Alibaba unveils Zhenwu V900 AI chip and Qwen roadmap scaling to 10 trillion parameters

Alibaba's Apsara keynote laid out a full-stack AI strategy spanning models, chips, and data centers. The centerpiece is the Zhenwu V900, an accelerator Wu positioned as China's most powerful, delivering 3x the performance of its predecessor for both high-precision training and ultra-low-precision inference, and designed to cluster up to 500,000 units. Mass production is slated for Q1 2027.

On models, Alibaba said Qwen 3.8 Max already underwent 33 iterative cycles via recursive self-improvement, and detailed a roadmap where Qwen 4.5 and Qwen 5 scale to 5–10 trillion parameters, a large jump from the current 2.4-trillion Qwen 3.8 Max. It also launched Qwen 3.8-LiveTranslate for real-time translation and continued pushing open-weight image models — Qwen-Image-2.1, which the community noted beats some closed models at just 7B parameters and grants output ownership.

Infrastructure ambitions are equally aggressive: 20+ gigawatts of data-center capacity by 2032, framed as a bid for compute self-sufficiency amid US export restrictions. Investors responded, with BABA shares up more than 3% in premarket trading.

The competitive frame is China's compute-independence push, alongside DeepSeek's bet on Huawei training chips. The V900's clusterability and the 10-trillion-parameter roadmap are direct signals that Chinese labs intend to train frontier-scale models on domestic silicon rather than NVIDIA. The caveat: performance claims are self-reported and unverified by third parties, mass production is still two quarters out, and clusters of 500,000 accelerators imply enormous power and packaging challenges that Alibaba has yet to demonstrate at scale.