system rather than as separate product layers.

Open at GeekPark Ask Claude ChatGPT

Hong Kong's July courts show the expanding reach of security law

# Initium 22h

Hong Kong authorities searched independent bookstores, convicted an autistic 18-year-old under the sedition provisions of the security law, and continued cases involving judicial conduct and press freedom.

National-security police searched two independent bookstores and arrested five people over allegedly seditious publications; the stores later announced closure amid economic pressure and uncertain red lines.
An 18-year-old was convicted after writing protest slogans in a public bathroom, although the court later imposed an 18-month probation order, the first non-custodial sentence under the Article 23 sedition offense according to the report.
The case shows how the security framework reaches low-scale expression with limited distribution, not only organized political activity or mass protest.
A judge was pushed into early retirement after appellate findings of repeated judicial plagiarism, while a private prosecution involving a Wall Street Journal parent company tests the boundary between employment decisions and union rights.
A judicial-review challenge by 87 mainland Chinese dependents over subsidized university places was rejected on the grounds that public resources should prioritize people with a close connection to Hong Kong.
Taken together, the cases show a legal system under simultaneous pressure to enforce political red lines, preserve institutional credibility, and manage scarce public resources.
The pattern matters to companies and civil society: bookstores, journalists, courts, students, and employers all face a more politically legible but less predictable operating environment.

Open at Initium Ask Claude ChatGPT

Tencent spends $7.4B on a quarter of AI infrastructure

# 36Kr 20h

Tencent's second-quarter capex jumped 176% to 52.78 billion yuan (about $7.4B), pushing free cash flow negative as it races to secure compute and memory for AI.

Tencent reported 204.785 billion yuan (about $28.8B) in quarterly revenue and 67.276 billion yuan (about $9.5B) in operating profit, giving its mature businesses room to fund an infrastructure bet.
The spending is not only a GPU purchase story: larger models and faster inference require high-bandwidth memory, storage, networking, and data-center systems, making memory a strategic bottleneck.
The analysis argues Tencent is buying scarce storage ahead of demand because capacity expansion can take one to three years, but the article's supply-tightness claims are not independently measured.
The company is trying to use games, advertising, cloud, office agents, and consumer assistants as distribution channels for its Hunyuan models rather than rely on API sales alone.
Investors reacted negatively as near-term cash flow and profit came under pressure, showing that even a cash-rich Chinese platform must persuade markets that AI infrastructure will become operating leverage.
The key test is deployment efficiency: if Hunyuan, WorkBuddy, CodeBuddy, and related products generate durable usage, the capex can strengthen Tencent's ecosystem; if not, it becomes a costly hardware inventory cycle.
Tencent's move also matters to China's supply chain because a major domestic buyer can reinforce demand for local accelerators, memory, and server integration even when leading components remain globally sourced.

Open at 36Kr Ask Claude ChatGPT

RISC-V bet reaches Hong Kong IPO despite $690M in losses

# 36Kr 50m

Eeasy Technology, founded by former BOE chairman Wang Dongsheng, is seeking a Hong Kong IPO after losing 4.9 billion yuan (about $690M) in three years while raising 9.3 billion yuan (about $13.1B).

The company sells domestic human-machine-interface chips and RISC-V controllers, but its reported market share is modest: 5.7% overall and 1.2% in China's RISC-V controller market.
Its core strategic pitch is architectural substitution, using RISC-V to reduce dependence on x86 and Arm with backing from state-linked capital and a BOE-derived management network.
The financial profile is difficult: 2025 gross profit of 453 million yuan (about $64M) did not cover 604 million yuan (about $85M) of administrative expenses, before 1.042 billion yuan (about $147M) of R&D.
The company reported cumulative losses of 4.9 billion yuan (about $690M) from 2023 through 2025 and expected another loss in 2026; a reported 968 million yuan (about $136M) monthly cash burn makes the IPO a financing necessity.
Customer concentration is falling but remains material: one unnamed customer supplied 82.1% of revenue in 2023 and 39.3% in the first quarter of 2026, with the description strongly suggesting BOE.
The case shows how China's semiconductor industrial policy can produce large funding pools around strategic substitution even when commercial scale and unit economics remain unresolved.

