TMTPOST(HANGZHOU) -- Alibaba Group's semiconductor division T-Head introduced its next-generation unified training and inference AI chip, the Zhenwu M890, at the 2026 Alibaba Cloud Summit on Wednesday. The new processor delivers three times the performance of its predecessor, the Zhenwu 810E, and is configured with 144GB of high-bandwidth memory alongside an inter-chip interconnect bandwidth of 800GB/s. Engineered with native data precision support ranging from FP32 to FP4, the hardware targets large-scale model training and mixed-precision inference tasks optimized for AI agents. Alongside the silicon debut, Alibaba Cloud launched its self-developed ICN Switch 1.0 interconnect chip—enabling full-bandwidth 64-card scaling—and rolled out the Panjiu AL128 supernode server to house the accelerators for enterprise cloud clients.