A Chinese energy company has opened a 2GW campus in Inner Mongolia centered on a claimed 120,000-square-meter AI data-center complex, making electricity delivery and thermal design the strategic constraint.
More
- Envision says its Star River base has a planned capacity of 2GW, support for one million accelerators, and a compute scale described as one million petaflops; those headline figures are company claims and should not be treated as demonstrated operating capacity.
- The engineering problem is no longer simply obtaining GPUs: large synchronized clusters require stable power, fast dynamic response, storage, high-efficiency conversion, and liquid or mixed-mode cooling.
- The site uses a forecast-decision-execution energy-management loop, storage, direct renewable-power connections, and an 800-volt direct-current architecture intended to reduce conversion losses between the grid and GPU racks.
- The choice of an energy company as the developer reflects a shift in Chinese infrastructure strategy, where access to power, land, grid capacity, and industrial construction can matter as much as server procurement.
- The location's reported renewable share is 67 percent, but renewable generation alone cannot guarantee the continuity and power quality required by long training runs; storage and grid integration are the harder test.
- If the campus operates at anything close to its stated scale, it would make power availability a first-order determinant of model capacity in China and increase the value of energy-aware scheduling and thermal engineering.
- The claims matter less as a finished benchmark than as a signal that Chinese AI infrastructure is being planned as an integrated energy system rather than a conventional server farm.
Manus is splitting from Meta after Beijing blocked the acquisition on national-security grounds, showing how Chinese regulatory sovereignty can reverse even a major AI transaction.
Also:
Hacker News
More
- The reported unwind turns Manus into a test case for whether Chinese-founded AI companies can be acquired by US technology groups without transferring control of data, models, or operating infrastructure.
- Manus says some post-acquisition user data will be deleted and that the independent service will store data in the United States and Singapore, underscoring the operational cost of regulatory separation.
- The episode reflects a broader pattern in which AI is treated as strategic infrastructure rather than an ordinary software asset when ownership crosses the China-US boundary.
- Meta loses a fast-growing agent platform, while Manus regains independence at the price of a disruptive data transition and a more limited corporate structure.
- The long-term significance depends on whether Manus can retain users and technical momentum after the breakup; otherwise, the deal becomes a warning to investors rather than a durable precedent.
- Chinese coverage presents the separation mainly as a compliance-driven return to independence, while the US-facing account frames it as another high-profile decoupling in AI.
After a stray dog was allegedly beaten and burned by four boys in Guangdong, Chinese activists used coded language and images to keep the case visible despite platform throttling and police pressure.
More
- The dog, known as Wangwang, was reportedly killed in late June after being beaten and doused with flammable material; the source cites online footage and local reporting.
- Authorities said all four suspects were under 14 and had been sent to special educational institutions, a response that failed to calm public anger.
- Activists substituted a snack brand, altered characters, dog-like clouds, and AI-generated images after the dog’s name appeared to trigger throttling on major platforms.
- One activist says police visited her home after she displayed a support video in a commercial district, illustrating how online pressure can become a local security matter.
- The episode shows censorship operating through visibility and reach rather than only deletion: users can post, but coded language and reduced distribution make sustained mobilization difficult.
- It also reveals a growing constituency for animal protection that can organize outside formal civil-society channels while remaining vulnerable to child-protection and public-order rules.
- The broader question is whether repeated animal-abuse controversies will produce legal reform or simply more sophisticated online workarounds.
ByteDance has made AI data and safety a standalone top-level unit, signaling that scarce training data, quality control, and compliance are becoming strategic functions rather than support work.
More
- The new division sits alongside Seed, Flow, and Douyin and is led by Wang Yinglei, formerly associated with TikTok platform responsibility and livestreaming.
- The reported organization combines Global Data, the group data center, and Flow's AI data platform, giving it responsibility across sourcing, synthetic data, cleaning, evaluation, and security compliance.
- Its remit spans ByteDance's foundation models and products, suggesting that data pipelines are being centralized as a shared internal platform rather than left to individual model teams.
- The move follows ByteDance's stated preference for self-developed models rather than relying on distillation, a strategy that requires more proprietary data and human evaluation.
- The source says ByteDance has used more than a thousand people to evaluate some video-model training data and has budgets in the multi-million-dollar range for specialized datasets, illustrating the labor intensity behind frontier-model development.
- The strategic question is whether a large centralized data organization can stay responsive when the value of a dataset may decay within months; scale alone does not guarantee useful training signal.