Open at 36Kr Ask Claude ChatGPT

SHEIN's IPO pressure comes from its capital structure, not cash

# 36Kr 47m

SHEIN generated $6B of operating cash in three years and held $14.8B in cash, yet its IPO appears driven by $17.3B of redeemable preferred stock and deadlines around investor exit rights.

The analysis says recent operating cash flow was 17.57 billion, 14.05 billion, and 28.38 billion dollars from 2023 through 2025, exceeding the roughly $3.8B raised in equity financing over the prior 11 years.
Its negative working-capital model is central: customers pay before fulfillment, while inventory turns in about 36 days, allowing growth to fund operations.
The pressure is contractual rather than a simple liquidity crisis: certain investors' return rate rose from 8% to 12%, and redemption rights can revive if an IPO is withdrawn, rejected, or left inactive.
SHEIN's valuation history shows a sharp reset, from a reported $98.2B pre-money valuation in one round to $64B in the next, even as revenue and customers kept growing.
The business is still expanding, but revenue growth slowed to 8% in 2025 and 1.1% in the first quarter of 2026; US revenue also declined, making a public listing a way to convert an earlier private-market price into a more durable financing channel.
For investors, the important distinction is between operational cash generation and the obligations embedded in preferred shares; a cash-rich company can still face strong IPO incentives.

Open at 36Kr Ask Claude ChatGPT

Horizon Starry wins a flagship Chinese EV design slot

# IT Home 2m

Horizon says its Starry chip will power a premium cockpit project, putting a domestically designed cockpit-and-driving platform into another major EV program.

The single-chip design pools compute across cockpit and driving functions instead of maintaining separate hardware domains, with shared data paths and a unified vehicle-computing base.
Horizon claims roughly 50% less in-vehicle hardware space, a per-vehicle hardware saving of 1,500-4,000 yuan (about $210-560), and development time falling from 18 to 8 months; these are vendor claims, not independently verified results.
The significance is architectural as much as commercial: Chinese automakers are testing whether centralized, domestically sourced compute can reduce integration work while supporting local multimodal perception and on-device model inference.
The project follows an earlier iCar launch partnership, suggesting Starry is moving from a showcase release into a broader vehicle-platform sales process.
What matters next is production validation: sustained automotive-grade reliability, software isolation between safety-critical driving functions and consumer AI, and whether the claimed schedule savings survive mass deployment.

Open at IT Home Ask Claude ChatGPT

DeepSeek Harness is an agent saddle built for customization

# AI QbitAI 13h

A Chinese hands-on review finds DeepSeek Harness unusually suited to developers who want to replace models, tools, policies, storage, and UI, while still trailing mature coding agents in polish.

Also: GeekPark, Leiphone, Hacker News, V2EX

The product's distinctive feature is event-level observability: its trajectory view exposes raw agent actions and token costs instead of hiding execution behind a chat summary.
Its plugin architecture turns the runtime into a customizable layer for models, tools, context management, and policies, making it closer to an agent operating environment than a finished assistant.
The reviewer says many tasks run longer but produce more complete results, suggesting that orchestration and persistence can matter as much as the underlying model for complex work.
The system remains developer-first: it lacks a native desktop app, has unfinished interface elements, and needs another multimodal model for image inputs.
Reported cache hit rates were often near 99%, but this is one reviewer's experience and does not establish broad cost performance.
The launch's significance is ecosystem design. DeepSeek is trying to make open agents extensible at the runtime layer, where permissions, traces, routing, and skills can become reusable infrastructure.

