--- date: 2026-09-09 subject: "Rubidium and cesium from lithium ore | Cambricon joins PyTorch's board | JD Cloud's 100,000 GPUs" --- **A Chinese Academy of Sciences line began production** separating rubidium and cesium from lithium ore at a Jiangxi demonstration facility, giving China a domestic source of two metals it has little directly minable ore for. **Cambricon joined the PyTorch Foundation board** as a Platinum member, giving the Chinese AI chip designer a vote on whether the framework most models are trained in supports new accelerators. **JD Cloud plans a 100,000-GPU cluster** running on chips from domestic designer Moore Threads, days after the industry ministry issued a five-year plan calling for clusters at that scale. # Top Stories - **Chinese Academy of Sciences line extracts rubidium and cesium from lithium ore at production scale** — A demonstration line in Jiangxi province has reached full, stable output separating rubidium and cesium, two metals aerospace and precision-instrument makers depend on, from lithium ore. Recovering them from a stream that lithium mining already produces converts a wasted byproduct into a domestic source of both. China lacks directly minable pollucite, the conventional cesium ore, and the two metals otherwise sit as low-grade traces in lithium ore and salt-lake brine. The Chinese Academy of Sciences Institute of Process Engineering built the line with the lithium producer Ganfeng Lithium, and state media calls it the world's first continuous tower-extraction line for the pair. One new extraction tower does the work of up to 15 conventional units while consuming less reagent. A back-extraction tower built around carbon dioxide does the stripping, and line data puts overall cesium recovery above 98%. Rubidium carbonate and cesium carbonate at 99.9% purity or better are already selling. [Global Times](https://china.huanqiu.com/article/4T8ju18lYhZ) [state media] - **AI chipmaker Cambricon takes a PyTorch Foundation board seat** — Cambricon has become a Platinum member of the PyTorch Foundation, which governs the open-source framework most AI models are built and trained in. It also took a seat on the foundation's board. Support for a new accelerator is decided inside that governance, so the seat gives the Chinese chip designer a vote on the layer that determines whether its parts can run other people's models. Wang Jin, Cambricon's senior director of AI frameworks, holds the board seat, and a second engineer, Zhu Jing, joined the foundation's Technical Advisory Council. The 2022 founding board was six American companies, among them Meta, NVIDIA and Google Cloud, and Alibaba Cloud reached Platinum membership in May 2026. Cambricon says its contributed code has landed in seven of PyTorch's trunk modules, including the torch.compile compiler and the device runtime. The company says open models such as DeepSeek-V4 and GLM-5 run on its accelerators the day they ship. [BAAI](https://hub.baai.ac.cn/view/57808) - **JD Cloud plans a 100,000-GPU cluster on Moore Threads chips** — JD Cloud said it plans to build a computing cluster of 100,000 graphics processing units running on chips from the Chinese designer Moore Threads, announcing it at JD's technology conference. A commitment at that scale points Moore Threads toward the anchor buyer a training grade GPU needs to reach volume, and gives the retailer's cloud arm training capacity it can source domestically. CSDN's account calls it the first time a domestic GPU has entered a leading Chinese AI cloud's 100,000-card core cluster. Workloads named are large-model training and inference plus embodied intelligence, which pairs an AI system with a physical body such as arms, grippers and sensors. The announcement follows the Information and Communications Industry Development 15th Five-Year Plan, newly issued by the Ministry of Industry and Information Technology, which sets industrial policy for the sector. That plan calls for orderly deployment of clusters at 10,000 cards, 100,000 cards and above. [CSDN](https://blog.csdn.net/csdnnews/article/details/164741238) - **Beijing government orders 100,000-card cluster capability and frontier chip work** — The Beijing municipal government issued a 15th Five-Year Plan for the digital economy that directs the city to raise its capability to build computing clusters of 100,000 accelerator cards. The chip targets it names lean toward gains from memory, light and packaging rather than from smaller transistors. Among the frontier chip technologies it wants mastered are compute-in-memory chips, silicon-based optoelectronics and chiplet-based advanced packaging, which stitches several small dies into one package. Compute-in-memory does the arithmetic