At its 2026 Apsara Conference, Alibaba laid out a full-stack AI roadmap covering chips, cloud infrastructure, models and agents, according to Alibaba’s own account published on September 24, 2026. Its chip unit T-Head unveiled the Zhenwu V900, a training and inference processor that Alibaba says delivers three times the performance of its predecessor, the Zhenwu M890, with 216 GB of memory, 1,200 GB/s of inter-chip bandwidth and native FP8 and FP4 support. Mass production and commercial release are scheduled for the first quarter of 2027.
Alibaba also showed an upgraded supernode server combining the V900 with its own ICN switch, Panmai SmartNIC and Zhenyue SSD controller, supporting clusters of up to 500,000 cards, and said T-Head’s Zhenwu chips already serve more than 650 customers in autos, finance, energy and manufacturing. Group CEO Eddie Wu set a target for the global data center capacity operated by Alibaba Cloud to surpass 20 GW by 2032. Alibaba Cloud also plans its first regions in Turkiye, Finland and the Netherlands over the next 12 months, and says it currently runs 107 availability zones in 31 regions.
The announcement matters because China’s large AI companies cannot freely buy NVIDIA’s leading accelerators, and an in-house chip line with its own networking and storage silicon is Alibaba’s route to scaling anyway. Publishing a multi-year gigawatt target puts Alibaba in the same capacity conversation as the U.S. hyperscalers, and the same event said its next Qwen generation is already in training.
What the source does not show: the performance comparison is only against Alibaba’s own previous chip, with no benchmark against NVIDIA or other accelerators, no process node or foundry, and no disclosed spending plan behind the 20 GW goal. The V900 is not yet shipping, and a 500,000-card cluster is a supported configuration, not a deployed system.