Open at QbitAI Ask Claude ChatGPT

Lenovo's AI infrastructure business drives a record quarter

# Signal QbitAI 23h

Lenovo reported record first-quarter revenue of 183.4 billion yuan (about $25.8B), up 43%, as infrastructure revenue nearly doubled and AI-related sales reached 63.4 billion yuan (about $8.9B).

Adjusted profit exceeded $1B for the first time, up 176%, while the infrastructure-solutions group generated 57.9 billion yuan (about $8.2B) of revenue and an operating margin of 9.1%.
AI-related revenue grew 60% and represented about 35% of group revenue; AI services grew 141%, according to the company.
The result reflects Lenovo's hybrid model of global supply, local delivery, and enterprise integration rather than a pure consumer-device rebound.
The infrastructure business is becoming a second profit engine alongside PCs, positioning Lenovo to capture demand from cloud operators and enterprises building AI systems.
Lenovo says its personal AI assistant reached a 40% quarterly weekly-active rate on PCs, while its Qira product had about 28% conversion; these adoption figures are company-reported and definitions are not supplied.
The strategic issue is execution under component inflation and supply-chain restructuring: Lenovo can benefit from AI capex only if it preserves margin while delivering systems across many local markets.
The quarter suggests Chinese hardware companies may monetize the AI wave through servers, integration, and managed services even when they do not control the leading accelerator designs.

Open at QbitAI Ask Claude ChatGPT

Why Chinese users feel unable to speak plainly online

# Policy China Digital Times 4h

A translated Chinese essay argues that blunt platform keyword filters suppress ordinary speech and push users toward coded language, even when national law does not specify those exact blocked terms.

The essay distinguishes formal legal prohibitions from platform enforcement: companies face penalties for harmful content, so they often over-filter to reduce institutional risk.
Users respond with homophones, initials, and euphemisms, creating a technical and cultural layer of indirect speech that can evade filters but also makes public discussion harder to interpret.
The author places responsibility partly on platform negligence and risk management rather than attributing every blocked phrase to a central list, a framing that is safer politically while still criticizing how censorship is experienced.
This distinction matters for operators: moderation systems are shaped not only by statutes but by asymmetric liability, opaque escalation paths, and incentives to remove ambiguous content.
The post remains online according to the source, but one surviving essay does not show that the wider claims are tolerated consistently.
The broader signal is a censorship regime experienced through product design and corporate risk controls as much as through explicit state orders.

Open at China Digital Times Ask Claude ChatGPT

Claude helps construct long-missing Hadamard matrices

# AI QbitAI 18h

Chinese coverage says an Anthropic researcher and Claude produced constructions for all previously missing Hadamard-matrix orders below 2,000, including the long-elusive order 668.

A Hadamard matrix is a plus-or-minus-one square matrix whose rows are mutually orthogonal; constructing one at order 668 had remained an open challenge for decades.
The reported solution includes 12 matrices and was checked with exact integer arithmetic, but the benchmark organization has only provisionally marked the problem as solved.
The important research question is whether Claude discovered a generalizable construction or merely helped search and decode a one-off result; the technical paper is still pending.
This is a stronger signal than a benchmark score because the output is a mathematical object that can be mechanically verified, although attribution between human researchers and model remains unclear.
If the method generalizes, it could affect combinatorial design and related coding or signal-processing work; if not, the result is still a notable human-model collaboration.

Open at QbitAI Ask Claude ChatGPT

Hangzhou startup unveils a standing personal aircraft

# AI QbitAI 19h

Hangzhou company Coolfly introduced a standing electric VTOL craft, a seated model, and a redundant flight-control system, suggesting low-altitude aviation is moving toward consumer pilots.

