Macro, Policy & Capital
🇨🇳CXMT sued the U.S. Department of Defense, challenging its designation as a Chinese military company. #
ChangXin Memory Technologies (CXMT, 长鑫存储) said it filed suit in the U.S. District Court for the District of Columbia on August 28. The complaint challenges the Department of Defense’s decision to place CXMT on its list of “Chinese military companies,” where it was first named in January 2025.
CXMT said it is not a military enterprise and develops and manufactures dynamic random-access memory (DRAM) products for commercial and civilian uses. The company asserted that its products are not used for military purposes and that the designation has harmed its reputation and commercial interests.
U.S. Defense Secretary Pete Hegseth was also named as a defendant, according to the cited report. The Department of Defense said that, as a matter of policy, it does not comment on pending or unresolved litigation.
🇨🇦Canada allocated over C$500 million across eight years to recruit 64 researchers, including 48 from U.S. universities. #
Canadian Industry Minister Mélanie Joly announced that Canada would recruit 64 scholars and provide more than C$500 million in funding over eight years. The funding will support their work at multiple Canadian universities.
Of the 64 researchers, 48 are moving from universities in the United States. The institutions identified in the report include Harvard University, Cornell University and the Massachusetts Institute of Technology.
The researchers’ fields include medicine, engineering and climate science. The source described the recruitment initiative in the context of deteriorating relations between Canada and the United States.
🇨🇳14 state-owned enterprises launched Data Open Month (数据开放月), offering over six billion records across more than 10 industries. #
Fourteen centrally administered and state-owned enterprises launched China’s first Data Open Month (数据开放月) during the 2026 China International Big Data Industry Expo in Guiyang. The initiative will make high-value data resources available to the public without charge.
The participating enterprises assembled more than 500 data resources, with the expected volume exceeding six billion records. The data covers more than 10 industries, including electricity, oil and gas pipelines, logistics, automobiles and agricultural machinery.
Trusted data spaces are intended to provide controlled environments for data circulation and development. The announcement also said these spaces would be combined with artificial intelligence applications to support high-value use cases and improve matching between data supply and demand.
🇨🇳Guoguang Quantum (国光量子) secured tens of millions of yuan in strategic financing for photonic integrated quantum computing. #
Beijing Zhongke Guoguang Quantum Technology Co., Ltd., referred to as Guoguang Quantum (国光量子), received a new strategic financing round worth tens of millions of yuan. The source did not disclose a more precise amount or valuation.
The round was led by a fund controlled by the management committee of the Zhengzhou Airport Economy Zone, a national-level development area. No other investors or financing terms were identified in the supplied report.
Guoguang Quantum and the investor plan to coordinate technology, capital and industrial resources around photonic integrated quantum computing. The source described the transaction as strategic financing rather than a credit facility or proposed fundraising plan.
Chips, GPUs & Systems
CXMT began mass-producing LPDDR6 memory, reaching 12,800 Mbps and capacities up to 16 GB. #
CXMT announced mass production of its independently developed LPDDR6, or sixth-generation low-power double-data-rate memory. The product is intended for mobile and other power-sensitive computing devices rather than serving as a processor or accelerator.
The disclosed peak data-transfer rate is 12,800 megabits per second, while the maximum chip capacity is 16 GB. The report attributed these specifications to CXMT’s architectural and data-transmission optimizations.
The cited report described this as the first commercial deployment of an LPDDR6 product worldwide and the first time a Chinese memory company had led mass production of a high-end memory standard. That characterization is a claim contained in the report rather than an independently supplied benchmark.
Xiaomi 18 Fold will debut CXMT’s LPDDR6 alongside Xiaomi’s Xring O3 mobile processor. #
CXMT’s first mass-produced LPDDR6 memory will be used in Xiaomi’s Xiaomi 18 Fold flagship foldable smartphone. Xiaomi founder Lei Jun said the device would be the first handset to ship with the new CXMT memory.
The handset will also use Xiaomi’s latest internally designed Xring O3 (玄戒 O3) flagship mobile processor. Lei said Xring O3 would be the first processor platform to support CXMT’s LPDDR6.
The source described the handset as a production flagship combining a domestically designed mobile processor with domestically produced next-generation memory. It did not provide the phone’s launch date, processor specifications, memory configuration or shipment volume.
- Xiaomi 18 Fold
- CXMT’s LPDDR6
- Xiaomi’s Xring O3
Cisco Secure AI Factory with NVIDIA added integrated Supermicro rack-scale systems for larger artificial-intelligence clusters. #
Cisco Secure AI Factory with NVIDIA is expanding to support integrated rack-scale systems supplied by Supermicro. The addition is specifically intended to address larger-scale artificial-intelligence cluster requirements.
The update concerns complete integrated racks rather than an announcement of a new graphics processing unit, central processing unit or standalone server. Rack-scale integration packages multiple compute and supporting components into a coordinated system for cluster deployment.
The supplied article description did not identify specific Supermicro rack models, NVIDIA accelerators, rack power levels or cluster sizes. It also did not disclose pricing, customer commitments or a deployment timetable.
