Inside China Unicom’s US$390m AI Centre with Domestic Chips
-
China Unicom is launching a US$390m AI data centre in Xining, Qinghai province, powered entirely by domestically manufactured processors. The telecommunications operator, which runs China’s second-largest mobile network, announced the facility on state broadcaster CCTV this week. The data centre currently operates 23,000 AI processors and delivers 3,579 petaflops of computing power, with plans to expand to 20,000 petaflops upon completion. A petaflop measures computer processing speed equivalent to one quadrillion floating-point operations per second. The metric indicates the system’s capacity to handle complex AI model training and inference workloads that require massive parallel processing. The project showcases Beijing’s strategy to build AI infrastructure using local semiconductor technology rather than foreign chips. This approach has gained urgency following trade restrictions that limit Chinese companies’ access to advanced processors from American suppliers including Nvidia. Alibaba’s semiconductor unit T-Head provides approximately 72% of the processors deployed in the facility. The company develops custom chips for cloud computing and AI applications, competing with international suppliers in the domestic market. Other local chip suppliers contributing to the data centre include MetaX, Biren Tech and Zhonghao Xinying. Future upgrades will incorporate processors from Tecorigin, Moore Threads and Enflame, expanding the base of domestic suppliers. Biren Tech designs graphics processing units for AI training, while Moore Threads develops processors for gaming and AI applications. These companies position themselves as alternatives to foreign GPU manufacturers whose products face increasing restrictions in China. The deployment reflects a government mandate announced in August requiring publicly owned data centres to source more than 50% of their chips from domestic suppliers. This directive supports China’s semiconductor ecosystem while reducing exposure to export controls. Furthermore, China’s internet regulator has reportedly banned technology companies from purchasing Nvidia processors, which dominate global AI systems. Nvidia’s H100 and A100 chips power most large language models (LLMs) developed by companies including OpenAI and Google, making domestic alternatives crucial for Chinese AI development. Alibaba announced earlier this month that it is developing a new AI inference chip, with testing already underway. The processor has been manufactured domestically as part of Beijing’s plan to strengthen AI-specific chip design and fabrication capabilities. AI inference chips process trained models to generate responses or predictions, requiring different optimisations than training processors. These chips must balance speed, power efficiency and cost to serve AI applications at scale. The development fits into China’s strategy to build complete AI infrastructure using local technology. Companies including Baidu, Tencent and ByteDance have invested in custom chip programmes to reduce dependence on foreign suppliers. As the leading chip provider for the China Unicom facility, T-Head’s role underlines connections between cloud operators, telecommunications firms and the emerging domestic semiconductor ecosystem. These partnerships create distribution channels for local chip companies while providing technology firms with alternative supply sources. The data centre’s target capacity of 20,000 petaflops would rank among the world’s most powerful computing facilities – though specifications do not clarify whether performance figures use double-precision floating-point standard common in high-performance computing or lower precision standards often used for AI workloads. The investment demonstrates the scale of resources directed toward AI infrastructure in China as companies build domestic capabilities amid international trade tensions.