The standing Urban uses a platform with protected rotors and is expected to fly for about 11 minutes with a 70-kilogram pilot; the seated Dream Pro is reported to carry 120 kilograms at up to 100 kilometers per hour for as long as 40 minutes.
The core engineering challenge is human-machine coupling: a standing pilot changes the aircraft's center of gravity, so control software must compensate for body movement as well as commands and disturbances.
Coolfly's NA80 flight-control system uses two independent control units and three aviation-grade IMUs in each, with cross-checking and claimed millisecond fault detection and takeover.
The company says it has secured nearly 100 paid overseas deposits and developed bundled insurance with Chinese insurers, but deposits are not evidence of certified or profitable operation.
The regulatory path is market-specific; US ultralight rules may simplify some requirements, while other jurisdictions could demand far more certification and operational control.
The product shows why making a personal aircraft easy to operate shifts complexity into redundancy, sensing, control law, and safety engineering rather than eliminating it.

Open at QbitAI Ask Claude ChatGPT

Didi's China rides grow for a 14th straight quarter

# Signal Leiphone 19h

Didi's China ride orders rose 8.1% to 3.648 billion in the second quarter, while the company expanded AI matching, charging, and overseas services.

China mobility gross transaction value reached 90.4 billion yuan (about $12.7B), while the core platform recorded 5.052 billion orders and 133.9 billion yuan (about $18.9B) in total transaction value.
Didi says its AI assistant uses more than 90 service tags and a multi-agent architecture to translate user preferences into matching decisions; satisfaction in core scenarios rose nearly 10 percentage points, according to the company.
The business is shifting from generic ride supply toward branded service tiers, with the selected-car category reporting a 46.3% fall in complaints and a 98% positive rating.
Didi's charging network covered more than 290 cities and 74,000 stations, showing how a mobility platform is using its demand base to extend into adjacent infrastructure.
The result is a useful example of Chinese platform maturity: growth now comes from service segmentation, operational reliability, and ecosystem cross-sell rather than only adding rides.
Company-reported growth does not resolve regulatory, driver-income, or profitability questions, but the sustained order trend is strategically meaningful.

Open at Leiphone Ask Claude ChatGPT

Didi links Chinese EVs to Brazil's driver economy

# Analysis Leiphone 19h

Didi's Brazilian platform 99 and rental company Movida will offer lower-cost vehicles to up to 50,000 drivers, creating an operating channel for Chinese EVs beyond direct car sales.

The plan addresses a real adoption barrier: Brazilian ride-hailing drivers face high fuel, maintenance, insurance, and depreciation costs but may not afford an electric vehicle outright.
The companies expect the rental program to reduce vehicle costs by about 25%, while a broader sustainable-mobility alliance has attracted more than 410 million Brazilian reais in investment.
This resembles China's own EV diffusion path, where ride-hailing platforms, financing, charging, and high-mileage drivers helped turn electric cars into everyday commercial tools.
The strategic model is platform-led industrial export: a local app, driver base, financial services, charging relationships, and Chinese vehicle supply reinforce one another.
The article says 99 has more than 60 million users and 2.5 million drivers and couriers in Brazil, but those company figures are not independently audited.
If the model works, Chinese automakers gain a route into recurring local usage and data, not merely a shipment channel; if rental economics fail, the vehicles remain expensive imports without a durable operating ecosystem.

Open at Leiphone Ask Claude ChatGPT

Chinese STEM students increasingly choose Singapore over the West

# Culture SCMP 7h

Chinese science and engineering students are redirecting toward Singapore as US, UK, Canada, and Australia tighten immigration and technology access, reshaping China's talent map.

The shift is driven by both pull and push factors: Singapore remains a regional research and business hub while Western study routes are seen as less predictable amid strategic competition.
For China, this is not simply an education preference; it affects where future researchers build networks, obtain advanced training, and potentially settle.
The article points to a more Asia-centered mobility pattern, but the supplied extract provides no enrollment totals or evidence that Singapore can absorb the displaced demand.
The long-term issue is whether restrictions intended to protect technology advantages instead redirect Chinese technical talent toward institutions and markets that are easier for China to access.

Open at SCMP Ask Claude ChatGPT

Zhejiang propaganda warns against anti-Western pseudo-history myths

# Culture SCMP 4h

A Zhejiang propaganda account has warned against online theories claiming Aristotle and ancient Greece never existed, revealing official concern about nationalist historical revisionism from the opposite direction.

