Alibaba unveils Zhenwu V900 AI chip, plans trillion-parameter Qwen 4

Alibaba used its Apsara Conference to lay out a full-stack AI strategy spanning silicon, models, and infrastructure. CEO Eddie Wu unveiled the Zhenwu V900, developed by Alibaba's T-Head unit and claimed to be among China's most powerful AI chips, with three times the performance of the prior Zhenwu M890, scaling to 500,000-unit clusters and entering mass production in Q1 2027.

On models, Alibaba confirmed Qwen 4 is coming and disclosed plans to train next-generation models at 5-10 trillion parameters — a major jump from Qwen3.8-Max's 2.4 trillion. It also rolled out multimodal updates including Qwen3.8-LiveTranslate and the Qwen-Audio-3.1 stack (TTS-Next for creation, ASR-Next for understanding) with 'big price cuts across the line,' per the Qwen team. On infrastructure, Alibaba committed to 20 gigawatts of global data-center capacity by 2032.

Taken together with DeepSeek's shift to Huawei training chips and $1B run-rate, this is the week's clearest signal of China pursuing full-stack AI independence — domestic silicon, frontier-scale models, and hyperscale capacity — as a challenge to US dominance and a hedge against export controls.

The caveats are standard for chip announcements: the 3x performance claim is vendor-reported without independent benchmarks, and mass production in Q1 2027 means real-world availability is quarters away. Training 5-10 trillion-parameter models is also an ambition, not a shipped result. Skeptics will watch whether the V900 delivers against NVIDIA's Blackwell-class parts in practice and whether Qwen 4 ships on the claimed timeline. Still, the breadth of the announcement makes Alibaba the most vertically integrated Chinese AI player on paper.