--- date: 2026-08-12 subject: "Alibaba rents out a 64-card super-node | SK Hynix restarts Dalian NAND | C919 flies abroad" --- **Alibaba Cloud began selling a 64-card compute instance** in its Ulanqab region, letting enterprise customers rent tightly linked accelerators for inference on mixture-of-experts models of up to ten trillion parameters. **SK Hynix will resume building its second Dalian NAND plant**, raising capacity there by roughly 50% for mature 100-layer chips while newer production stays in Korea. **Air China launched the C919's first scheduled international route** today, a daily Beijing-Ulaanbaatar round trip that puts China's mainline jet into global ticketing systems. # Top Stories - **Alibaba Cloud puts a rentable 64-card AI compute unit on sale** — Alibaba Cloud opened first-batch sales of its Lingjun Zhenwu M890 super-node instance in the Ulanqab region. Renting the unit gives an enterprise customer 64 tightly linked accelerator cards on demand, without building a facility of its own. The company says the instance supports inference for mixture-of-experts models of up to ten trillion parameters, designs that activate only a fraction of their parameters for each token. The instance also supports FP8 and FP4 low-precision compute, number formats that cut the memory each parameter occupies. Against the prior-generation Zhenwu 810E, Alibaba Cloud says it delivers three times the training performance on intelligent-driving and embodied-intelligence workloads. Kimi K3 and Qwen3.8 Max are already served through the instance. [Sina Tech](https://finance.sina.com.cn/tech/shenji/2026-08-12/doc-inimzivu5426848.shtml) [state-affiliated] - **SK Hynix restarts its second Dalian NAND plant, lifting local capacity by half** — SK Hynix will resume construction of its long-halted second NAND flash plant in Dalian, raising its capacity there by about 50%, per Seoul Economic Daily. The added output goes to the mature end of its range. Dalian is expected to make roughly 100-layer chips built on Intel's floating-gate structure, while 300-plus-layer parts stay at Korean sites. NAND flash holds data when the power is off, the storage inside solid-state drives. The money came from Solidigm, the subsidiary SK Hynix formed out of its 2021 purchase of Intel's NAND business. Solidigm resumed investing in the plant in the first half of this year. NAND prices have risen nearly tenfold in a year, pulled up by demand for enterprise drives in AI data centers. Equipment could arrive as early as November, with a mass-production system in place in the first half of next year. [Sina Tech](https://finance.sina.com.cn/tech/digi/2026-08-12/doc-inimyaak5781264.shtml) [state-affiliated] - **Air China opens the C919's first scheduled international route** — Air China began a daily C919 round trip today between Beijing Capital International Airport and Ulaanbaatar, Mongolia, as flights CA723 and CA724. It is the first route on China's own mainline passenger jet that is open to ordinary passengers and loaded into the global ticketing system, taking the aircraft out of domestic route-proving. How much further it flies rests with the European Union Aviation Safety Agency, the EU regulator whose approval would open more than 90 other countries to the type. Information the agency has disclosed puts full certification at three to six years, with approval possible as early as 2028 and as late as 2031. C919s flown by Air China, China Eastern and China Southern have carried more than 7.5 million passengers to date. [Sina Tech](https://finance.sina.com.cn/tech/shenji/2026-08-12/doc-inimzivr6975975.shtml) [state-affiliated] [Yicai](https://www.yicai.com/brief/103315216.html) [state-affiliated] - **Tencent starts testing an AI agent inside WeChat** — Tencent said it has begun small-scale prototype testing of an AI agent called Xiaowei inside WeChat, released to a limited slice of users. Running on Tencent's in-house WeLM model, Xiaowei takes text and voice instructions to summarize a group chat, generate text or set a reminder. It can also call up mini-programs, the small apps that run inside WeChat, to order food delivery, book travel or schedule a medical appointment. Sending a message and any payment step each require manual confirmation by the user, which Tencent said guards against privacy and fund-security risks. WeChat and its international edition together carry 1.44 billion monthly active accounts. [Mydrivers](https://m.mydrivers.com/newsview/1143213.html) # Policy & Regulation - **XPeng says it helped write China's new battery and assisted-driving standards** — XPeng chairman and chief executive He Xiaopeng said that the Chinese electric vehicle maker helped formulate two new national standards, one covering power-battery safety and one covering autonomous-driving safety. The standards would be set at a maker's weakest systems rather than its best, on his account, since the aim is to confirm no weak points rather than reward strengths. Assisted-driving systems handle some road scenarios well and others poorly, he said, and owners often cannot tell which from the outside. By XPeng's own figures, the G9L, which opened for pre-sale, exceeds the national standard on chassis anti-collision performance by nearly five times. Its highway assisted driving records a few tenths of a safety takeover per 10,000 kilometers, a takeover being a moment when the system hands control back to the driver. [Sina Tech](https://finance.sina.com.cn/jjxw/2026-08-11/doc-inimyhiz0553840.shtml) [state-affiliated] # AI & Foundation Models - **Ant Group open-sources a compact model built to run on a laptop** — Ant Group's Ling team published Ling-3.0-tiny on Hugging Face yesterday, a laptop-scale addition to the Ling-3.0 family whose flash version CTD covered [August 7th](https://chinatechdaily.org/archive/china-daily/2026-08-07/). The model is built for local deployment, and activating 1.3 billion of its 7.9 billion parameters per token keeps its compute cost near that of a far smaller model. Local deployment means it runs on the machine in front of the user rather than in a data center. Ant released it in successively lower-precision weight formats, which trade a little accuracy for less memory. The company says it validated the model on NVIDIA DGX Spark, Apple Silicon MacBooks and Mac mini