The warning targets a growing online school that says Western civilizations were fabricated or that major achievements were stolen, according to the report.
This is unusual because Chinese state messaging often emphasizes civilizational confidence and Western decline; here it is policing a form of excess nationalism that undermines accepted historical knowledge.
The intervention shows that propaganda authorities are not only suppressing politically sensitive history but also trying to manage the boundaries of acceptable national pride and factual discourse.
The report does not say whether the warning triggered deletions, prosecutions, or a broader campaign, so its practical reach is unclear.

Open at SCMP Ask Claude ChatGPT

Chinese missile AI reportedly identifies stealth-fighter heat signatures

# Analysis SCMP 15h

SCMP reports that Chinese researchers trained missile AI to distinguish F-22 and F-35-like infrared signatures from flares with more than 90% accuracy.

The system targets a fundamental stealth limitation: aircraft can reduce radar visibility but still emit heat from engines and aerodynamic friction.
The reported approach uses machine learning to classify infrared patterns, potentially helping seekers distinguish an aircraft from decoy flares rather than simply detect a hot object.
The result comes from Chinese military research reporting and is not independently validated in the supplied text; test conditions, sensor range, and adversarial robustness are unknown.
If transferable from controlled data to contested environments, the work would shift part of the stealth competition from radar cross-section to thermal signature management and countermeasure design.
The strategic value is therefore conditional: accuracy against representative signatures matters only if the model remains reliable under changing aspect angles, weather, backgrounds, and deliberate deception.

Open at SCMP Ask Claude ChatGPT

Black-box attacks expose visual navigation weaknesses in robots

# AI Leiphone 20h

A Hong Kong university team built AdvNav, a black-box adversarial framework that perturbs a robot's visual input and reportedly achieved attack success rates up to 87.3%.

AdvNav needs no model weights or gradients; it observes trajectories and actions, then optimizes image perturbations using trajectory-level and action-level feedback.
The method combines adaptive perturbation strength with a genetic algorithm, addressing the delayed and sparse feedback of long-horizon navigation rather than a single image classification decision.
On the R2R benchmark, reported attack success rates against tested navigation systems were 49.7%, 65.96%, and 87.3%, but the article does not map each number clearly to a model and environment.
The architectural lesson is that a vision-language robot can be attacked at its sensor boundary even when its internal model is inaccessible, making perception integrity part of deployment security.
Real-world transfer remains the missing test: lighting, camera motion, occlusion, physical perturbations, and the robot's ability to recover could materially change the result.

Open at Leiphone Ask Claude ChatGPT

The cost of refusing model distillation

# AI 36Kr 3h

Chinese researchers argue that ByteDance's refusal to distill other models may preserve a more independent training system but slows coding and agent progress by removing a powerful cold-start shortcut.

Distillation lets a weaker model learn answers or tool-use trajectories from a stronger one, compressing the path to a useful baseline before further training.
Researchers cited in the article say a no-distillation strategy demands more data, compute, and post-training iteration, making it especially costly for long-horizon coding and agent tasks where successful trajectories are sparse.
ByteDance's Seed model reportedly ranked well behind Kimi K3 and GLM-5.2 in a coding benchmark, while its video model has succeeded without external-model distillation because real video data is harder to replace with synthetic outputs.
The strategic tradeoff is independence versus time-to-market: an internally generated data flywheel may produce a stronger long-term moat, but only if the team reaches enough real usage to collect feedback before competitors pull ahead.
The article links ByteDance's new AI Data and Safety department to this problem, suggesting data sourcing, cleaning, synthesis, and evaluation are becoming organizational pillars rather than support functions.
The choice is therefore workload-specific, not ideological; a policy that makes sense for video generation may be damaging for language agents.

