Alibaba unveils Zhenwu V900 AI chip and plans up to 10-trillion parameter model

Zhenwu V900 processor and semiconductor development

Alibaba Group presented its latest artificial intelligence accelerator, the Zhenwu V900, on Tuesday during the annual Apsara conference held in Hangzhou. Developed by T-Head, Alibaba's dedicated semiconductor division, the processor delivers three times the computing performance of its predecessor, the Zhenwu M890. Alibaba described the hardware as the most powerful AI chip produced in China. The company released details of the hardware through its corporate portal Alizila, framing the launch around a strategy focused on three pillars: artificial intelligence models, semiconductor chips, and cloud computing services. Chief Executive Eddie Wu introduced the chip as part of a push to build full-stack control over both the hardware and software operating across Alibaba's digital ecosystem.

Scaled clusters and domestic chip production

The Zhenwu V900 is engineered to operate within large-scale server installations tailored for demanding computational workloads. Eddie Wu stated that an infrastructure cluster utilizing the V900 architecture can support up to 500,000 accelerator cards to handle both advanced model training and real-time inference tasks. Alibaba expects a sharp increase in its annual delivery volumes of proprietary artificial intelligence chips as data centres deploy the hardware. The semiconductor push aligns with a broader financial commitment reported by the Wall Street Journal, in which Alibaba pledged to invest more than $53 billion over three years to integrate artificial intelligence capabilities into its traditional e-commerce business. On the same day as the Apsara opening, rival technology company Tencent Holdings launched its Hy Image 3.5 Preview generation system in an effort to close the gap with ByteDance and Alibaba.

Expanding Qwen models and multi-trillion parameters

In addition to hardware, Alibaba detailed expansion plans for its flagship Qwen artificial intelligence model series, which anchors the company's cloud computing offerings. The Qwen engineering team plans to train a next-generation foundation model with a scale ranging between 5 trillion and 10 trillion parameters. This initiative builds directly on the previous release in August 2026 of the Qwen3.8-Max model, which operates with 2.4 trillion parameters. The expanded parameter count represents a fourfold scaling from the August release at the upper end of the projected range. Alibaba is positioning these scaled models to support complex enterprise applications and automated tasks across its expanding commercial network.

Cloud capacity targets and semiconductor export limits

To sustain long-term computing requirements, Alibaba announced a target for Alibaba Cloud's global data centre capacity to exceed 20 gigawatts by 2032. Eddie Wu described current artificial intelligence demand as exceptionally strong, noting that medium- and long-term industry demand far exceeds Alibaba's existing supply capacity. The expansion takes place as Chinese technology companies seek alternatives to foreign suppliers such as Nvidia, whose advanced semiconductors remain subject to United States export controls on national security grounds. These regulatory restrictions have established computing power as the central operational constraint for Chinese artificial intelligence developers. Liang Wenfeng, the founder of Chinese AI firm DeepSeek, observed in recent months that computational resources, rather than engineering talent or core technology, constitute the primary gap separating Chinese developers from competitors in the United States.