Alibaba has introduced Zhenwu V900 – a new artificial intelligence accelerator developed by the T-Head division for training and inference of large language models. According to the company, the chip's performance is three times that of the previous M890, released in May. The V900 is equipped with 216 GB of memory, supports computations in FP8 and FP4 formats, and the inter-chip connection bandwidth reaches 1200 GB/s. Mass production and commercial use of the accelerator are planned for the first quarter of 2027. Alibaba also showcased a server supernode based on it: such systems can be combined into clusters of up to 500,000 chips for training and deploying large-scale AI models. According to the company, solutions based on Zhenwu through Alibaba Cloud are already used by more than 650 external clients from over 20 industries.
The V900 is set to become part of Alibaba's unified computing platform, which the company is developing alongside new generations of Qwen models. Currently, the corporation is training Qwen 4, and plans to scale subsequent Qwen 4.5 and Qwen 5 to 5–10 trillion parameters. For comparison, Qwen3.8-Max contains 2.4 trillion parameters. Alibaba's CEO Eddie Wu noted that future models should better handle complex tasks with long-term planning horizons, calling it part of the movement towards artificial superintelligence. Alibaba is also exploring methods for automatic model improvement: systems independently identify weaknesses, conduct experiments, and create data for subsequent training cycles. In one experiment, Qwen3.8-Max underwent 33 consecutive cycles of automatic optimization in a month. In another test, the model spent over 60 hours on chip design and made more than 10,000 calls to EDA tools. According to Alibaba, as a result, the area of the developed block was reduced by 42% without compromising the declared performance.
Simultaneously, Alibaba is increasing the volume of available computing resources. The company expects that by 2032, the total capacity of its data centers will exceed 20 GW. Alibaba Cloud intends to begin commercial deployment of new AI supernodes this quarter. The company's management links the expansion of infrastructure with the growing demand for computing for training and operating AI models, as well as with supply chain constraints. In August, Alibaba raised $10.2 billion for AI development: the funds are directed towards computing infrastructure, proprietary chips, models, and applications. Proprietary accelerators allow Alibaba to control part of the hardware platform within its ecosystem. The development of the V900 is happening simultaneously with the increase in Qwen parameters and experiments with automated training, so further scaling of models will require a corresponding growth in computing power, memory, and network infrastructure.
Related articles