machines, reporting roughly 86 to 90 tokens a second on an M4 Pro MacBook. [Sina Tech](https://finance.sina.com.cn/tech/digi/2026-08-12/doc-inimxvum4250371.shtml) [state-affiliated] - **SenseTime opens a free trial of an agent model built for long tasks** — SenseTime opened SenseNova 6.8 Flash Lite Preview for trial on its token plan platform, at 1,500 free uses every five hours. The pitch is endurance rather than answer quality, and the company says the model holds its goals, constraints and key facts across tasks that run several hours and many stages. If a step fails, SenseTime says, the model rolls back and replans on its own. A main agent schedules more than ten specialist sub-agents in parallel, covering information retrieval, data analysis, fact-checking and others. SenseTime says the model works on text, images and application interfaces within one task, including driving a browser directly. The company has published a code repository on GitHub under the OpenSenseNova organization. [PingWest](https://www.pingwest.com/w/316356) - **Alibaba opens an AI agent contest scored on real e-commerce listing work** — Alibaba's Qwen platform launched Qwen AI Arena, a contest whose tasks come from live business operations rather than a fixed test set. Entrants are scored on finishing a job end to end rather than on answering prompts. The first challenge, run with Aidge, asks for an agent that turns raw product information into a complete listing package for the U.S., Korean and Brazilian marketplaces in one unattended pass. That package means copy in English, Korean and Portuguese, plus a main image, detail images and a product video. An automated engine grades the output for business fit, and the top 30 entries go to expert blind review from Aug. 31. The gold award carries $10,000 in token grants. [GeekPark](https://geekpark.net/news/368717) # Chips & Semiconductors - **Three firms plan an advanced chip-packaging center in Wuxi** — Youzu Interactive, Kangying Semiconductor and Sunrise agreed to build a 2.5D/3D advanced packaging center in the Wuxi New District. Advanced packaging joins several finished chips into one module instead of printing a single larger chip, and it decides how processors and memory sit together in an AI accelerator. The stated objective is a volume-production foundry platform for CoWoS-L, which would package other companies' chips to order rather than only its own. The partners said the project would break foreign technological monopolies in back-end packaging. Build-out runs in three phases and is to cover bumping, the solder connections a chip sits on, along with wafer-level and fan-out packaging. [Yicai](https://www.yicai.com/brief/103315350.html) [state-affiliated] - **Huawei's next flagship phone to carry a stacked Kirin chip at a higher price** — Huawei's Mate 90 series is expected to launch Sept. 23 as the first phone with a Kirin processor built on the approach Huawei calls Taodinglu, according to unnamed sources. Stacking logic in two vertical layers, rather than shrinking transistors to fit more on a flat die, raises performance without a more advanced lithography process. Counterpoint's tracking shows smartphone memory prices rose more than 80% quarter on quarter in the second quarter of 2026. Huawei executive Yu Chengdong said rising memory and component prices could force industry-wide smartphone price increases. The prior-generation Mate 80 raised specifications while cutting its price, and its 4,699 yuan ($660) entry point is not expected to survive into the new series. [Mydrivers](https://m.mydrivers.com/newsview/1143207.html) # Robotics & Autonomous Systems - **Japanese teardown of a Unitree humanoid frames Tokyo's own robot AI program** — Nikkei xTECH, a Japanese technology outlet, published video of engineers taking a Unitree G1 humanoid robot apart from head to foot, concluding that Chinese robot development has moved past being able only to demonstrate. Japan's answer is Noetra, a state-funded program bringing together dozens of companies including SoftBank, Honda and Sony to build foundation models for physical AI and robotics. The government put more than 380 billion yen ($2.5 billion) into its first year. What Japan cannot copy is deployment volume: a database counted at least 90 humanoid data-collection and training centers in China in use, planned or under construction as of April 30. Entertainment and commercial performances took the largest share of Chinese humanoid demand in 2025 at 36.8%, ahead of research and education. [Huxiu](https://www.huxiu.com/article/4882573.html) - **Robotics startup FEAGINE releases three cable-driven arms and a control model** — FEAGINE released three flexible robot arms, the A01, A02 and A03, alongside Fi0, a foundation model meant to drive different robot bodies rather than one machine. The arms bend along flexible segments pulled by cables instead of pivoting at rigid joints. The smallest, the A01, weighs 750 grams and lifts 200 grams at its tip, while the A03 reaches 50 centimeters. FEAGINE says Fi0 takes on an unfamiliar task from a single demonstration. A person performs the task once wearing the company's head-mounted rig, and the model uses that demonstration as context without updating its parameters. All three arms ship with support for ROS 1 and ROS 2, the standard open-source frameworks for robot software, plus Python and C++. [QbitAI](https://www.qbitai.com/2026/08/470874.html) # Strategic Context - **Chinese transformer makers deliver in months where Western rivals take years** — A Chinese manufacturer can build a power transformer and hand it to a customer in six to nine months, against two to three years for European and American makers, according to a Wallstreetcn analysis. The gap is in build time rather than price, and transformers are the equipment that steps voltage up and down on the electricity grid. The same gap runs through robotics, where Unitree can update a generation of humanoid robot within half a year. Both sit inside a third phase of Chinese export competitiveness that Goldman Sachs calls Going Global 3.0. The two earlier phases were cheap consumer goods and then electric vehicles, lithium batteries and photovoltaics; the third is the power, cooling and robotics capability underneath AI. [Wallstreetcn](https://wallstreetcn.com/charts/41959564)