Alibaba Releases Zhenwu M890 AI Semiconductor Chip and Panjiu AL128 128-Card Supernode at 2026 Cloud Summit

At the 2026 Alibaba Cloud Summit held on 20th May 2026, Alibaba released a 128-card supernode server based on Pingtouge’s (T-Head) new-generation AI Semiconductor chip, the Zhenwu M890. Equipped with the ICN Switch 1.0 interconnect chip,  delivering communication latency  in 100s of nanoseconds level. This allows 128 AI semiconductor chips to operate as a single computer, supporting massive agentic inference and training of LLMs. The supernode server is now available on Alibaba Cloud’s Bailian platform and supports mainstream models including Qwen, DeepSeek, and Kimi.

In the Agent era, computing clusters must handle thousands to tens of thousands of Agents running simultaneously, with each Agent potentially initiating dozens of model calls in a single task. This creates stringent requirements for communication latency and bandwidth. The Panjiu AL128 supernode server, constructed with self-developed AI and interconnect chips, uses tight 128-card interconnection per rack. It provides P2P latency below 150 ns and single-rack bandwidth at Pb/s levels to support high-volume concurrent Agent requests.

The Zhenwu M890 adopts a self-developed parallel computing architecture and integrates 144 GB of HBM memory. Its performance is three times that of the previous-generation Zhenwu 810E, with ...