Open at 36Kr Ask Claude ChatGPT

Moonshot AI needs an IPO story stronger than Kimi's latest benchmark

# Analysis 36Kr 1h

Chinese analysis says Moonshot AI may pursue a Hong Kong IPO after Kimi K3's strong benchmark showing, but warns that compute limits and fast-moving rivals could erase its valuation window.

The article argues that an IPO would provide a continuing financing channel, not merely one cash injection, citing Zhipu's roughly 31.4 billion Hong Kong dollar placement and MiniMax's 6.5 billion Hong Kong dollar convertible-bond deal.
Kimi K3's reported performance near top Anthropic and OpenAI models could support a premium narrative, but DeepSeek's newer model is said to offer less than one-tenth the per-task cost.
Moonshot's temporary suspension of new consumer subscriptions after K3 demand approached cluster capacity exposed a more concrete constraint: access to compute may limit monetization even when model quality is strong.
The model industry remains structurally difficult to differentiate because common advances in MoE, long context, reinforcement learning, and agent tooling are rapidly copied or distilled.
The IPO timing therefore matters: Moonshot must turn a temporary technical lead into financing before users, competitors, or infrastructure costs change the story.

Open at 36Kr Ask Claude ChatGPT

SCALEFORCE raises to build embodied-AI data infrastructure

# AI QbitAI 17h

Chinese embodied-intelligence data startup SCALEFORCE raised several tens of millions of yuan (roughly $4-7M) to build a real-world interaction data network and its MatrixOS platform.

Unlike language models, embodied systems need synchronized vision, touch, joint trajectories, forces, and environment state that cannot simply be scraped from the web.
The company says its platform spans collection, processing, distribution, and deployment, with a reported million-scale data order for a robot maker entering an automotive factory.
The strategic thesis is that data production capacity, labeling quality, and hardware-specific alignment may become the scarce infrastructure layer as robot companies move from demonstrations to continuous operation.
The article cites a huge industry funding total and a severe data shortage, but those figures are promotional estimates rather than an independently measured market dataset.

Open at QbitAI Ask Claude ChatGPT

China's canal link to Southeast Asia passes its first full-transit test

# Analysis SCMP 1h

SCMP's weekly roundup highlights China's first full transit test of a 134.2-kilometer, $10.8B canal project linking inland regions toward Southeast Asia, alongside regional security and space concerns.

The canal is presented as strategic infrastructure, not merely a transport project: it could reshape inland logistics, regional connectivity, and the economic geography around China's southwest.
The roundup format gives only a partial account, so cost, construction risk, environmental impact, and actual commercial demand remain unclear.
The value for a US technical reader is seeing infrastructure, regional influence, and state capacity discussed together rather than as isolated engineering milestones.

Open at SCMP Ask Claude ChatGPT

AI opens a new chapter for ancient Chinese literature

# Culture SCMP 1h

Chinese researchers are using AI to preserve, punctuate, compare, and interpret ancient texts, shifting scholarship from slow manual reconstruction toward searchable computational analysis.

The work addresses a specific data problem: manuscripts often lack punctuation, contain variant characters, and require comparison across records separated by centuries.
AI can help recover fragile material and surface relationships across large corpora, but specialist verification remains essential because errors in transcription or attribution can propagate into scholarship.
The story shows a culturally specific AI application in which preservation, national heritage, and research productivity reinforce one another.
It also illustrates a broader Chinese pattern: AI adoption is often framed as a tool for extending institutional capabilities and cultural continuity, not only as a consumer product.

Open at SCMP Ask Claude ChatGPT

Zhipu claims GLM-5.3 is the strongest open coding model

# AI IT Home 7m

Zhipu says GLM-5.3 gains sharply through post-training alone, with large benchmark improvements and weights promised after a safety review.

