🇨🇳Beijing E-Town (北京亦庄) launched the AI4Chip policy, targeting 3–5 globally influential leaders and 10 benchmark applications by 2028. #
Beijing Economic-Technological Development Area, also known as Beijing E-Town (北京亦庄), formally released the AI4Chip Action Plan for 2026–2028 (《“AI4Chip”人工智能赋能集成电路产业跨越式发展行动计划（2026-2028 年）》). Described by the source as China’s first dedicated AI4Chip policy, it contains five major actions and 20 priority tasks centered on coordinating artificial intelligence with the integrated-circuit industry.
The plan targets the cultivation of three to five internationally influential AI4Chip ecosystem leaders by 2028. It also calls for more than 10 benchmark applications capable of wider adoption across the industry.
At the design stage, the policy covers AI-chip architecture research, AI-supported design workflows, chip lifecycle management, coordination among design, manufacturing, packaging and testing, and intelligent-agent chip-design assistants. Its manufacturing provisions include AI-driven semiconductor production lines, intelligent packaging systems and end-to-end automated testing.
The plan also calls for an AI-native innovation system spanning design, manufacturing, packaging, testing and verification. These measures are presented as a framework for improving factory automation and embedding AI throughout the semiconductor production chain.
- Beijing E-Town (北京亦庄)
- AI4Chip
🇨🇳Eight ChiNext Compute Infrastructure ETFs opened subscriptions; Tianhong’s fund ended fundraising on its first day. #
Eight exchange-traded funds tracking the ChiNext Compute Infrastructure Index began accepting subscriptions simultaneously on September 4. Their managers were E Fund, China Asset Management, GF Fund, Fullgoal Fund, Dacheng Fund, Penghua Fund, Harvest Fund and Tianhong Fund.
The products received regulatory approval on August 28, leaving seven days between approval and the start of fundraising. According to the source, their original fundraising periods were generally six to seven trading days.
The fundraising ceilings were generally set between RMB 2 billion and RMB 5 billion per ETF. Tianhong ChiNext Compute Infrastructure ETF announced that it would close subscriptions early on the first day, making it the first of the eight products to complete a one-day fundraising period.
The source characterized these as the first ETFs focused on the ChiNext compute-infrastructure theme. It presented them as index products providing exposure to listed companies associated with the compute-infrastructure sector, rather than direct investments in physical data-center capacity.
🇺🇸LG-controlled Bear Robotics seeks ₩300–400 billion in pre-IPO funding at a target valuation of ₩2 trillion before a Nasdaq listing. #
Bear Robotics, a U.S. robotics startup controlled by LG Electronics, is preparing for a potential Nasdaq listing, according to a report citing unidentified investment-banking sources. The company is currently seeking pre-initial public offering financing rather than announcing a completed funding round or filed IPO.
Bear Robotics reportedly selected Bank of America to arrange the financing. It is approaching domestic and international private-equity institutions as prospective participants.
The company is targeting a valuation of approximately ₩2 trillion and seeking to raise between ₩300 billion and ₩400 billion. These figures are fundraising objectives reported by the source, not the terms of a closed transaction.
The proposed proceeds would be used for robotics research and development. The source did not provide a Nasdaq filing date, offering size, financing conditions, investor commitments or final valuation.
- LG-controlled Bear Robotics
- Nasdaq
🇨🇳Thinkon Semiconductor (神工股份) subsidiary Zhongjingxin (中晶芯) plans a RMB 195 million acquisition that would make Fujian Jinggong (福建精工) wholly owned. #
Thinkon Semiconductor (神工股份) disclosed that its wholly owned subsidiary Zhongjingxin (中晶芯) plans to acquire an additional 20.0002% interest in Fujian Jinggong (福建精工). The announced consideration is RMB 195 million.
Fujian Jinggong is already a controlled subsidiary of Zhongjingxin. Completion of the proposed purchase would raise Zhongjingxin’s ownership to 100%, making Fujian Jinggong a wholly owned subsidiary.
