Alibaba unveils Zhenwu M890 AI chip

- Alibaba Group unveiled the Zhenwu M890 AI chip on May 20, saying the T-Head processor is built for training, inference and AI agent workloads. - Alibaba said the M890 delivers three times the performance of the Zhenwu 810E, with 144GB memory, after shipping 560,000 Zhenwu units. - Alibaba said the next chips in the line are the V900 in third-quarter 2027 and J900 in third-quarter 2028.

Alibaba Group on May 20 unveiled the Zhenwu M890, a new artificial intelligence chip from its T-Head semiconductor unit, as the Chinese company expands its in-house hardware stack for AI training and inference. Alibaba said the chip delivers three times the performance of its predecessor, the Zhenwu 810E, and is designed for AI “agent” workloads that need large memory capacity and fast communication between processors. Reuters and CNBC reported the launch came as U.S. export curbs continue to restrict Chinese access to advanced Nvidia chips. The announcement was made at an Alibaba Cloud event in Hangzhou, where the company also outlined a multi-year chip roadmap. Alibaba said it plans to follow the M890 with the V900 in the third quarter of 2027 and the J900 in the third quarter of 2028. Bloomberg reported Alibaba aims to update the Zhenwu line on an annual cadence. (money.usnews.com) ### What exactly did Alibaba launch? T-Head introduced the Zhenwu M890 as a unified training-and-inference AI processor, according to Alibaba and TMTPost. Alibaba said the chip is intended for workloads tied to AI agents, where models need to retain long context windows and coordinate tasks in real time. (economictimes.indiatimes.com) Bloomberg and CNBC reported the M890 is configured with 144GB of GPU memory, while CNBC said interchip bandwidth is 800 GB per second. TMTPost reported the chip supports data precision formats ranging from FP32 to FP4. ### Why is Alibaba emphasizing “agent” workloads? Alibaba said the M890 is purpose-built for AI agents, a category of software systems that can carry out multi-step tasks with limited human oversight. (money.usnews.com) Reuters reported the company framed the chip as suited to the memory and communication demands of those workloads. (bloomberg.com) Bloomberg reported the hardware will handle both training and inferencing jobs and is particularly suitable for agentic tasks. That places the chip inside Alibaba’s broader effort to pair its own computing hardware with its Qwen model family and cloud services. ### How much traction does the Zhenwu line already have? (money.usnews.com) Alibaba said it has already delivered 560,000 Zhenwu units to more than 400 customers across 20 industries. CNBC reported those shipments could strengthen Alibaba’s position in China’s domestic AI processor market, where it competes with companies including Huawei and Cambricon. (bloomberg.com) Myron Xie, an analyst at SemiAnalysis, told CNBC that Alibaba-designed AI chips are “making headway with external customers” and becoming one of the more popular domestic AI hardware platforms in China. He also said Alibaba had not yet disclosed some other metrics, including compute performance, and that the announced memory and bandwidth figures still trail leading Western chipmakers. (cnbc.com) ### How does this fit the pressure created by U.S. curbs on Nvidia? Reuters reported Alibaba’s launch comes as Chinese technology companies intensify efforts to build domestic alternatives to Nvidia processors amid tightening U.S. export restrictions. CNBC said Chinese AI developers have faced long-running limits on buying cutting-edge Nvidia chips. (cnbc.com) Leonid Mironov, a portfolio manager at Gavekal, told CNBC that Nvidia is unlikely to remain a long-term supplier into all of China. Brady Wang, associate director at Counterpoint Research, said questions remain about how much manufacturing capacity Alibaba can secure at domestic foundries such as SMIC. (money.usnews.com) ### What else did Alibaba announce alongside the chip? TMTPost reported Alibaba Cloud also launched its ICN Switch 1.0 interconnect chip and the Panjiu AL128 supernode server system to support larger-scale deployments of the accelerators for enterprise cloud customers. Reuters reported Alibaba simultaneously introduced a new Qwen large language model at the event. (cnbc.com) Alibaba’s March 19 company update said T-Head’s proprietary GPUs had already entered scaled production and were being integrated with Qwen models and Alibaba Cloud infrastructure. The next scheduled milestones Alibaba named for the Zhenwu line are the V900 in the third quarter of 2027 and the J900 in the third quarter of 2028. (alibabagroup.com) (en.tmtpost.com)

Get your own daily briefing

Scout delivers personalized news, insights, and conversations tailored to your role and industry.

Download on the App Store

Shared from Scout - Be the smartest in the room.