- Chinese reporting treats the reorganization as evidence of a data arms race, with the emphasis on internal capability-building rather than on a single model launch.
Unitree’s CEO said robot demand has a long runway but acknowledged slowing growth, heavier research spending, and a 52.55 percent first-quarter drop in adjusted profit.
More
- At an IPO roadshow, Wang Xingxing answered more than 200 questions about Unitree’s valuation, products, funding, and commercialization. The company’s disclosed 2025 figures imply adjusted earnings of 590.75 million yuan and a price-to-earnings ratio of 92.92 times, while the article also cites a 219-times figure under another calculation. First-quarter revenue growth slowed to 68.49 percent from 332.64 percent in 2025, and adjusted profit fell 52.55 percent year on year. Wang attributed the decline to a higher base, more research spending, and marketing, while offering no near-term earnings guidance beyond long-term industrial and household adoption.
Ant Group led a multimillion-dollar round for Daimon Robotics, its first reported investment in tactile sensing, as the Chinese robot startup links visual-touch sensors, interaction data, and control models.
Also:
Leiphone
More
- Daimon Robotics says it raised several hundred million yuan, roughly $30 million to $60 million, in a strategic round led by Ant Group. The company had completed a yuan-denominated Series A about two months earlier, making the timing notable even though exact totals are undisclosed. Daimon is building tactile sensors, tactile data infrastructure, and models for dexterous manipulation rather than another robot body. Its new Daimon-TWM is described as a tactile-grounded world model that combines physical state estimation, planning, and rapid control. The company says the model reached 95% success in demonstrations involving fruit packing and pencil-case organization, including disturbances and flexible objects. It claims tens of thousands of hours of multimodal tactile-operation data and more than 4.4 million downloads for an initial 10,000-hour open dataset. A planned partnership with China Mobile will deploy 1,000 collection devices and potentially generate 1 million hours of real-world operation data annually. The engineering rationale is sound in outline: vision estimates appearance and position, while touch reveals contact, slip, force, and misalignment. The difficult question is whether high-frequency tactile signals can be fused with language and action models without making deployment too sensor-specific. The source is largely company material and does not provide a controlled comparison with vision-only systems, task definitions, or long-term production failure rates. Ant's investment is strategically interesting because it extends its robotics portfolio from bodies and brains into the physical interface where manipulation data is created. If the data loop scales, tactile sensing could become a valuable platform layer; if not, it may remain an expensive collection of custom sensors and demonstrations.
The US has lifted its federal-device TikTok ban after a new joint venture separated data protection and algorithm security from ByteDance, showing how ownership structures can reshape national-security policy.
More
- The White House rescinded a 2023 directive that barred TikTok from federal devices, according to the report.
- The decision followed a determination that the restructured US operation was no longer covered by the earlier prohibition aimed at applications controlled by ByteDance.
- The new entity reportedly handles US data protection, algorithm security, content review, and software assurance, while ByteDance retains a minority stake.
- This is not a return to unrestricted trust: it is a governance experiment in which technical and operational separation is used to manage perceived foreign-control risk.
- The arrangement may become a template for other cross-border platforms, but its credibility depends on enforceable access controls, auditability, and whether the new entity can make consequential product decisions independently.
- Chinese coverage presents the change mainly as a result of corporate restructuring, while the policy question for US officials is whether formal separation can reduce state-access risk.
A US State Department report faults China for insufficient disclosure of overseas lending, turning debt transparency into another arena of strategic competition.
More
- The report says China failed to meet minimum transparency requirements for domestic finances and claims against foreign borrowers.
- The central concern is not simply the size of Chinese lending but the difficulty of determining repayment priorities, collateral, restructuring terms, and exposure when borrowers encounter stress.
- For emerging markets, opaque bilateral lending can complicate coordination with other creditors and make debt crises harder to resolve.
- The report is a US government assessment, not an independent audit, and the provided text does not quantify the loans or defaults at issue.
- The dispute reflects the broader contest over whether Chinese state-backed finance should be treated as commercial lending, development assistance, or strategic influence.
A former Kimi CLI leader argues that stronger models will not eliminate harnesses; they will shift complexity from patching model weaknesses to coordinating agents, state and permissions.
More
- The analysis defines a harness as the runtime layer governing execution loops, context and state, tools, resources, safety and boundaries rather than merely a system prompt.
- As models absorb basic tool-use and formatting skills, the remaining engineering burden moves upward to multi-agent coordination, cross-session memory, handoffs and dynamic permissions.
