Alibaba Revolutionizes AI with Its New Zhenwu M890 Chips

Alibaba Unveils the Zhenwu M890 for AI Agents

Alibaba has recently unveiled its latest AI processor, the Zhenwu M890, developed by its semiconductor subsidiary T-Head. This new processor is specifically designed for AI agents, offering performance three times that of its predecessor, the Zhenwu 810E. However, the real interest lies in the architecture of the M890, which is designed to maintain long contexts, coordinate with other models in real-time, and execute complex multi-step tasks with limited human intervention. This marks a significant departure from standard inference chips, which are not optimized for these memory bandwidth and inter-model communication-intensive requirements.

An Ambitious Roadmap for the Future

Alongside the launch of the M890, Alibaba presented a detailed roadmap for its future processors. Next on the list is the V900, scheduled for the third quarter of 2027, which is expected to deliver a performance boost of about three times compared to the M890. Following that, the J900 is anticipated for the third quarter of 2028. This development strategy mirrors the tick-tock product cycles used by Nvidia for its AI accelerators, allowing Alibaba to maintain a steady pace of innovation.

This approach is part of a broader trend among Chinese tech companies, such as Huawei, which also announced a similar roadmap for its Ascend chips the previous year. These initiatives underscore the determination of these companies to reduce their reliance on foreign semiconductors, even if export restrictions were to ease. For Alibaba, semiconductor development is viewed as a long-term capability-building exercise rather than merely a supply chain issue.

Massive Investments and Early Success

Alibaba's commitment to AI development is backed by a colossal investment of $53 billion in cloud and AI infrastructure over three years, the largest in its history. To date, T-Head has shipped over 560,000 Zhenwu units, used by more than 400 clients across 20 different industries, including automotive manufacturers and financial services companies. This demonstrates a significant hardware production footprint and provides Alibaba with real-world deployment data at scale ahead of the M890 launch.

The new M890 chip will be available to Chinese enterprises via Alibaba Cloud's Bailian platform, integrated into the Panjiu AL128 server system, which can house 128 M890 accelerators in a single rack.

A Robust Software Ecosystem

In addition to the hardware, Alibaba has launched Qwen 3.7-Max, the latest version of its large language model. This model is designed for advanced coding and long-duration agent tasks, capable of running continuously for up to 35 hours without performance degradation. The timing of this launch is not coincidental. By simultaneously introducing a chip and a model optimized for the same class of workload, Alibaba is establishing an integrated platform strategy.

The goal is to create a closed ecosystem where each component, from silicon to language model, reinforces the others, thereby reducing customers' dependence on external suppliers. With 500,000 chips shipped and successors planned for 2027 and 2028, T-Head is not taking any risks. At some point, building around U.S. export controls ceases to be a workaround and becomes a strategy. Alibaba seems to have crossed that line.