Alibaba launches new homegrown AI chip

Alibaba has unveiled a new artificial intelligence processor and a flagship large language model engineered specifically for autonomous AI agents. The announcements, made at the Alibaba Cloud Summit in Hangzhou on May 20, spanned silicon, software, and server systems.

The chip, named the Zhenwu M890 and designed by subsidiary T-Head, delivers three times the performance of its predecessor, the 810E, according to an official blog post by the company. With 144 GB of on-chip memory and 800 GB/s of inter-chip bandwidth, it is purpose-built for the context-heavy, communication-intensive workloads that agent-based applications require. T-Head also laid out a multi-year roadmap: the V900 is scheduled for the third quarter of 2027, followed by the J900 a year later, each targeting another significant performance jump.

Alibaba complemented the processor with the Panjiu AL128 server, a rack-scale system that houses 128 M890 accelerators and delivers petabyte-per-second internal bandwidth. It is immediately available to domestic customers through the company’s Bailian platform.

On the software side, Alibaba introduced Qwen 3.7-Max, a model tuned for extended agent operations. In an internal test by the company, the new model ran continuously for 35 hours on an unfamiliar M890 chip, executing over a thousand tool calls without human intervention, and produced an optimised computing kernel that outperformed the manufacturer’s official version by tenfold.

The announcements land amid an intensifying race among Chinese technology firms to develop homegrown AI silicon, as US export curbs restrict access to Nvidia’s most advanced processors. Huawei laid out its own domestic chip ambitions last year, and Alibaba has pledged more than 380 billion yuan ($53 billion) for cloud and AI infrastructure over the next three years.