Alibaba Unveils Full-Stack AI Push: Zhenwu V900 Chip, New Qwen Models, 20GW Cloud Plan

Alibaba used its Apsara Conference to lay out a full-stack AI strategy — silicon, models, and cloud — that positions it as China's most vertically integrated AI player. The centerpiece hardware is the Zhenwu V900 accelerator from Alibaba's T-Head division, delivering triple the performance of the prior M890 with 216GB of memory and 1,200 GB/s chip-to-chip bandwidth, and designed to scale into 500,000-chip clusters. Mass production begins in Q1 2027, and it is explicitly framed as an alternative to Nvidia processors amid export constraints.

On models, Alibaba refreshed its multimodal Qwen portfolio: Qwen-Audio 3.1 arrived alongside audio-API price cuts of up to 95%, Qwen3.8-LiveTranslate targets simultaneous interpretation, Qwen-Audio-3.1-TTS-Next promises cinematic soundscapes, and Qwen-Image 3.1 refreshes image generation. Alibaba also signaled ambitions toward Qwen 4, 4.5, and 5 models scaling from 5 trillion to 10 trillion parameters.

The infrastructure vow underpins it all: CEO Eddie Wu committed to growing Alibaba Cloud capacity past 20 gigawatts by 2032 — a build that, alongside xAI's Colossus 2 and Anthropic's Akamai deal, defines the week's compute-arms-race theme.

The aggressive audio-API price cut is the most immediately consequential move for developers, extending the price war that has characterized Chinese model providers and pressuring Western TTS/audio vendors. Combined with DeepSeek's surging revenue and the CNBC-reported global adoption of Chinese models, Alibaba's full-stack push reinforces Washington's growing unease that US benchmark leadership no longer guarantees real-world usage dominance. Watch whether the V900 can actually ship at cluster scale, and whether the 10-trillion-parameter Qwen roadmap materializes or remains aspirational.