Zhenwu M890 debuts as Alibaba's domestic datacenter GPU for training and inference

Alibaba Group on May 20 unveiled the Zhenwu M890, a new AI accelerator built by its semiconductor arm T-Head, as the Chinese company expanded its in-house stack for data-center training and inference. Alibaba said the chip delivers three times the performance of its previous-generation Zhenwu 810E and is aimed at high-concurrency AI workloads, including agentic applications. The company disclosed 144 GB of memory and 800 GB per second of interchip bandwidth for the new processor. Reuters, CNBC and Alibaba Cloud all reported the launch on Wednesday. ### What exactly did Alibaba launch on May 20? Alibaba’s T-Head unit introduced the Zhenwu M890 at an Alibaba Cloud event in Hangzhou on Wednesday, adding a domestic data-center accelerator to the group’s cloud and model lineup. Alibaba said the chip is designed to handle both model training and inferencing jobs, extending the Zhenwu line beyond earlier products more closely associated with inference workloads. (money.usnews.com) Reuters reported that the launch comes as Alibaba intensifies efforts to build domestic alternatives to Nvidia processors amid tighter U.S. export curbs on advanced AI chips into China. CNBC similarly framed the announcement as part of a broader expansion of China’s local AI hardware options. (alibabacloud.com) ### What are the core specifications that matter here? Alibaba said the M890 delivers three times the performance of the Zhenwu 810E, the prior generation in the line. The company also said the new chip carries 144 GB of GPU memory and 800 GB/s of interchip bandwidth, two figures that point to the scale of models and multi-chip workloads it is meant to support. (money.usnews.com) Bloomberg reported that the chip is intended for both training and inferencing, while Alibaba said it is particularly suited to agentic tasks. That combination matters because many domestic Chinese accelerators have been positioned either for narrower inference use or for more limited deployment environments. (money.usnews.com) ### Why is Alibaba emphasizing both training and inference? Alibaba Cloud said the M890 is part of a broader push to rebuild its cloud stack for the “agentic era,” and paired the chip launch with a new flagship model and cloud updates. In Alibaba’s framing, the hardware is meant to serve not only model creation but also deployment at scale inside its cloud platform. (bloomberg.com) CNTechPost and Bloomberg both reported that Alibaba presented the chip as suitable for high-concurrency workloads, including agentic AI use cases. That suggests Alibaba wants the processor to be used across the full life cycle of AI services rather than as a single-purpose accelerator. (alibabacloud.com) ### How does this fit into China’s domestic chip push? Reuters said the launch underscores China’s effort to develop local substitutes for Nvidia chips as U.S. export restrictions tighten. Alibaba is one of several Chinese technology groups trying to secure more of the AI hardware stack at home, rather than relying entirely on imported merchant GPUs. (bloomberg.com) China Daily reported that the Zhenwu series has shipped more than 560,000 units cumulatively to over 400 customers across more than 20 industries, including telecom, automotive and finance. Those figures, if sustained, would give Alibaba a larger installed base from which to place the M890 inside domestic cloud and enterprise deployments. (money.usnews.com) ### Who is this likely to pressure? Nvidia remains the dominant supplier of AI accelerators globally, but Reuters said Alibaba’s move adds to domestic alternatives being built inside China as access to top-end U.S. chips becomes more constrained. The pressure is most direct in Chinese cloud and enterprise accounts that need local supply, policy alignment or a fallback to restricted imports. (chinadaily.com.cn) Alibaba Cloud said on May 20 that the M890 launch was part of a wider product rollout at its Hangzhou event, where it also introduced a new flagship model and updated cloud services. The next concrete step will be adoption inside Alibaba’s cloud platform and by the company’s existing enterprise customer base. (alibabacloud.com) (money.usnews.com)