where the data already sits, instead of shuttling it to a processor and back. The plan calls for work on an AI computing instruction set built on the open RISC-V standard, and for deeper open-source development around the FlagOS system software stack. It also tells the city to speed validation of domestic compute through the Beijing-Tianjin-Hebei hub node of the national computing network. [Yicai](https://www.yicai.com/brief/103357887.html) [state-affiliated] [Yicai](https://www.yicai.com/brief/103357889.html) [state-affiliated] - **Chipmaker Yaoxinwei claims the world's first silicon carbide super-junction transistor** — Yaoxinwei Electronics unveiled Monday in Shanghai what it calls the world's first silicon carbide super-junction power transistor. A super junction is a change in the structure inside the die rather than a finer manufacturing node, so the performance it buys does not depend on more advanced lithography. The part is a MOSFET, the switching transistor at the heart of power electronics, and silicon carbide versions hold off higher voltages and run hotter than ordinary silicon ones. Measured data reported by the company put breakdown voltage at 1,635 V in the 1,200 V device class, with high-temperature current density 166% above what it calls the international top level. Hao Yue, a Chinese Academy of Sciences academician at Xidian University whose team developed the device with Yaoxinwei, called it the first of its type anywhere to complete productization validation. First customers are in data-center high-voltage power systems and AI server power supplies. [EV Industry Supply Chain](https://www.evchanye.com/2026/0909/52360.shtml) # AI & Foundation Models - **Beijing plans to meter AI model output through a token settlement system** — The Beijing municipal government issued a second 15th Five-Year Plan, this one for high-end, precision and advanced industry, calling for full-stack self-reliance in artificial intelligence. The instrument it names is a token settlement system covering the production, distribution and consumption of tokens, which would treat model output as a metered municipal resource rather than a private cloud contract. Tokens are the units of text an AI model reads and generates, and the plan calls for "token factories" to supply them. The same document orders work on AI system software stack capability and faster deployment of new domestic compute architectures. It also directs the city toward agents that execute tasks on their own, and toward the standards and protocols they would run on. [Yicai](https://www.yicai.com/brief/103357850.html) [state-affiliated] - **DeepSeek cuts prices across its cheaper Flash models from Sept. 10** — DeepSeek will cut prices across its Flash model line, the company said on its open platform. Cutting the cheaper tier's price while testing whether it can take over the expensive tier's work puts DeepSeek's competition on cost per token. Off-peak output falls to 4 yuan (about $0.60) per million tokens, and input prices fall further, with the deepest cut applying to text the model has already processed and can reuse. At peak hours Flash costs 9 yuan (about $1.30) per million output tokens against 27 yuan (about $4) for the V4 Pro model now in service. Peak rates stay at twice the off-peak rates once the new schedule takes effect. [TMTPost](https://www.tmtpost.com/8133525.html) - **Ant Group open-sources the finance benchmark behind its equity-research model** — Ant Group's Ling team has open-sourced FinFIRST, the benchmark it grades its finance-tuned model on, releasing it alongside the weights for Ling-3.0-flash-Fin ahead of the Bund Summit. Publishing the test with the model puts Ant's own definition of good investment-research work into the open, where rival labs are measured against it. FinFIRST, short for Financial Information Retrieval, Sourcing and Traceability, was built with the investment-banking team at China International Capital Corporation and more than 50 finance professionals. Its questions span mainland China, U.S. and Hong Kong markets, and 75.6% name no source at all, leaving the model to find and judge the material itself. Ant says the model was tested on FinFIRST, FinSearchComp Verified and Finance Agent, and that its overall performance beats some larger flagship models. [QbitAI](https://www.qbitai.com/2026/09/486288.html) - **HONOR puts a system-level AI agent into its next phone software** — HONOR will release MagicOS 11, its next phone operating system, on Sept. 15, with the Magic9 series as the debut hardware. Reports circulating yesterday call it the first phone operating system in commercial use to put the agent below the apps rather than inside one, so an assistant can work across