China Unicom Reaches 45 EFLOPS in AI Computing, Plans 17.5B Yuan Investment
China Unicom announced that its intelligent computing capacity has reached 45 EFLOPS, supported by over 1.1 million standard racks and 7 hundred-megawatt-level AI data centers (AIDCs). The telecom giant also committed to investing over 17.5 billion yuan in computing power, representing 35% of its 50 billion yuan capital expenditure plan for 2026. This massive infrastructure expansion aligns with China's national "East Data, West Computing" strategy to balance computing resources across regions. Furthermore, achieving high efficiency in cross-regional training using heterogeneous hardware helps mitigate GPU supply constraints by enabling the use of diverse chip architectures. China Unicom successfully conducted cross-regional mixed training across four locations using three types of heterogeneous computing cards over thousands of kilometers. The training efficiency for 10-billion and 100-billion parameter large language models reached 90% and 85% respectively.
## BACKGROUND
China's "East Data, West Computing" initiative aims to channel computing workloads from populous eastern provinces to western regions rich in renewable energy and land. Heterogeneous AI training involves coordinating different types of hardware accelerators (like various GPU brands or architectures), which is technically challenging due to communication latency and software compatibility. EFLOPS measures computing speed, with 1 EFLOPS representing one quintillion (10^18) floating-point operations per second.