China's AI Power: Rent a Supernode

Alibaba Cloud has made its Lingjun Zhenwu M890 supernode instance commercially available, allowing businesses to rent powerful 64-card computing units. This offering is notable as it exclusively uses in-house designed silicon, developed by Alibaba's T-Head chip design unit. The supernode architecture tightly interconnects accelerators, treating them as a single large chip to efficiently handle massive AI models.

The Lingjun Zhenwu M890 is specifically designed to address limitations imposed by state restrictions and sanctions on accessing advanced foreign chips. It features the Zhenwu M890 processor with 144GB of HBM memory and uses an in-house ICNSwitch 1.0 chip for high inter-chip bandwidth. This configuration is reported to be capable of handling AI models with up to 10 trillion parameters, powering commercial services for notable models like Kimi K3 and Qwen3.8-Max.

Alibaba has outlined a future roadmap, with the Zhenwu V900 slated for the third quarter of 2027 and the Zhenwu J900 for the third quarter of 2028, promising further performance increases. Currently, the service is exclusively available to users within the Ulanqab region of Inner Mongolia, China, underscoring the nation's drive for AI self-sufficiency. Pricing and power consumption details have not been publicly disclosed.