- This is an architectural version of complexity migration: a model can internalize local behaviors while the surrounding system still has to provide determinism, observability, isolation and policy.
- The former Kimi leader’s examples include replacing dedicated subagent scheduling with shell-driven decomposition, showing where model capability can simplify a lower layer without removing the runtime.
- For technical executives, the implication is that model selection alone will not produce reliable long-running agents; orchestration and governance become more important as task scope expands.
- The article is an expert opinion from a Chinese AI engineering context, not a measured comparison of harness designs or a universal law of model scaling.
A Chinese military analysis argues that Beijing's carrier force should be judged by doctrine and mission, not by a simple count of hulls versus the US Navy.
More
- The report places China's carrier development within the People's Liberation Army's centenary modernization timeline and its broader military strategy.
- The central analytical distinction is between platform numbers and the operational concepts that determine how carriers support maritime, air, and regional missions.
- China's carrier force is developing under different geography, industrial constraints, and strategic priorities from the US fleet, so numerical comparison can obscure rather than clarify capability.
- The provided excerpt does not give the article's estimates or concrete doctrinal conclusions, limiting how far the comparison can be taken.
- The value for observers is methodological: Chinese military power should be read through force integration, logistics, training, and mission design rather than procurement headlines alone.
Guoao Technology raised nearly 100 million yuan, roughly $14 million, after deploying its dual-axis motor robots in Bosch and other industrial supply chains.
More
- Guoao's core product is a dual-axis motor that combines linear and rotary motion in one actuator without conventional mechanical transmission. The company says its robots have been sold to Bosch and are in trial operation at several major factories. Its G-Tools robot reportedly achieved plus-or-minus 2 micrometers of positioning accuracy and plus-or-minus 0.05 newtons of force control while inserting a 0.5-millimeter needle into a hole. The company says its motors improved precision, reliability, and throughput by three to five times in die-bonding equipment while lowering system cost. Guoao's production capacity is reported at 1,800 to 2,000 units per month, with additional factories planned. Its integrated robot business currently contributes less than 10% of revenue, while precision motors remain the financial base. The company explicitly positions itself against ABB-style industrial automation rather than general-purpose humanoids such as Optimus. That is a meaningful distinction: fixed-task industrial robots can win through repeatability, force control, and uptime without solving open-world autonomy. The report does not provide Bosch acceptance criteria, production uptime, or independently measured accuracy. The financing is company-reported, but the reported industrial deployment makes this more consequential than another humanoid demonstration.
Prosecutors say intermediaries defrauded Tzu Chi of about $30 million during a 2021 BNT purchase attempt, reigniting Taiwan's dispute over private vaccine procurement.
More
- Taiwanese prosecutors said on August 6 that a lawyer and broker falsely claimed to have access to BNT vaccines and persuaded Tzu Chi to commission 5 million doses. The foundation reportedly paid a $30 million commission, roughly NT$1.06 billion, excluding the vaccines and cold-chain transport, but received no doses from the intermediaries; it later signed for 5 million doses with Shanghai Fosun Pharma. Prosecutors allege that about NT$740 million was transferred into a personal account, with 232 kilograms of gold purchased and NT$200 million lent to a Hong Kong businessman. The case has reopened a political dispute over whether Taiwan's government obstructed private vaccine purchases, with the ruling party demanding an opposition apology and the opposition refusing; the supplied account reports allegations, not final convictions.
Shanghai startup Yingsu says it can convert ordinary monocular video into dynamic 3D scenes at one or two orders of magnitude lower cost than physical data collection.
More
- The company argues that robotics data from teleoperation and first-person capture is expensive, slow, and narrow. Its proposed pipeline reconstructs 3D or 4D structure from internet video, aiming to produce training data for world models and action models. The founder says the company will begin batch conversion of monocular video into 360-degree dynamic scenes in 2026. The technical advantage would be scale: video already contains motion, occlusion, and physical context, while traditional 3D capture requires people and equipment. The source offers no public reconstruction benchmark, data samples, or proof that generated scenes preserve physical consistency. The distinction between a visually plausible scene and an action-useful simulator is substantial. If the conversion works, it could become a data-supply layer for Chinese spatial-intelligence companies that cannot afford massive robot fleets. If it fails on contact dynamics and hidden state, it may only produce attractive training imagery.
ChinaTalk proposes testing AI models on strategic decisions, including simulations where a Chinese model was reportedly more accepting of a Taiwan invasion than Claude.
