Alibaba’s 10k‑chip AI centre
Reports say Alibaba has launched an AI data centre built around 10,000 domestically developed AI chips to rival Nvidia and to work around export controls. The move was presented as part of a broader push by Chinese cloud players to reduce reliance on foreign accelerators and to support heterogeneous hardware stacks. (archyde.com)
Alibaba and China Telecom have opened an artificial intelligence data center in Shaoguan, Guangdong, built around 10,000 of Alibaba’s own Zhenwu chips. (cnbc.com) The companies said the site will handle both model training, which is the heavy computing used to build systems, and inference, which is the work of running them after they are built. Alibaba said the cluster can support models with hundreds of billions of parameters, the internal settings that shape how a model behaves. (techrepublic.com) Alibaba built the project with China Telecom, one of China’s three state-backed telecom carriers, and reports on April 8 said it was the first commercial carrier-linked deployment of a domestic artificial intelligence chip at that scale. Quartz reported the site is in the Greater Bay Area supply chain around southern China. (qz.com) A data center like this is a warehouse of servers, networking gear and cooling systems that rents out computing power. In artificial intelligence, the most sought-after part is the accelerator chip, a processor built to do the matrix math behind training and serving models much faster than a general-purpose central processing unit. (alibabacloud.com) Chinese cloud companies have been trying to replace some Nvidia demand with local chips after Washington tightened controls on advanced artificial intelligence processors. In April 2025, Nvidia said the United States had imposed a license requirement on exports of its H20 chip to China, and the Commerce Department later said in February 2026 that H200-class exports would be reviewed case by case. (techcrunch.com) (bis.gov) Alibaba has been putting money behind that shift. The company said on February 24, 2025 that it would invest at least 380 billion yuan, about $53 billion at the time, in cloud computing and artificial intelligence infrastructure over three years. (alibabagroup.com) The chip at the center of the project comes from T-Head, Alibaba’s semiconductor design arm. Reports on the new site said Zhenwu was designed to reduce dependence on Nvidia and to give Alibaba Cloud a domestic option for both its own services and customer workloads. (msn.com) Alibaba has also been pitching a mixed-hardware approach rather than a single-chip stack. At its Apsara Conference in September 2025, Alibaba Cloud said customers would be able to build on a full stack that spans models, platforms and infrastructure, with support for different underlying hardware. (alibabagroup.com) That does not mean Alibaba has displaced Nvidia in China. The company is showing that a domestic chip can run at commercial data-center scale, while Nvidia still supplies widely used software tools and a large installed base that many developers already know. (cnbc.com) (techrepublic.com) The immediate test is whether Alibaba can turn a 10,000-chip showcase into steady cloud demand. If the Shaoguan center fills up with paying training and inference jobs, it gives Alibaba a working answer to the export-control squeeze that pushed Chinese cloud groups toward homegrown compute in the first place. (cnbc.com) (alibabagroup.com)