University of Alberta researchers introduced PPAPlace, optimizing placements against post-route performance, power, and area rather than wirelength alone. #
University of Alberta researchers published a technical paper titled “PPAPlace: Differentiable Cross-Stage Objectives for Chip Placement Optimization.” PPA refers to the post-route performance, power and area characteristics of a chip design.
The researchers focused on macro placement, which materially affects those post-route results. They noted that most existing placement approaches use half-perimeter wirelength (HPWL) as their primary optimization objective.
The paper’s abstract states that recent benchmarking found a near-zero correlation between HPWL and post-route timing metrics. PPAPlace instead uses differentiable cross-stage objectives designed to connect placement optimization more directly with post-route PPA outcomes.
Wuhan Xinxin targets quadrupling specialty-wafer and three-dimensional-integration capacity under its 10-year Chip Light Plan (芯光计划). #
Wuhan Xinxin Semiconductor Manufacturing announced its 10-year Chip Light Plan (芯光计划) at a company development conference in Wuhan’s Optics Valley. General manager Sun Peng presented the program on August 28.
The plan uses “optical sensing and interconnection” and “three-dimensional storage and computing” as its strategic themes. Wuhan Xinxin intends to build a specialty-process wafer-manufacturing and three-dimensional-integration industrial platform centered on “3D Link–optoelectronic integration–silicon on insulator.”
The company is targeting a fourfold increase in capacity over the plan’s 10-year period. It also aims to support an upstream and downstream industrial ecosystem exceeding RMB100 billion and attract or develop more than 10,000 high-end semiconductor professionals; these figures are corporate targets, not completed results.
Power, Grid & Data Centers
Sungrow reported approximately 2 gigawatt-hours of committed AI-data-center storage orders and several additional gigawatt-hours under pursuit. #
Sungrow (阳光电源) said during an August 28 institutional investor call that its artificial-intelligence data-center energy-storage business had approximately 2 gigawatt-hours of orders on hand. These were described as existing orders rather than a broader prospective pipeline.
The company separately said it was following potential projects totaling more than 10 gigawatt-hours. That figure represents projects under pursuit and should not be treated as contracted or committed capacity.
Sungrow divides its artificial-intelligence data-center business into power-supply systems and colocated energy storage. Management forecast that growth in these businesses could become very high beginning in the next two years, but the source presented this as the company’s expectation rather than a realized result.
Sungrow delivered several solid-state transformers, targeting fourth-quarter operation of an 800-volt direct-current supply in a live compute environment. #
Sungrow said it had delivered several solid-state transformers (SSTs) to customers and expected them to enter service during the fourth quarter. An SST is power-conversion equipment, not a conventional server or computing accelerator.
The planned deployment uses an 800-volt direct-current power-supply architecture. Sungrow said the systems could become the first SST-based 800-volt direct-current solution to operate under load in a real computing environment.
The company also said that, at the time of its investor call, no SST had yet entered genuine commercial operation globally. Accordingly, the fourth-quarter date and first-deployment characterization remain company expectations rather than reported operating results.
Jerry reported over $2.5 billion in gas-turbine contracts, with large-volume equipment deliveries expected during 2027. #
Jereh Group (杰瑞股份), rendered as “Jerry” in the digest, said the contract value of its existing gas-turbine orders exceeded US$2.5 billion. The figure refers to contracted orders, while the additional projections provided by management are forecasts.
The company expects its group-wide 2026 orders for gas-turbine generator sets and supporting equipment to exceed RMB20 billion. It also said it expected the industry’s strong conditions to continue through at least 2031, which is a management outlook rather than a secured revenue figure.
Large-volume deliveries are expected in 2027 because of constraints on the availability of the main turbine units. A company representative said those main units would begin arriving in large quantities around year-end and that Jereh had previously locked in upstream orders, limiting the effect of subsequent main-unit price increases.
Zaporizhzhia Nuclear Power Plant has lacked external power for over one week; onsite diesel reserves cover approximately 10 days. #
International Atomic Energy Agency Director General Rafael Grossi said the Zaporizhzhia Nuclear Power Plant had been without external electricity for more than one week. Emergency diesel generators were supplying the electricity required for nuclear safety and security functions.
The plant reported that its onsite diesel reserves could sustain operations for approximately 10 days. That inventory was described as meeting the minimum regulatory requirement.
The International Atomic Energy Agency warned that the plant could face a complete loss of power after roughly 10 days if external electricity was not restored or additional diesel delivered. Its most recent diesel delivery occurred in March, while a subsequent shipment planned for May was delayed because of frequent military activity in the area.
- Zaporizhzhia Nuclear Power Plant
Dagang Oilfield’s storage sites injected over two billion cubic meters of gas, completing 90% of their seasonal plan. #
Dagang Oilfield’s underground gas-storage group had injected more than two billion cubic meters of natural gas during the current injection cycle. The sites are operated through Dagang Oilfield’s Tianjin gas-storage business.
The completed volume represented 90% of the overall seasonal injection plan. It was also 8% higher than the corresponding period last year.