More
- The project focuses on evaluating models in high-stakes policy and national-security settings rather than ordinary coding or knowledge benchmarks. It cites Civilization V experiments, strategic second-order reasoning failures, and different model tendencies in simulated crises. A PresidentBench example reportedly found a Chinese model relatively nonchalant about a Taiwan invasion while Claude favored preserving Taiwan's independence. These are simulated behaviors, not evidence that a model or government would act the same way in reality. The source gives no model names, prompts, sampling protocol, or reproducibility package for the Taiwan result. The project is still soliciting essays and evaluation ideas. It is relevant to China because model behavior on Taiwan-related scenarios may reflect training data, safety alignment, or policy constraints. The central lesson is that strategic evaluation needs scenario design, uncertainty measurement, and adversarial testing rather than a single ideological answer.
A Chinese research team used stable isotopes from human and animal remains in Xinjiang to reconstruct 4,000 years of diet, including shifts from millet back toward wheat and barley.
More
- The study analyzed carbon and nitrogen isotopes from remains at the Baiyang River cemetery near the northern Tianshan range, combined with radiocarbon dating. It reports a broad transition from C3 foods such as wheat, barley, meat, and dairy in the Bronze Age to greater C4 millet use during the early Iron Age through the Han period. During the Han to Tang period, millet consumption declined and C3 crops again dominated. Regional differences suggest millet was a staple in some areas but an elite status marker in others. Higher nitrogen values often appeared in burials with richer grave goods, implying a correlation between status and animal-protein intake, though not an absolute rule. The authors link changes to climate, pastoral exchange, and grain-processing technology. Isotope evidence shows consumption patterns but cannot by itself identify recipes, trade routes, or cultural identity. The study is a China-linked scientific result with broader implications for migration and agricultural exchange across Inner Asia.
Researchers’ three-bar tensegrity robot kept moving after a 5.7-meter fall, showing how compliant structures can absorb impacts in hazardous or remote environments.
More
- Tribar uses a network of elastic cables and rigid struts rather than a conventional rigid frame. Stretchable sensor tendons and onboard motion sensors estimate its changing shape and orientation after impact. The robot reportedly fell from a bridge onto asphalt and continued moving, a record height for a tensegrity robot. The design trades precision and payload for resilience, light weight and passive shock absorption. The supplied report gives no locomotion speed, energy budget or long-term reliability data, but the mechanical architecture is relevant to planetary exploration, disaster response and machines that cannot rely on careful landing surfaces.
A US-China Business Council survey says licensing delays under the Trump administration cost exports and market share, highlighting the gap between control policy and commercial outcomes.
More
- The survey was conducted in July among businesses affected by the US export-control licensing regime. Respondents reportedly cited months-long delays, lost exports, and weaker global market share. The source does not state the sample size, response rate, sector breakdown, or whether the respondents were weighted. It therefore measures business sentiment rather than proving the full economic cost of controls. The China dimension is direct: companies are assessing rules designed to restrict Chinese access to advanced technology and related supply chains. Export controls can impose costs through delayed approvals even when the targeted transaction is eventually licensed. The report does not establish whether the controls achieved security benefits that respondents could not observe. Chinese readers may see the survey as evidence that Washington is hurting its own firms; US policy coverage tends to focus on enforcement and technological denial. The practical issue is whether a licensing regime can be selective enough to preserve commercial scale while limiting strategic capability transfer. Executives should treat the dollar impact as unverified until the survey methodology and company-level data are available.
A Chinese video tool adds shot breakdown, automatic reference linking, and editable five-minute outputs around Seedance 2.5, shifting the bottleneck from model power to production workflow.
More
- Seedance 2.5 reportedly accepts up to 50 multimodal reference items, compared with nine images, three videos, and three audio clips for version 2.0. LibTV's shot-breakdown feature extracts shots, camera movement, and music from a reference film. Its AutoLink feature matches script elements to characters, scenes, props, and style references. The platform claims to extend a single generation from 30 seconds to five minutes while preserving editable shot nodes. That is more useful than simply concatenating clips because creators can revise local sections rather than regenerate a whole sequence. The source is product marketing and offers no independent quality or cost comparison. The broader lesson is that stronger video models create new orchestration work: reference management, continuity, revision, and asset provenance.
A Chinese entrepreneur's long African expansion shows why the market is attractive but slow: local infrastructure, regulation, language, and consumer differences defeat one-size-fits-all e-commerce.
