Alibaba deploys 10,000 chips

Alibaba and China Telecom have opened an artificial intelligence data center in southern China built around 10,000 Alibaba-designed Zhenwu chips. (cnbc.com) The site is in Shaoguan, Guangdong, and Alibaba said it is meant for both training large models and running them after deployment, a step the industry calls inference. Reports on the launch said the cluster is sized for models with hundreds of billions of parameters, the internal settings a model learns from data. (cnbc.com) (techrepublic.com) Zhenwu comes from T-Head, Alibaba’s chip design arm, which has spent years building processors for cloud and artificial intelligence workloads. The new facility is the first large commercial deployment of that chip line at this scale, according to reports on the project. (msn.com 1) (msn.com 2) A data center like this is the warehouse behind modern artificial intelligence services: racks of chips do the math needed to train chatbots, image tools and recommendation systems. The key shift here is not just scale, but that Alibaba is trying to supply that computing power with domestic silicon instead of relying on Nvidia processors imported from the United States. (techrepublic.com) (cset.georgetown.edu) That push accelerated after Washington tightened export controls on advanced computing chips and chipmaking tools in October 2022 and expanded those rules again on October 17, 2023. Those controls were designed to limit China’s access to the highest-end hardware used for artificial intelligence and supercomputing. (cset.georgetown.edu) (csis.org) Alibaba has a direct business reason to build more of its own compute stack. In its fiscal year 2025 annual report, the company said Alibaba Cloud’s artificial-intelligence-related product revenue posted triple-digit growth for seven straight quarters, and public cloud revenue growth accelerated as demand for artificial intelligence services rose. (hkexnews.hk) The company is also pairing chips with its own models. Alibaba Cloud says its Qwen family includes large language and multimodal models, meaning systems that work with text, images, audio and video, and it markets them through its cloud platform. (alibabacloud.com) (github.com) Other Chinese groups are making similar moves. The South China Morning Post reported that Shenzhen activated a 10,000-card computing cluster built with Huawei Ascend 910C chips in late March, putting Alibaba’s rollout into a broader race to build large domestic pools of artificial intelligence compute. (scmp.com) Outside China, Nvidia still dominates the global market for artificial intelligence accelerators, and Alibaba has not published enough technical detail to show Zhenwu matches Nvidia’s top systems on speed, software support or energy use. What Alibaba has shown is that it can put 10,000 homegrown chips into one commercial site and use them inside its cloud business. (cnbc.com) (techrepublic.com) For now, the Shaoguan project looks less like a single product launch than a test of whether China’s biggest internet groups can keep expanding artificial intelligence services with hardware they control themselves. (cnbc.com) (hkexnews.hk)