Alibaba's New AI Chip Is Built Specifically for Autonomous Agents — and It's Already Shipping

As US export restrictions continue to limit Chinese firms' access to Nvidia hardware, Alibaba has made its most significant hardware move yet: the Zhenwu M890, a chip purpose-built for the memory demands of autonomous AI agents.

Developed by Alibaba's T-Head semiconductor subsidiary and unveiled on May 20, the M890 delivers three times the performance of its predecessor and is designed to handle the kind of complex, stateful memory operations that agentic AI workloads require — tasks that standard inference chips struggle with.

Infrastructure at Scale

Alongside the chip, Alibaba introduced the Panjiu AL128 server system, which integrates 128 M890 accelerators into a single rack. The system is already available to Chinese enterprise clients through Alibaba Cloud's Bailian platform. T-Head has shipped over 560,000 total Zhenwu units to date, spanning automotive, finance, and other sectors.

To complete the stack, Alibaba also released the Qwen 3.7-Max large language model, optimized for complex coding and agent-driven tasks — giving enterprise customers both the silicon and the model layer in one ecosystem.

A Long Roadmap, Backed by Real Money

The announcement came with a declared 380 billion yuan (~$53 billion) infrastructure investment and a multi-year chip roadmap: the V900 successor is targeted for 2027, followed by the J900 in 2028. This is not a one-off chip announcement — it's a sustained domestic AI infrastructure build-out.

Key Takeaways

• The Zhenwu M890 is optimized specifically for autonomous agent workloads — a hardware design choice that signals where enterprise AI is heading
• Alibaba's Panjiu AL128 server system (128 chips per rack) is immediately available through Alibaba Cloud
• A $53B infrastructure commitment and a multi-year chip roadmap indicate this is a long-term strategic play, not a reaction to export restrictions alone

Read the full article on Insider Monkey