The Zhenwu M890 comes with 144GB of GPU memory and is designed to handle agentic AI workloads as China pushes to develop Nvidia alternatives
SOPA Images / Getty Images
Alibaba unveiled a new AI chip on Wednesday, the Zhenwu M890, which delivers three times the performance of its predecessor, the Zhenwu 810E, as the company builds out its domestic semiconductor capabilities amid U.S. export restrictions on advanced chips.
Featuring 144GB of GPU memory and interchip bandwidth of 800GB per second, the M890 was built by T-Head, Alibaba's chip design subsidiary, the company said. Among its intended uses are training and inferencing workloads, with a particular emphasis on agentic AI — a class of software capable of executing intricate, multi-step operations autonomously.
Alongside the chip, Alibaba unveiled the Panjiu AL128, a server system that packages 128 of the accelerators into a single rack. Qwen 3.7-Max, a new large language model geared toward coding and extended agentic workflows, was also introduced at the event. Alibaba noted the model sustains consistent output across sessions of as long as 35 hours.
Alibaba said the Zhenwu product line has reached a cumulative shipment total exceeding 560,000 units, with deployments spanning more than 400 enterprise clients in 20 different industries. Looking ahead, Alibaba mapped out successive generations of in-house silicon: the V900, slated for mid-2027, is projected to roughly triple the M890's performance, while a subsequent design, the J900, is scheduled to arrive in the third quarter of 2028. Alibaba also said it plans to list T-Head to tap investor demand for Chinese alternatives to Nvidia $NVDA, according to Bloomberg.
Myron Xie of SemiAnalysis, who covers AI accelerators, said the Zhenwu platform has become a popular choice among enterprise buyers in China, but cautioned that key specifications such as compute performance have not been disclosed and that the chip's memory and bandwidth numbers remain below the benchmarks set by leading Western vendors. Counterpoint Research's Brady Wang, who serves as associate director at the firm, said that while the M890 cannot match Nvidia's H200 in raw processing terms, it nonetheless functions as "a believable replacement for H200" within China's domestic market.
Alibaba launched a data center in Shaoguan, Guangdong province, earlier this year built around 10,000 Zhenwu chips in partnership with China Telecom, with plans to expand to 100,000 chips. The facility, characterized by Alibaba Cloud as a fully domestic computing deployment, reflects broader pressure on Chinese technology companies to build homegrown AI infrastructure. U.S. export restrictions have blocked Chinese firms from purchasing the most advanced Nvidia processors, and Beijing has separately issued guidance barring foreign-made AI chips from state-funded data center projects.
Alibaba has committed to spending more than 380 billion yuan ($53 billion) on cloud and AI infrastructure over three years, the company said. Eddie Wu, Alibaba's chief executive, has made building end-to-end AI capabilities a stated priority, with ambitions covering the company's own silicon, infrastructure operations, foundational models, and cloud offerings.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
