Alibaba unveils Zhenwu M890 AI chip with 144GB HBM3 and a roadmap through 2028

Alibaba has introduced its Zhenwu M890 AI chip as China’s AI hardware market continues to grow under pressure from US restrictions on advanced Nvidia GPUs. The new chip is built on Alibaba’s in house Parallel Processing Unit architecture and is designed for agentic AI workloads, including training and inference.

The Zhenwu M890 targets Nvidia’s Hopper based H20 chip, with Alibaba claiming around three times the H20’s performance. The chip offers 0.6 PFLOPs of FP16 compute, which the report says puts it close to Nvidia’s A100 class performance. Alibaba also says the M890 delivers three times the compute performance of its previous generation chip.

Alibaba is building a wider AI stack around chips, interconnects, servers, and models

The Zhenwu M890 comes with 144GB of HBM3 memory, up from 96GB on the earlier Zhenwu 810E. Interconnect bandwidth has also increased from 700GB per second to 800GB per second. The chip supports FP32, FP16, FP8, and FP4 formats, which are important for modern AI workloads that need a balance of performance, memory use, and efficiency.

Alibaba is also adding supporting hardware around the chip. Its ICN Switch 1.0 interconnect chip offers 25.6Tb per second speeds with peer to peer latency below 150ns. The company is pairing this with its Yitian Arm based host CPU and Panmai networking cards inside the Panjiu AL128 Supernode Server, which can integrate 128 AI accelerators in one rack.

Alibaba says T Head has shipped about 560,000 Zhenwu AI chips so far, with more than 400 external customers across 20 industries. The roadmap also points to faster chips over the next two years. The V900 is planned for Q3 2027 with another claimed three times performance increase, while the J900 is expected in Q3 2028 with further architecture upgrades.

Alongside the chip news, Alibaba Cloud also introduced Qwen3.7 Max, a new model aimed at coding, complex reasoning, and long running agentic tasks. Alibaba says the model can run autonomous tasks for up to 35 hours and handle more than 1,000 tool calls without performance degradation. It will be made available through Alibaba’s Model Studio platform for developers and enterprises.

This is connected to the earlier Nvidia China revenue story, but it is a different update. Nvidia’s report focused on uncertainty around H200 shipments to China. Alibaba’s update shows how local Chinese companies are trying to fill that gap with their own chips, servers, interconnects, and AI models.