At its 2026 summit, Alibaba Cloud launched the Zhenwu M890, a new training‑inference AI chip from T‑Head, alongside the ICN Switch 1.0 interconnect and the Panjiu AL128 128‑card supernode server.

The Zhenwu M890 features 144GB HBM, 800GB/s inter‑chip bandwidth, and delivers 3x the performance of its predecessor. It supports full precision from FP32 to FP4, covering both training and low‑precision inference.

The Panjiu AL128 server integrates the ICN Switch 1.0 with sub‑150ns P2P latency and Pb/s per‑rack bandwidth, allowing 128 Zhenwu M890 chips to work as one supercomputer. It is already live on Alibaba Cloud Bailian, supporting Qwen and DeepSeek.

ICgoodFind: Alibaba Cloud’s homegrown chip‑to‑supernode stack delivers a strong domestic AI compute foundation for high‑concurrency agentic workloads.