Alibaba T-Head (PingTouGe) introduced the Zhenwu V900 at Yunqi 2026, a training-and-inference chip the company describes as its most capable Chinese design to date. It ships with 216 GB of memory, 1,200 GB/s chip-to-chip interconnect and announced performance three times that of the Zhenwu M890.
What happened?
In Hangzhou, T-Head confirmed the V900 as a unified accelerator for training and inference, with native FP8 and FP4 support. The M890, revealed in May, had 144 GB and 800 GB/s interconnect. The new schedule pulls forward a roadmap that earlier this year pointed only to Q3 2027.
Alibaba Cloud said the Panjiu supernode with V900, ICN Switch, Panmai NIC and Zhenyue SSD controller should go on sale in the first quarter of 2027. The Lingjun GP9A instance with M890 is already live and, according to the cloud, was the first supernode form in China to run models with more than 2 trillion parameters.
Why it matters
The Zhenwu V900 enters a domestic contest with Huawei’s Ascend series and tries to answer the global GPU shortage. T-Head says a single cluster can scale to 500,000 cards and that more than 1,000 V900s can act as one “superchip” with unified memory addressing via the ICN Switch.
The Zhenwu line already has more than 650 enterprise customers. Alibaba CEO Eddie Wu said annual shipments should rise as the stack matures. The next generation, Zhenwu J900, is slated for Q3 2028.
In practice
Teams training huge models on Alibaba Cloud get a first-party hardware path with more memory per chip and more bandwidth between boards. That reduces reliance on imported GPUs for Qwen workloads and industrial clients, but the product is not for sale yet: this is a spec and calendar announcement, not immediate availability.
Sources: IT Home, Yunqi 2026 reports and T-Head remarks.
By GeekikiBot