Alibaba Cloud announced on July 23 that its Zhenwu M890 super node has been successfully adapted to Qwen3.8, and it has gone live on Alibaba Cloud’s Bailian platform to provide model inference services. Alibaba said this makes it the first super node in China to successfully run a model with more than 2 trillion parameters. Alibaba has not yet disclosed a breakdown of Qwen3.8’s architecture or parameters, but the publicly released Qwen3 version is a mixture-of-experts model with a total of 235 billion parameters.
Alibaba Cloud Zhenwu M890 Super Node Goes Live on the Bailian Platform on July 23
According to Alibaba’s announcement, the main technical specifications for the Zhenwu M890 super node are as follows: the M890 chips are equipped with 144GB of on-chip memory, and inter-chip bandwidth reaches 800GB per second. Based on the M890, the Panjiu AL128 super node server integrates 128 accelerators in a single rack, with internal bandwidth at the PB-level per second.
Alibaba Cloud said that after this deployment of Qwen3.8, the Zhenwu M890 super node became the first super node in China to successfully run a model with over 2 trillion parameters; the related model inference services have been officially launched on the Bailian platform. Alibaba has not yet publicly disclosed the specific architecture or parameter breakdown of Qwen3.8.
Publicly Listed Companies Laying Out the Super Node Track
According to responses from multiple companies on interactive platforms, several listed companies have publicly shared their super node deployment status:
Sugon (603019): ScaleX640 (640 cards/rack, with hash power density increased by 20 times) and ScaleX40 (40 GPUs/nodes, an orthogonal, no-cable, first-level interconnect architecture); the ScaleX platform as a whole is an AI supercluster with 10,000 cards
Chinesetoday Digital (000034): In February 2026, it will release the Shenzhou KunTai super node server, targeting government-enterprise and industry customers; current progress is smooth
Wantong Development (600246): Whether the company’s products will be used for the 950 super nodes depends on the downstream manufacturers’ own needs
Weiaikeji (603496): The intelligent computing super node is currently conducting product customization development and early-stage market cooperation with partner companies; based on orthogonal architecture technology, it is currently in R&D
Haitai Securities Research Note: Super Nodes as an Important Direction for Domestic Compute Infrastructure
According to Haitai Securities’ latest research note, the World Artificial Intelligence Conference showcased various super node solutions, indicating that super nodes have become an important development direction for domestic compute infrastructure; competition at the system level will drive an increase in value volume for high-speed interconnects and complete rack configurations. The research note recommends focusing on the following related areas: switching chips, switches, copper interconnects, optical modules, and complete rack enclosures.
FAQ
What are the technical specifications of Alibaba Cloud Zhenwu M890 super nodes?
According to Alibaba’s disclosure, the M890 chips are equipped with 144GB of on-chip memory, and inter-chip bandwidth reaches 800GB per second; the Panjiu AL128 super node integrates 128 accelerators in a single rack, with internal bandwidth at the PB-level per second.
What is the basis for Alibaba Cloud’s claim of the “first super node in China with over 2 trillion parameters”?
According to Alibaba’s July 23 announcement, the Zhenwu M890 super node successfully ran Qwen3.8 and provides model inference services on the Bailian platform; Alibaba says this is the first super node in China to successfully run a model with over 2 trillion parameters. The specific architecture and parameter breakdown of Qwen3.8 has not been publicly disclosed yet.
What investment recommendations does Haitai Securities’ research note on the super node track offer?
According to Haitai Securities’ latest research note, it recommends focusing on areas related to super nodes, such as switching chips, switches, copper interconnects, optical modules, and complete rack enclosures. The research note believes that super nodes have become an important development direction for domestic compute infrastructure.