At the 2026 Alibaba Cloud Summit held on May 20, T-Head, Alibaba's semiconductor division, officially unveiled the Zhenwu M890, its next-generation AI chip designed for both training and inference. This release marks a significant milestone in Alibaba Cloud's hardware infrastructure, optimized to meet the intensive computational demands of the burgeoning Agentic AI era.
In terms of technical specifications, the Zhenwu M890 features 144GB of high-capacity memory and an impressive chip-to-chip interconnect bandwidth of 800GB/s. Compared to its predecessor, the Zhenwu 810E, the M890 offers a substantial 3x performance improvement. Furthermore, it provides native support for multiple data precisions ranging from FP32 to FP4, allowing it to excel across diverse scenarios including high-precision training and ultra-low-precision inference.
To further enhance the efficiency of large-scale intelligent computing clusters, T-Head also introduced its self-developed ICN Switch 1.0. This switching chip enables 64-card full-bandwidth interconnection, drastically improving the stability and communication efficiency of massive computing clusters during complex model processing.
The debut of the Zhenwu M890 is a cornerstone of Alibaba Cloud’s newly launched "Chip-Cloud-Model-Inference" integrated technology ecosystem. As we transition into the Agentic Era, full-stack optimization from underlying hardware to top-level models has become the competitive frontier, with the Zhenwu M890 providing the essential compute foundation for this shift.
