Alibaba unveiled its Zhenwu V900 AI processor Tuesday, with commercial release planned for early 2027, and set a target of operating more than 20 gigawatts of global data-center capacity by 2032.
Chief Executive Eddie Wu outlined the expansion at the Apsara conference in Hangzhou. Alibaba also plans to train models with between 5 trillion and 10 trillion parameters, the numerical values a model learns during training.
The V900 comes from Alibaba’s T-Head chip-design subsidiary. Mass production and commercial release are scheduled for the first quarter of 2027. Alibaba claims three times the performance of its M890 processor, introduced in May, without specifying a benchmark or test configuration in the announcement.
The new chip has 216 gigabytes of memory and 1,200 gigabytes per second of bandwidth for communication between processors, according to Alibaba. It is designed for both training AI models and running them to answer requests, a process known as inference. It supports lower-precision formats including FP8 and FP4.
For comparison, Alibaba specified 144 gigabytes of memory and 800 gigabytes per second of inter-chip bandwidth for the M890. The V900’s stated figures are 50 percent higher on both measures.
Alibaba says Qwen 4 is still in training. Its larger, 5 trillion to 10 trillion parameter targets concern the subsequent Qwen 4.5 and Qwen 5 series. Model size alone does not establish performance.
T-Head says its Zhenwu processors now serve more than 650 customers in sectors including automotive manufacturing, finance and energy. In May, it reported more than 400 external customers across 20 industries and cumulative shipments of more than 560,000 chips. That shipment total covered the Zhenwu family before the V900 was announced.
Alibaba’s May hardware launch also included a networking chip designed to connect groups of 64 accelerators and software for running workloads on its processors. The company is developing both the chips and the systems that connect them.
The Panjiu AL128 server, introduced at the same event, combines 128 AI accelerators in a single system. Alibaba designed it to handle the uneven bursts of computing requests generated by AI agents working through multistep tasks. It made the server available through its model-service platform.
Huawei has also brought forward two Ascend 960 launches, targeting the first and third quarters of 2027. The company’s roadmap includes the Ascend 970 in 2028 and the 980 in 2029.
Alibaba’s 20-gigawatt target covers the global data-center capacity operated by Alibaba Cloud. Wu set 2032 as the deadline for that expansion.
Sources: Alibaba, Alibaba, Reuters
–
By the Control Plane Editorial Team