The transaction was among merger-and-acquisition updates disclosed by 13 A-share companies during the week from August 31 through September 6. The cited source describes the deal as a proposed acquisition and does not state that it has closed.
No payment schedule, valuation methodology, financing structure, closing conditions or regulatory approvals were provided in the source excerpt. The disclosed transaction therefore remains distinct from a completed transfer of ownership.
- Thinkon Semiconductor (神工股份)
- Zhongjingxin (中晶芯)
- Fujian Jinggong (福建精工)
China produced 53 billion integrated circuits in July; seven-month exports reached $216 billion, increasing 99.5%. #
Chinese industrial enterprises above the designated size produced 53 billion integrated circuits in July. The source calculated that output as approximately 1.7 billion chips per day during the month.
China’s integrated-circuit exports totaled $216 billion during the first seven months of 2026, an increase of 99.5% year over year. That total had already exceeded the full-year 2025 export value of $201.9 billion cited by the source.
The Economic Daily commentary said current capacity growth is concentrated in memory chips and mature process nodes. It also identified duplicated construction and concentrated low- and mid-range capacity alongside insufficient high-end chip capacity.
The article further said high-end chips remain affected by foreign technology restrictions and that localization remains insufficient in key materials and semiconductor equipment. These points were presented as continuing structural constraints despite the production and export growth.
QNAP showed QSW-M2130 managed switches featuring 24 2.5-gigabit Ethernet ports and six 10-gigabit Ethernet ports. #
QNAP displayed two new managed enterprise switches, the QSW-M2130 and QSW-M2130P, at Computex 2026. Both models were shown with 24 ports supporting 2.5-gigabit Ethernet, or 2.5GbE, and six ports supporting 10-gigabit Ethernet, or 10GbE.
The six 10GbE connections provide higher-speed backhaul connectivity relative to the 2.5GbE access ports. The switches support both RJ45 copper and SFP+ connectivity for that backhaul role.
The source identifies the products as managed enterprise switches rather than servers, accelerators or data-center racks. It does not provide pricing, availability, switching capacity, power consumption or shipment timing.
The QSW-M2130P was presented alongside the base QSW-M2130 as the second model in the pair. The supplied article description does not specify the operational differences between the two versions.
SEMI expects glass substrates to enter limited production around 2028, supporting approximately tenfold higher interconnect density. #
A research report cited SEMI as expecting glass substrates to enter limited production around 2028. The forecast concerns semiconductor packaging substrates, not conventional display glass or complete semiconductor devices.
Compared with organic packaging substrates, glass substrates were described as offering extremely low warpage and stronger thermal and mechanical stability. These characteristics support larger package dimensions as compute chips move toward larger and more highly integrated designs.
The report said glass could improve wiring design rules by approximately one order of magnitude, corresponding to roughly ten times the interconnect density. It linked that density improvement to placing and connecting more transistors within a single package.
SEMI’s cited forecast estimated a compound annual growth rate of approximately 67.2% for the market from 2028 through 2040. The corresponding market size was projected at approximately $13 billion in 2040, but these figures remain forecasts rather than current production or revenue.
Sidike (斯迪克) began supplying optical adhesive and anti-reflective films for selected domestic foldable-phone models. #
Sidike (斯迪克) stated on an investor-interaction platform on September 7 that it had begun supplying materials for selected domestic foldable-screen terminal models. The disclosure covers only some models and does not identify the device manufacturers.
The supplied products are optically clear adhesive, or OCA (OCA 光学胶), and anti-reflective film, or AR film (AR 减反膜). These are component materials used in display assemblies rather than complete foldable displays or finished phones.
The source confirms that supply has commenced but gives no shipment volumes, contract values, customer concentration or revenue contribution. It also does not specify whether the supply agreements are exclusive or how many models are covered.
A separate source reported that Huawei, Xiaomi and Apple were scheduled to introduce foldable flagship phones within a 72-hour period from September 7 to early September 10. That report did not identify Sidike as a supplier to any of those specific devices.
