Alibaba Cloud launches Lingjun Zhenwu M890 supernode for trillion-parameter AI models

Alibaba Cloud has launched its Lingjun Zhenwu M890 supernode instance, initially available in Ulanqab, Inner Mongolia. This service lets enterprise customers access a 64-GPU high-speed computing unit in the cloud without building their own data centers, supporting inference for AI models with up to a trillion parameters. It's the first supernode in China to run models with over 2 trillion parameters, and Kimi K3 and Qwen3.8-Max are already using it. The supernode supports FP8/FP4 computing and uses ICN Switch 1.0 chips to boost GPU connectivity from 16 to 64 cards, with 800GB/s inter-card speed. Training performance is three times better than the previous generation, ideal for autonomous driving and embodied AI. The Ulanqab data center uses about 90% green energy and a modular design that supports at least three chip generations. Alibaba Cloud plans to double its modular data center capacity globally and roll out this supernode in other regions soon.