Alibaba rolls out China’s most powerful AI chip to boost data centers

Alibaba just unveiled the Zhenwu M890, a chip its semiconductor arm T-Head says triples the performance of the previous-generation Zhenwu 810E. Announced at the Alibaba Cloud Summit in Hangzhou, the accelerator is designed for training and running the kind of autonomous, multi-step AI agents that the industry has started calling “agentic AI.”

The timing is not subtle. With US export controls continuing to block Chinese companies from accessing Nvidia’s most advanced processors, Alibaba is betting that building its own silicon is the only reliable path forward.

What the M890 actually brings to the table

The headline specs are notable. The Zhenwu M890 packs 144 GB of on-chip memory and delivers 800 GB/s of inter-chip bandwidth.

The chip is already shipping inside Alibaba’s new Panjiu AL128 supernode server. Paired with what the company calls the ICN Switch 1.0, the system can link up to 64 accelerators into congestion-free clusters.

T-Head isn’t a newcomer to the chip game, either. The subsidiary has shipped more than 560,000 Zhenwu series chips to over 400 customers, including major names like China Telecom and automaker FAW Group. Roughly 60% of T-Head’s production capacity now serves external clients.

The Nvidia-shaped elephant in the room

Alibaba’s roadmap now includes the Zhenwu V900, planned for Q3 2027, which Alibaba estimates will be three times more powerful than the M890 itself. Following that, the J900 is targeted for Q3 2028.

Alibaba Cloud’s CEO has publicly emphasized that self-developed chips improve operational efficiency and reduce supply-chain vulnerabilities.

What this means for the AI hardware landscape

The fact that external customers already account for the majority of T-Head’s output suggests a growing domestic ecosystem of companies willing to bet on Chinese-made AI silicon.

The next major milestone to watch is whether Alibaba can deliver on the V900 timeline. If a chip roughly nine times more powerful than the 810E arrives on schedule in Q3 2027, it would represent a pace of improvement that even Nvidia would respect.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.