The base model is unchanged from GLM-5.2; the reported gains come from longer post-training runs, more task environments, reinforcement learning, and Zhipu's evolving training stack.
Reported scores rose from 4.6 to 28.3 on Terminal Bench 3.0, from 46.2 to 66.9 on DeepSWE, and from 23.8 to 28.5 on Agents' Last Exam.
Zhipu says its internal coding test reached 31.4% accuracy versus 29.5% for Claude Opus 4.8, while using about 50,000 output tokens per task versus roughly 120,000 for the comparator; these are company-run results.
The important research signal is that better post-training can still unlock substantial capability without changing the pretraining base, which puts pressure on teams that treat model size as the main moat.
Weights are scheduled to open two weeks after launch following security hardening, making the release a test of whether an open model can offer strong agentic coding without materially increasing misuse risk.
Chinese model companies are increasingly competing on tool use, long-horizon execution, and token efficiency rather than only on static knowledge benchmarks.

Open at IT Home Ask Claude ChatGPT

Washington labels Chinese tariff routing a Great Transshipment Scam

# Policy SCMP 18h

SCMP reports that the US accuses Chinese exporters of systematically routing goods through more than 40 third countries to evade tariffs, escalating scrutiny of global supply chains.

The White House report reportedly alleges that limited assembly and relabeling in lower-tariff jurisdictions allow Chinese goods to avoid direct treatment as China-origin exports.
The accusation could make supply-chain transformation and customs documentation strategic issues for manufacturers in Mexico, Southeast Asia, Latin America, and elsewhere.
The source summarizes a US government document and does not independently establish the scale of illegal activity or distinguish fraud from lawful production relocation.
The policy signal is broader than one enforcement action: Washington is treating third-country manufacturing networks as a potential extension of China's export system.

US accuses Chinese exporters of routing goods through third countries SCMP

Open at SCMP Ask Claude ChatGPT

China may increase missile capacity on its Type 054B frigate

# Policy SCMP 16h

Satellite imagery suggests a new Type 054B hull under construction in Shanghai has a larger vertical-launch system than the first two ships, potentially raising missile capacity by 50%.

The reported change concerns the shape and apparent capacity of the forward vertical-launch system, which could strengthen air defense and missile reserves.
The evidence is commercial satellite imagery interpreted by SCMP, not an official specification or sea-trial result.
A larger launcher would signal iterative adaptation of a naval platform rather than a wholly new class, but the article does not establish the final weapon mix or operational doctrine.
The broader pattern is China's rapid shipbuilding feedback loop: successive hulls can incorporate changes before an initial class has fully matured.

Open at SCMP Ask Claude ChatGPT

Shanghai trade reaches $423B in seven months, led by AI hardware

# Signal Jiemian 32m

Shanghai's imports and exports reached 3 trillion yuan (about $423B) in the first seven months, up 17.7%, with chips, memory, servers, new-energy products, and equipment leading growth.

Three AI-related product groups, including integrated circuits, memory components, and servers, totaled 926.7 billion yuan (about $131B), up 61.9% and contributing 8.9 percentage points to overall growth.
Exports of electric vehicles, batteries, and solar products rose 88.7%, 20.8%, and 79.2%, respectively; robot exports reached 330 million yuan (about $46M).
Shanghai's trade has grown for 18 consecutive months, with July trade up 12.6%, suggesting that the city's industrial and logistics base is still expanding despite external pressure.
Trade with ASEAN rose 54.3% in July, while both foreign-invested and private firms posted double-digit growth, pointing to a more diversified engine than state-linked exporters alone.
The data are customs aggregates, not proof that every AI segment is healthy; the key question is how much of the growth reflects re-exporting, component prices, or durable end demand.

Open at Jiemian Ask Claude ChatGPT

Microsoft is shrinking in China but keeping an engineering foothold

# Policy Solidot 1h

Chinese coverage says Microsoft has closed at least 15 China offices and joint ventures while retaining profitable services for firms operating overseas and access to Chinese engineering talent.

