Alibaba Announces Zhenwu V900 AI Accelerator and Expanded Cloud Plans

V900 claims and production schedule

Alibaba introduced the T-Head Zhenwu V900 accelerator on September 22 at its Apsara Conference in Hangzhou. The chip is intended for AI-model training and inference, and Dow Jones Newswires reported that mass production and commercial release are scheduled for the first quarter of 2027.

Chief Executive Eddie Wu said the V900 provides three times the performance of Alibaba’s four-month-old Zhenwu M890 and can support clusters of up to 500,000 accelerator cards. Reported V900 specifications include 216 GB of high-bandwidth memory, 1.2 TB per second of inter-chip bandwidth, and FP8 and FP4 support, according to a keynote transcript published by InfoQ China.

Alibaba has not released independent benchmark results, workload details, power consumption, software configuration, or pricing for the V900. Its claim that the product is China’s most powerful AI chip lacks a disclosed independently audited comparison with accelerators from Huawei, Cambricon, Baidu, or other domestic suppliers. A design target for 500,000 cards also does not establish that Alibaba has operated a cluster at that scale with sustained utilization and acceptable failure rates.

Cloud and model expansion

Wu said the Qwen team plans models with 5 trillion to 10 trillion parameters. Alibaba’s current Qwen3.8-Max was reported at 2.4 trillion parameters. The company has not disclosed the planned model’s architecture, active parameter count, training budget, dataset, release date, or benchmark targets. Parameter count alone does not establish model reliability, training efficiency, or commercial value.

Alibaba also targets more than 20 GW of global data-center capacity under operation by 2032, the Associated Press reported. The company did not provide a project-by-project construction timetable, an updated capital-expenditure total, or a definition specifying whether the capacity figure represents electrical capacity, IT load, or another measure.

The target would require significant electricity, construction, cooling, networking, and semiconductor supply. Alibaba previously said it would invest at least RMB380 billion in AI and cloud infrastructure over three years, in a February 2025 announcement. Its fiscal-2026 annual report said Cloud Intelligence Group external revenue grew 40% in the final fiscal quarter and AI-related products accounted for 30% of that revenue.

Export controls and timing

A commercially available V900 at Alibaba’s claimed performance and volume could give Alibaba Cloud a domestic accelerator option for training and inference workloads. Alibaba’s control of chip design, cloud infrastructure, and Qwen models could also support system-level testing and deployment. Manufacturing volume, customer availability, yields, component sourcing, cloud pricing, and the first-quarter 2027 schedule remain untested.

US controls still cover advanced computing chips, high-bandwidth memory, and semiconductor-manufacturing equipment. A January 13 policy revision permits case-by-case review of applications to export Nvidia H200, AMD MI325X, and similar products to China under specified conditions, according to the US Bureau of Industry and Security. Alibaba has not identified the V900’s fabrication process, foundry, packaging partners, or memory supplier.

The announcement came one day before Xi Jinping’s September 23-25 US state visit and two days before his September 24 meeting with President Donald Trump. The White House confirmed those dates. No evidence in the available record establishes that Alibaba coordinated the launch with the summit or timed it to influence negotiations.