Alibaba has unveiled the Zhenwu V900, an AI accelerator developed by its T-Head semiconductor division. The chip, presented at the Apsara conference, is designed for both neural network training and inference. It features 216 GB of high-speed memory and supports FP8 and FP4 calculations.
The V900 offers a three-fold increase in performance compared to the Zhenwu M890, which was released in May 2026. The new accelerator provides an inter-chip connection bandwidth of 1200 GB/s, an improvement over the 800 GB/s offered by its predecessor, which featured 144 GB of HBM.
Alibaba plans to integrate the V900 into its upcoming Panjiu Supernode server platform, scheduled for release in the first quarter of 2027. This system will incorporate proprietary components, including the ICN Switch, Panmai, and Zhenyue, to support large-scale AI infrastructure.
The company is also developing the next-generation Zhenwu J900 chip with a redesigned parallel computing architecture, expected to launch in March 2028. These hardware efforts are part of a broader strategy involving an investment of ¥380 billion, or approximately $52 billion, in cloud and AI over three years.