The cycle’s injection volume set a new record for the storage group. The stored gas is intended to support winter natural-gas supply in the Beijing-Tianjin-Hebei region.
Cloud & Compute Infrastructure
China reported 2.45 million FP16 PFLOPS of intelligent-computing capacity, with 1.45 million PFLOPS monitored nationally. #
A National Data Administration official said China’s aggregate intelligent-computing capacity reached 2.45 million petaflops as of the end of July 2026. The measurement was explicitly stated at FP16, meaning 16-bit floating-point precision.
Of that total, 1.45 million petaflops had been incorporated into a national monitoring and scheduling platform. The monitored figure is therefore a subset of reported national capacity, not additional capacity.
The National Data Administration said it would continue optimizing compute deployment, increasing supply, strengthening monitoring and scheduling, and deepening coordination between computing and electricity. The agency also identified improved utilization efficiency and industry-ecosystem development as priorities.
China’s eight computing hubs and three compute-power coordination zones host over 85% of national intelligent-computing capacity. #
China’s eight national computing hubs and three computing-and-electricity coordination development zones account for more than 85% of the country’s installed intelligent-computing capacity. The figure was reported by the National Data Administration as of the end of July 2026.
The concentration metric covers intelligent-computing capacity hosted in those designated hubs and zones. It does not mean that 85% of capacity is necessarily in operation, nationally monitored or simultaneously available.
The National Data Administration separately described computing networks as interconnected with power grids, communications networks, water networks, urban utility networks and logistics networks. It said computing networks can support grid dispatch, renewable-energy consumption, data circulation, model training and infrastructure monitoring.
Baidu AI Cloud moved Model-as-a-Service and Qianfan (千帆) into its infrastructure division, separating agent products organizationally. #
Baidu AI Cloud completed a research-and-development organizational adjustment in which Model-as-a-Service and the Qianfan (千帆) brand and related products moved into the Infrastructure Business Group. The former Platform Products Business Group was renamed the Agent Business Group.
Hou Zhenyu now leads the Infrastructure Business Group, which manages artificial-intelligence infrastructure and agent infrastructure. Yin Dawei leads the Agent Business Group, which focuses on agent products including Baidu Dazi (百度搭子), Miaoda (秒哒) and Famou (伐谋).
The Data Platform Department was renamed the Data Intelligence Department and moved into the Intelligent Applications Business Group led by Ruan Yu. That group manages products including Baidu Yijian (百度一见), Hogee and Shengsuan (胜算), as well as applications for transportation, finance and other industries; all three executives report directly to Baidu executive vice president and AI Cloud head Shen Dou.
- Baidu AI Cloud
- Model-as-a-Service
- Qianfan (千帆)
AI Models, Software & Research
Northeastern University researchers introduced ROBBIN, a device-aware Rowhammer attack that injects neural-network backdoors during inference. #
Northeastern University researchers published a paper titled “ROBBIN: Rowhammer-Based Backdoor Injection during Inference.” The work presents a hardware-aware attack that incorporates device-specific memory vulnerabilities into neural-network backdoor construction.
Rowhammer induces bit flips in dynamic random-access memory by repeatedly accessing selected memory rows. ROBBIN first characterizes the bit-flip patterns of the target DRAM device rather than assuming that vulnerabilities are identical across memory hardware.
The attack then uses the characterized device information to iteratively select DRAM targets for backdoor injection during model inference. The supplied abstract identifies ROBBIN as an attack method; it does not report a product deployment or operating-system vulnerability patch.
- Northeastern University
- ROBBIN
- Rowhammer
OpenAI plans to terminate Cursor model services on November 12 following SpaceX’s acquisition of the coding platform. #
OpenAI said it had notified Cursor that it planned to wind down the companies’ contract, with a proposed termination date of November 12, 2026. The decision followed SpaceX’s acquisition of Cursor, an artificial-intelligence coding platform developed by parent company Anysphere.
OpenAI said it made the decision because, based on its previous experience with alleged contract violations by companies controlled by Elon Musk, it could not be confident that SpaceX would use OpenAI technology within the company’s terms of service. This was OpenAI’s stated rationale rather than an adjudicated finding included in the supplied sources.
A separate report said SpaceX announced the acquisition in June as a US$60 billion all-stock transaction and completed it in early August. Musk responded that he did not care and criticized OpenAI chief executive Sam Altman and president Greg Brockman.
- OpenAI
- Cursor
- November 12
- SpaceX
Kuaishou’s AI livestream assistant served over 500,000 creators within six months, averaging over 180 million daily calls. #
Kuaishou livestream-business head Zhang Yifan said the platform’s AI livestream assistant had served more than 500,000 creators during its first six months. Average daily calls to the assistant exceeded 180 million.
Kuaishou has also integrated Kling AI (可灵 AI) into AI Wanxiang (AI 万象), a personalized special-effects service. More than six million of these personalized effects were sent during the second quarter.
Separately, Kuaishou said it had allocated an additional 1.2 billion units of traffic per month to high-quality creators and 600 million units to new creators over the past year. The company reported that average public-domain exposure per new creator increased 20.6% year over year.
- Kuaishou’s AI livestream assistant
