T-Head, the chip subsidiary of Alibaba, introduced the Zhenwu V900 AI chip at the 2026 Apsara Conference. The new hardware is designed for both training and inference tasks and reportedly delivers triple the performance of the previous Zhenwu M890 model.
The V900 features 216GB of memory and 1,200GB/s inter-chip bandwidth, with native support for FP8 and FP4 computing formats. Using proprietary ICN Switch interconnects, the chips can be linked into supernodes, allowing systems to scale up to 1,000 units.
Alibaba demonstrated a server integrating the V900 with ICN Switch, Panmai smart NIC, and Zhenyue SSD controller. This architecture enables clusters to scale up to 500,000 accelerators. Mass production of the V900 is scheduled for the first quarter of 2027.
The company also updated its Yitian server CPU roadmap. The Yitian 720 and 730 are expected in the third quarter of 2027, with the 750 model planned to support direct interconnects with Zhenwu AI chips to improve system-level coordination.
