Alibaba has introduced the Zhenwu M890 AI accelerator, developed by its T-Head semiconductor division. The new chip offers triple the performance of the previous Zhenwu 810E model. It features 144 GB of on-chip memory and an inter-chip bandwidth of 800 GB/s, specifically designed for training agentic AI systems.
The M890 is currently being integrated into the new Panjiu AL128 supernode server. When combined with the ICN Switch 1.0, the hardware allows for the connection of up to 64 accelerators into a single cluster. This development comes as Chinese firms seek domestic alternatives to Nvidia hardware amid ongoing US export restrictions.
T-Head has already distributed over 560,000 chips from the Zhenwu series to more than 400 clients, including China Telecom and FAW Group. External customers currently account for approximately 60% of the subsidiary's total production capacity.
Looking ahead, Alibaba has outlined a clear development roadmap for its hardware. The company plans to release the Zhenwu V900 in Q3 2027, which is expected to be three times more powerful than the M890. A subsequent model, the J900, is scheduled for launch in Q3 2028.
