Alibaba has unveiled a new artificial intelligence accelerator, the Zhenwu V900, in what amounts to China's clearest statement yet that it intends to build frontier AI systems on domestic hardware rather than American chips. The chip was announced by Alibaba chief executive Eddie Wu at the company's annual Apsara Conference in Hangzhou. Developed by T-Head, Alibaba's semiconductor unit, the V900 is designed for both AI training and inference. Alibaba says it delivers three times the performance of its predecessor, the Zhenwu M890, which was released in May.
According to T-Head, the V900 carries 216 gigabytes of memory and offers 1,200 gigabytes per second of inter-chip bandwidth, with native support for low-precision computing formats including FP8 and FP4. The specifications are the ones that matter most for training large models: more memory reduces the overhead of splitting models across chips, while lower-precision formats improve efficiency and cut inference costs.

Clusters of up to 500,000 chips

The most striking claim concerns scale. Alibaba says a computing cluster built around the V900 can link up to 500,000 accelerator cards for training and inference tasks involving advanced AI models. The company's upgraded supernode server combines the accelerator with its ICN Switch interconnect, Panmai SmartNIC networking, and Zhenyue SSD storage controller to reach that scale.
Numbers of that order decide whether a country can train frontier models at all, and the claim is aimed directly at the constraint that US export controls were meant to impose on China's AI industry. DeepSeek founder Liang Wenfeng has said in recent months that the key gap between China and the United States lies not in talent or technology but in computing resources.
Mass production and commercial release of the V900 are scheduled for the first quarter of 2027. Alibaba says its existing Zhenwu chips are already in use with more than 650 customers across automotive, finance, energy, and manufacturing sectors.

20 gigawatts and a 10-trillion-parameter roadmap

The chip arrived alongside an infrastructure target: Alibaba plans to expand the total capacity of data centers operated by Alibaba Cloud to more than 20 gigawatts by 2032. The investment is intended to meet what Wu described as exceptionally strong demand for AI services, though he acknowledged that medium- and long-term demand far exceeds the company's current supply capacity.
The Qwen model roadmap was equally ambitious. Alibaba disclosed that its next-generation Qwen 4 model is currently in training, with current flagship systems holding around 2.4 trillion parameters. Future Qwen 4.5 and Qwen 5 models are projected to scale to between 5 trillion and 10 trillion parameters — a multiple of today's size and comparable in ambition to the largest models anywhere in the world.

China's self-contained AI stack takes shape

The announcements came shortly after Alibaba's domestic rival Huawei unveiled new AI infrastructure that it said can scale to as many as one million processors. Taken together, the moves signal that China's technology giants are racing to reduce dependence on foreign suppliers such as Nvidia, on whose advanced semiconductors Washington has imposed export restrictions citing national security concerns.
The point of announcing the chip and the model roadmap together, analysts note, is to describe a self-contained stack: Chinese hardware training Chinese models at frontier scale. Independent reporting from Reuters and the Associated Press covered the launch this week, making it one of the major AI hardware stories of the period.
For the global AI market, Alibaba's push adds a new front to the infrastructure race — one in which the competition is no longer only about models and cloud credits, but about who controls the silicon and the power plants that run them.