The report describes a two-sided relationship: Chinese software preferences and US export controls narrow Microsoft's local growth, while Chinese companies still need Western tools and management for global operations.
Microsoft's stated position is that it operates under the same regulatory environment as international suppliers and remains committed to the market; the article does not quantify current revenue or headcount.
The strategic signal is not a clean exit but a narrower China presence focused on enterprise services, engineering, and relationships with regulators.
This is a useful counterpoint to simple decoupling narratives: foreign technology firms can lose domestic expansion opportunities while preserving a specialized, commercially valuable foothold.

Open at Solidot Ask Claude ChatGPT

How Nongfu Spring overtook Master Kong in China's beverage market

# Culture Jiemian 42m

A Chinese business analysis says Nongfu Spring surpassed Master Kong in tea drinks and beverage revenue as a founder-led challenger outpaced an incumbent preparing for a family succession.

Master Kong reported first-half revenue of 40.545 billion yuan (about $5.7B) and profit of 2.433 billion yuan (about $343M), with tea revenue up 9.1%.
Nongfu Spring's tea revenue reached 21.596 billion yuan (about $3.0B) in 2025, overtaking Master Kong's 20.603 billion yuan (about $2.9B); its total beverage revenue also exceeded Master Kong's.
The analysis frames the shift as a difference in operating posture: Nongfu Spring remains aggressive on channel, pricing, and brand positioning, while Master Kong is trying to recover speed under the founder's successor.
This is a useful consumer-market case because it links succession, distribution discipline, and category renewal rather than treating market share as a pure product contest.

Open at Jiemian Ask Claude ChatGPT

China's maker-tool exporters face a growth ceiling

# Analysis Leiphone 21h

A Chinese industry analysis sees 3D printers, laser engravers, desktop CNC, and UV printers becoming a global consumer market, but warns that price, safety, and ease of use still block mass adoption.

The article cites a global DIY maker-tool market of $18.9B in 2025, projected to reach $36.2B by 2033, and says 3D-printer leader Bambu Lab has raised its annual sales target to 24 billion yuan (about $3.4B).
China's advantage is not only manufacturing cost; brands are turning specialist machines into consumer products through automated calibration, faster workflows, retail distribution, and local content.
Desktop CNC shows the harder path: dust, noise, fixturing, CAM programming, tool wear, and cleanup make it much less appliance-like than 3D printing.
The analysis says overseas prices would need to fall below roughly $500 for laser engravers to reach broader adoption, but key lasers, print heads, and other components still depend on foreign supply chains.
The strategic question is whether AI can remove enough setup and process complexity to turn maker hardware into a general household tool rather than a niche workshop product.

Open at Leiphone Ask Claude ChatGPT

A new Mecca defense pact reshapes the Middle East, with China watching

# Policy SCMP 8h

SCMP analyzes a Saudi-Turkish-Pakistani defense agreement that treats an attack on one as an attack on all, placing China's regional relationships in a changing security architecture.

The pact links Gulf capital, Turkish military reach, and Pakistani security commitments into a structure that could reduce reliance on older US-centered arrangements.
China's relevance is indirect but substantive: Beijing has economic, diplomatic, and energy interests across the participating states and has positioned itself as a broker in regional reconciliation.
The supplied extract does not establish China's role in the agreement or predict a specific policy response, so the story is strategic context rather than a China action.

Open at SCMP Ask Claude ChatGPT

ChinAI	403	0/0		HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk	200	1/20
Pekingnology	200	1/20
Ginger River Review	200	0/20
Sinocism	200	0/20
QbitAI	200	10/10
36Kr	200	9/30
Solidot	200	11/20
Leiphone	200	14/20
GeekPark	200	4/30
IT Home	200	12/60	+43 over cap
SCMP	200	10/50	+7 over cap
Initium	200	4/15
Jiemian	200	8/30	+18 over cap
China Digital Times	200	1/15
The Wire China	200	1/50
Techmeme	200	4/15
Hacker News	200	1/20
V2EX	200	8/50	+38 over cap

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

What Beijing is signalling

The AI race

The vibe

Read differently

Threads we are pulling