Alibaba Group has unveiled its latest artificial intelligence chip, the Zhenwu M890, marking an aggressive step forward by the Chinese technology giant to construct viable domestic alternatives to Nvidia processors amid intensifying U.S. export restrictions. Developed by Alibaba’s dedicated semiconductor design subsidiary, T-Head, the new processor boasts a threefold performance increase over its predecessor, the Zhenwu 810E. It is specifically engineered to handle the heavy memory and real-time communication demands required by AI “agents”—sophisticated software systems capable of executing complex, multi-step operations with minimal human intervention.
To demonstrate its commitment to sustained in-house silicon development, Alibaba also detailed a multi-year hardware roadmap. The company plans to follow the M890 with a successor dubbed the V900 in the third quarter of 2027—aiming for another three-fold performance leap—and a subsequent J900 chip scheduled for late 2028. This long-term strategy mirrors similar moves by domestic rivals like Huawei and is backed by Alibaba’s massive three-year, 380 billion yuan ($53 billion) commitment to cloud and AI infrastructure. The aggressive capital expenditure highlights a widespread belief across China’s tech sector that corporate demand for processing power will skyrocket as enterprises adopt agent-based AI systems.
Announced at the annual Alibaba Cloud Summit, the hardware launch also featured the introduction of the Panjiu AL128, a new server system that packages 128 of these new accelerators into a single rack. Chinese enterprise clients can access the system immediately via Alibaba Cloud’s domestic model platform, Bailian. T-Head revealed it has already shipped over 560,000 units from the Zhenwu lineup to date, serving more than 400 external clients across 20 distinct industries, including finance and automotive manufacturing. Alongside the hardware reveals, Alibaba introduced Qwen 3.7-Max, the newest iteration of its flagship large language model. Designed for advanced coding and prolonged agent workflows, the company claims the model can run continuously for up to 35 hours without any drop in performance.
Click here for more on Technology