South Korea selected the proposed $22.3 billion Encinal project, a 6.3-gigawatt Texas gas plant for AI data centers, pending parliamentary approval; it plans 1.4 gigawatts of gas turbines followed by 4.9 gigawatts of combined-cycle capacity. #
South Korea selected a proposed gas-fired power project in Encinal, Texas, as the first project under its $350 billion U.S. investment plan, according to a report citing unnamed government sources. The reported final project cost was approximately $22.3 billion, compared with an initial U.S. request of $25 billion and an initial South Korean consideration of $19.8 billion.
The project is designed for 6.3 gigawatts of total generating capacity intended to supply electricity for AI data centers. This figure refers to planned power-generation capacity rather than data-center information-technology load, GPU capacity or contracted customer demand.
Development is planned in two stages: 1.4 gigawatts of gas-turbine generating units followed by 4.9 gigawatts of combined-cycle generating capacity. The source did not provide construction dates, turbine suppliers, fuel contracts or data-center customers.
The South Korean government planned to report the project to the National Assembly on Monday. A formal announcement was expected later in the month after parliamentary reporting and approval, so the source did not describe the project as finally approved or under construction.
Planned but unstarted Chinese energy-storage cell projects face paused approvals, while registered projects already under construction remain unaffected. #
Multiple energy-storage industry participants told the cited outlet that relevant authorities were surveying existing and planned production capacity. Energy-storage cells were identified as one of the main areas under review amid concerns about excess capacity.
For projects still in planning and not formally under construction, approvals and project advancement were reportedly being paused. The reported measure therefore applies to proposed capacity rather than all Chinese energy-storage cell manufacturing.
Projects that have already been registered and are under construction remain unaffected, according to the same industry sources. The report explicitly characterized the action as something other than a comprehensive halt.
The industry participants said future treatment could be adjusted according to actual market demand. The source did not identify a formal policy document, affected companies, aggregate capacity, review timetable or enforcement agency.
Malkang Dazang Solar Project (马尔康大藏光伏项目) achieved full-capacity grid connection and entered comprehensive commercial operation. #
The Malkang Dazang Solar Project (马尔康大藏光伏项目) completed full-capacity grid connection after its final photovoltaic array was connected and placed into operation. The update was attributed to PowerChina’s official account.
PowerChina served as the project’s general contractor. Following the final connection, the project formally moved into the full production and operating stage.
The source confirms grid connection of the project’s entire planned generating capacity rather than completion of only an individual array or construction package. It does not provide the project’s megawatt capacity, investment cost, location details beyond its name or expected annual generation.
No power-purchase agreement, electricity price, storage configuration or transmission arrangement was disclosed in the cited message. The reported milestone is therefore limited to full-capacity grid connection and the transition to operations.
OpenAI said GPT-6 Astra was its first model trained using 100,000 GPUs. #
A Telegram market roundup attributed to OpenAI’s president the statement that GPT-6 Astra was the first model trained on 100,000 graphics processing units, or GPUs. The same report said the model had crossed an “application threshold,” without supplying a technical definition or benchmark for that phrase.
NVIDIA chief executive Jensen Huang separately said GPT-6 Astra was trained using approximately 100,000 or more NVIDIA Grace Blackwell NVLink72 systems. His statement described the hardware platform but did not provide training duration, energy consumption, cluster location or total cost.
The source material does not clarify whether “100,000 GPUs” refers to the maximum concurrently active GPU count, the installed cluster size or the aggregate number used across the training run. It also does not provide model parameters, training-token count or measured utilization.
Huang stated that the progression from ChatGPT to o1 and then Astra occurred over four years. His additional statement that artificial general intelligence had arrived was his characterization, not a performance result established by metrics in the supplied sources.
Shanghai’s state-owned enterprises achieved full AI-application coverage, emphasizing networked compute, technical development, and operational outcomes. #
The Shanghai State-owned Assets Supervision and Administration Commission said the enterprises under its supervision had moved from selected AI pilots to full coverage. The update was presented during an AI-focused training seminar for the commission’s system.
