Alibaba unveils Zhenwu V900 data center GPU, claiming triple the performance of its M890 predecessor

Alibaba has unveiled the Zhenwu V900, a new data center GPU that the company says delivers three times the performance of its previous-generation Zhenwu M890, as it expands its AI chip and computing infrastructure roadmap. Chief executive Eddie Wu introduced the chip at the Apsara Conference in Hangzhou, according to a w.media report.

Alibaba plans to use the V900 for large-scale AI model training and inference, with clusters capable of supporting up to 500,000 GPUs. That figure points at hyperscale training installations rather than single-rack deployments, and signals that the chip was designed from the start around Alibaba's own frontier model workload.

The V900 is part of T-Head's Zhenwu GPU series, which Alibaba positions alongside its Yitian CPU line and its Panmai smart network interface chip. The company is also developing interconnect chips, the components that bind thousands of accelerators into a single training system.

Deployment plans for the current generation are already public: Alibaba Cloud intends to put M890-based AI Supernodes into commercial-scale service from the third quarter of 2026. Alibaba did not provide a specific target for annual AI chip shipments, leaving the pace of volume ramp-up as the open question.

Beyond silicon, Alibaba says it is optimising chips, servers, networks, models and inference software as a single system rather than as separate layers. That integration goal has been the through-line of its infrastructure strategy.

On the model side, Alibaba's Qwen team is researching recursive self-improvement and plans to train a model with between 5 trillion and 10 trillion parameters, targeting more complex and longer-running tasks. Multimodal models are a parallel area of development.

Capacity is the third leg. Alibaba Cloud plans to increase its global data center capacity to more than 20 GW by 2032, capacity the company says will support growing demand for AI model training, inference and autonomous AI agents.

Taken together, the V900, the frontier-scale model plans and the data center build-out form Alibaba's longer term strategy to build an integrated AI infrastructure stack spanning chips, models and cloud computing. What to watch next is whether the V900 reaches volume production, and how quickly Alibaba's software stack can make a homegrown accelerator a first-class target for its own model teams.