Alibaba Unveils Zhenwu V900 AI Chip and Next-Gen Panjiu Super-Node Servers
Alibaba's chip division PingTouGe unveiled the Zhenwu V900, a unified training and inference AI chip featuring triple the compute power of its predecessor, 216GB VRAM, 1200GB/s interconnect bandwidth, and native FP8/FP4 precision support. The chip will power Alibaba's upcoming Panjiu super-node servers set to release in Q1 2027, alongside a teaser for the next-generation Zhenwu J900 planned for 2028. The release underscores China's accelerating efforts toward domestic AI hardware self-reliance and scaling compute infrastructure to handle trillion-parameter LLMs. By adopting ultra-low precision formats like FP4 and custom high-speed interconnects, Alibaba aims to significantly boost training and inference efficiency in cloud data centers. The Zhenwu V900 combines 216GB of video memory with 1200GB/s inter-chip interconnect bandwidth and hardware acceleration for FP8 and FP4 precisions. The upcoming Panjiu super-node server integrates the V900 alongside custom ICN Switch, Panmai, and Zhenyue chips, building upon Alibaba Cloud's existing M890 clusters that already support models exceeding 2 trillion parameters.
## BACKGROUND
Super-node server architectures consolidate high-performance computing chips and dedicated interconnect fabrics into a single tightly coupled system, minimizing bandwidth bottlenecks during distributed AI workload scaling. Lower-precision numerical formats like FP4 drastically cut memory footprints and energy consumption per operation, enabling large language models to run faster without requiring proportional hardware expansion.