Its second stated transition was from concentrating on compute supply toward equal emphasis on networked compute, technical research and operational results. The commission said AI systems should be matched with enterprises’ real businesses, data and processes.
A third transition is from question-and-answer assistance toward intelligent execution. Intelligent agents and digital labor are being embedded in areas including investment research and advisory services, architectural design and supply-chain scheduling.
The fourth transition is from fragmented implementation toward ecosystem collaboration. Compute, corpora, models, application scenarios and talent are being coordinated as connected resources, although the source supplied no spending, compute-capacity or application-count figures.
Insilico Medicine and MobiDrop signed an agreement integrating AI-ready single-cell multiomics data into the Pharma.AI platform. #
Insilico Medicine and MobiDrop signed a strategic cooperation agreement on September 7. Their collaboration covers the production, integration and use of high-quality, traceable and AI-ready single-cell data.
Multiomics data produced at scale by MobiDrop will be systematically added to the data warehouse of Insilico Medicine’s Pharma.AI platform. The source describes a data-infrastructure integration rather than an equity investment, acquisition or completed drug-development milestone.
The data will support training, validation and model iteration for PandaOmics, Insilico Medicine’s generative-biology platform. It will also be used for the company’s recently released multi-agent-driven Virtual Aging Cell platform, abbreviated VAC.
The source does not disclose the agreement’s financial value, data volume, exclusivity, duration or ownership terms. It also provides no performance results showing the effect of the added data on model accuracy or drug-development outcomes.
- Insilico Medicine
- MobiDrop
- Pharma.AI
China conditionally approved Aipusiwei (艾普司韦), its first AI-assisted original drug, after three-and-a-half years of development and trials. #
China’s National Medical Products Administration conditionally approved Yisitewei hydrochloride tablets (盐酸伊司特韦片), sold under the brand Aipusiwei (艾普司韦). The drug was jointly developed by Westlake University, Westlake Laboratory and Westlake Pharmaceutical (Hangzhou).
The approved indication is treatment of mild and moderate COVID-19 infections in adults. The source identifies Aipusiwei as China’s first domestically originated innovative drug approved with assistance from artificial intelligence in its research and development.
The development process took three and a half years from initial discovery through completion of clinical trials. The source does not provide the trial population, study phases, efficacy endpoints, safety results or details of the AI methods used.
The approval is conditional rather than an unrestricted approval. No post-approval study requirements, pricing, production capacity or commercialization timetable were included in the supplied source.
National Anti-Fraud AI (国家反诈 AI) launched with language, multimodal, and agent technologies for scam-risk assessment. #
The National Anti-Fraud AI application (国家反诈 AI) formally launched under the guidance of the Criminal Investigation Bureau of China’s Ministry of Public Security. It was independently developed by the Shanghai Public Security Bureau.
The application combines large-language, multimodal-model and intelligent-agent technologies. Users can enter a suspicious situation on a mobile device and receive an assessment of the associated fraud risk.
The system is designed to identify scam patterns and provide relevant examples of typical cases. It can explain fraudulent techniques through text, voice and video and offer corresponding prevention guidance.
The sources describe the product as an anti-fraud assistant that users can consult and search at any time. They do not provide model names, inference infrastructure, user counts, accuracy rates or false-positive and false-negative measurements.
The Jier (吉儿) breeding robot added embodied intelligence for flower identification, positioning, pollination, and phenotype collection. #
The intelligent breeding robot Jier (吉儿) received an embodied-intelligence version at a September 4 launch event in Beijing. The robot had previously been described as capable of autonomous navigation for crossbreeding and pollination.
The upgraded system can identify flower structures and determine their positions. It can then perform pollination and collect plant-phenotype information as part of the breeding workflow.
These functions combine visual recognition, physical positioning and manipulation in an agricultural setting. The source describes the robot’s task capabilities but provides no success rate, operating speed, supported crop list or autonomy duration.
No pricing, commercial deployment count, manufacturer shipment target or customer information was disclosed. The reported development is therefore a product-version launch rather than evidence of large-scale deployment.