several. The agent is HONOR's YOYO assistant, which pairs proactive suggestions with automatic execution and can run tasks of more than 100 steps. Those reports credit it with more than 100 categories of closed-loop service tools and more than 40 conditions that can trigger a task. By the same account, HONOR has finished development and test validation and is seeking certification before it supports the functions on the Magic9 and its RobotPhone. Beta recruitment on the Magic9 and a preview build of an "AgenticOS" open in October. [GeekPark](https://geekpark.net/news/370028) # Robotics & Autonomous Systems - **Former regulator puts China at 97% of humanoid robot manufacturing and exports** — China accounts for 97% of humanoid robot manufacturing and exports worldwide, Jin Bing said at an industry conference. That share rests on a small base, since global shipments of general-purpose humanoids ran to about 18,000 units in 2025. Jin, formerly a deputy director-general at the State Post Bureau, which regulates postal and delivery services, credited whole-machine mass production, full supply-chain support and the ability to validate robots in real deployment settings. Five Chinese companies, among them Unitree, AgiBot and UBTech, together account for 74% of global whole-machine humanoid shipments, by his account. He also named where the supply chain still falls short, listing high-end RV reducers, the precision gearboxes in robot joints, alongside tactile electronic skin and chips built specifically for embodied AI. [EV Industry Technology](https://www.evchanye.com/2026/0909/52369.shtml) - **ACE Robotics turns ordinary video into simulation-ready robot training data** — ACE Robotics and its partners released HSImul3R, which they call the first system to rebuild human interactions with a scene from ordinary video in a form a physics simulator can run. Generating that data from footage rather than a motion-capture stage attacks the shortage of physically valid manipulation examples humanoid training depends on. On the team's own benchmark, interaction stability on the easiest tier reaches roughly five times the earlier HSfM method's score. Interpenetration, where a reconstructed body passes through a solid object, falls from nearly 70% to about 23%. The team retargeted recovered motions onto a Unitree G1 humanoid, trained a whole-body controller in simulation, then ran it on a physical G1. The framework came from ACE Robotics, Nanyang Technological University's S-Lab and Shanghai AI Lab, and was accepted at ECCV 2026, a computer vision conference. [PingWest](https://www.pingwest.com/w/317202) # Military Watch - **China and Laos train drones and robot dogs to move casualties** — The China-Laos "Peace Train-2026" joint exercise in humanitarian medical rescue entered its mixed-formation training phase today, and state media says unmanned equipment was combined as a system there for the first time. Running casualty search, evacuation and supply delivery at the same time tests those machines as one chain rather than as individual pieces of equipment. The equipment includes drones flying wide-area searches, quadruped robot dogs sweeping the simulated disaster site and folding unmanned stretcher carts, among others. Participant Xu Chao said the training follows a "human-led, unmanned-equipment-enabled" approach, with the unmanned systems running from front-line casualty search through to treatment at rear medical facilities. Chen Peizheng, deputy leader of a first-tier medical facility search-and-rescue team, called the exercise a key step toward using unmanned intelligent medical rescue forces as a system. [Global Times](https://mil.huanqiu.com/article/4T8kQWWTx22) [state media] - **Army brigade corrects artillery fire with digital burst tracking in Gobi test** — A brigade of the 77th Group Army, a corps-level formation of China's army, recently ran a two-day live-fire assessment on a Gobi desert range to test whether its artillery can strike independently. The assessment tested the correction loop rather than the guns, since the first rounds fell well off target and the unit had to fix its own aim. Howitzers, mortar-howitzers and rocket artillery maneuvered as a single fire unit, with a forward command post computing firing data and pushing corrected coordinates and weather readings to the crews. The reconnaissance team then captured the burst point with digital equipment and corrected the aim, destroying a simulated enemy ammunition store on the second attempt, per the Global Times account. Officer Huang Peng said the brigade dispersed its equipment into separate fire clusters that support each other. [Global Times](https://mil.huanqiu.com/article/4T8g8j6daYg) [state media]