Alibaba's Zhenwu V900 chip targets 10-trillion-parameter AI models with 500,000-chip cluster
Alibaba unveiled its Zhenwu V900 AI accelerator at the Apsara Conference, claiming it is China's most powerful AI chip with triple the performance of its predecessor. The chip features 216GB of memory and 1,200GB/s inter-chip bandwidth, and can scale to a supercluster of 500,000 cards. Mass production is slated for the first quarter of 2027, supporting future Qwen models with up to 10 trillion parameters.
The Zhenwu V900's launch timeline was accelerated from an earlier third-quarter 2027 roadmap to the first quarter of 2027, reflecting Alibaba's push to close the gap with domestic rival Huawei, whose Ascend 960PR is slated for the third quarter of 2027. Alibaba's chip arm T-Head has shipped over 560,000 Zhenwu processors to more than 650 external customers, with annual refreshes planned.
The chip's specifications omit several key metrics, including FLOPS, process node, and memory bandwidth, leaving its "most powerful" claim without an absolute baseline. Alibaba's broader infrastructure ambitions include exceeding 20 gigawatts of data center capacity by 2032, though supply chain constraints currently limit compute scaling. The V900 is designed to support Qwen models scaling from the current 2.4-trillion-parameter Qwen3.8-Max toward 10 trillion parameters.
This chip could reshape China's AI infrastructure landscape by reducing reliance on foreign accelerators, potentially accelerating domestic model development and lowering costs for Chinese enterprises. However, the lack of disclosed performance metrics means real-world impact remains uncertain until deployment. If the 500,000-card cluster materializes, it may enable frontier-scale research within China, affecting global AI competition dynamics. Conversely, supply chain bottlenecks and the 2027 timeline could delay benefits, leaving smaller firms dependent on existing hardware while larger players consolidate advantages.