More
- Liao Xuhui began sourcing African trade in 2000 and started selling VCD players from Shenzhen to Togo in 2003 with 300,000 yuan, roughly $42,000. He later expanded across more than 20 countries and built the LEADDER electronics brand. His Amanbo platform now says it operates in more than 30 African markets, connects 10,000 Chinese suppliers with over 200,000 local business buyers, and maintains local centers and warehouses. The early business faced internet speeds of about 5 kilobytes per second, frequent power cuts, poor roads, and unreliable server access. The entrepreneur says he built offline databases and local wireless points before learning that moving too early can make a company infrastructure's unpaid pioneer. He stresses that Africa is 54 separate markets, not one market, with different plugs, religions, languages, regulations, and product preferences. China's trade with Africa has doubled over the past decade, and China has remained Africa's largest trading partner for 16 years, according to the article. Those trade figures and company reach claims are not independently audited in the supplied text. The story is valuable because it explains the operating layer behind Chinese exports: local warehouses, informal distribution, market education, and adaptation matter more than simply listing products online. It also cautions against treating demographic growth as an immediate e-commerce opportunity; the market may be large and young while still requiring years of slow infrastructure and trust-building.
Yuexing Tong raised nearly 50 million yuan, roughly $7 million, to reserve e-paper capacity and add software and AI to a deliberately limited phone-back reader.
More
- The company says its devices attach magnetically to a phone and avoid open apps, feeds, and advertising. It has reportedly raised more than 100 million yuan, roughly $14 million, in total. The new capital will fund display supply, hardware, software, AI, content licensing, and overseas expansion. The source is a company profile and gives no audited sales or margin data. Its strategic tension is whether AI can improve content delivery without destroying the low-distraction proposition that attracted users.
The daughter of Ho Feng-shan objected after a Chinese property listing used his rescue of Jewish refugees to market homes, exposing the commercialization of wartime memory.
More
- Ho Feng-shan issued visas to Jews fleeing Nazi persecution while serving as China's consul-general in Vienna. A removed English-language property page used his name and rescue record to advertise homes priced at $40,000. His daughter said she was outraged by the use of the family legacy. The case concerns heritage marketing and historical memory in China rather than technology. The source provides no broader evidence about the developer, sales, or official response.
A Chinese editorial examines extreme heat as a literal public-health threat, but the supplied text contains only the headline and broken webpage code.
More
- The title frames heat as an increasingly direct threat to human life rather than a rhetorical description of discomfort.
- Because the source body is unavailable, the article’s evidence, location, and policy argument cannot be established from the supplied material.
- Its likely relevance is to China’s exposure to heat, public-health capacity, and the social meaning of climate risk, but that context is not reported here.
A Chinese morning roundup links Unitree's tightly allocated IPO, Manus's independent restart, ByteDance's new data division, and several AI developments into a broad domestic technology mood board.
More
- Unitree's reported IPO allocation gave 19,414 winning numbers, each eligible for 500 shares at 150.80 yuan, roughly $21.20, requiring about 75,400 yuan, or $10,600. The roundup repeats Manus's planned data deletion during its separation from Meta. It also reports ByteDance's new AI data and safety division and DeepSeek's infrastructure hiring. Other entries cover OpenAI departures, NVIDIA model rumors, and Tencent's WorkBuddy synchronization, but the source does not provide original reporting for all of them. The IPO allocation is a stronger China-specific signal than the promotional AI product items because it shows retail appetite for robotics exposure. The roundup's claims are sourced to multiple outlets and should be checked individually. Its value is not a single event but a snapshot of how Chinese technology media groups AI, robotics, finance, and organizational restructuring into one competitive story.
Redmi's K100 Pro Max combines a 185 Hz screen, dedicated AI graphics processor, and 9,070 mAh battery, but the launch is a routine hardware refresh.
More
- The Pro Max starts at 4,499 yuan, roughly $630, and uses Qualcomm's flagship mobile processor. Redmi claims support for more than 30 games at 185 frames per second and an average of 184.4 frames over one hour in a selected title. The phone also offers 120-watt wired charging and 50-watt wireless charging. These figures are manufacturer tests, and the source provides no independent thermal or battery-life verification. The design reflects Chinese handset competition around extreme gaming specifications rather than a major platform shift.
An Initium headline raises biosafety concerns about AI-assisted virus creation, but the supplied text is incomplete and contains no verifiable scientific details.
More
- The excerpt consists of page code and does not identify the virus, researchers, laboratory, methods, or publication. It is therefore impossible to distinguish a real experiment from a headline about a study or commentary. The China angle is present in the source context, but the item should be treated as an unverified signal only.