A machine-learning approach to identifying drivers of CO2 emissions in China's computing power infrastructure
研究构建三阶段分析框架，识别2015-2030年中国算力基础设施CO₂排放的驱动因素，发现政策情景下排放增速从23%降至6%，算力规模仍是主导因素，但发展模式转向规模与效率并重。
Identifying the key determinants of CO₂ emissions from computing power infrastructure (CPI) is essential for designing effective decarbonization pathways. This study develops a three-stage analytical framework—emissions mechanism analysis, driver decomposition, and scenario forecasting—to examine the evolution of CO₂ emissions drivers and the mitigation potential of China's CPI over 2015–2030.Under the policy scenario for 2023–2030, the average annual growth rate of CO₂ emissions declines markedly from 23% during 2015–2023 to approximately 6%, with total emissions projected to reach about 219 Mt. by 2030. Computing scale remains the dominant driver; however, the development pattern shifts from scale-oriented expansion toward a more balanced emphasis on both scale and efficiency. Compared with 2023 levels, the contributions of power usage effectiveness (PUE) and computing energy intensity (CEI) decrease by 1.83% and 1.57%, respectively. In provinces such as Gansu, Xinjiang, Sichuan, Chongqing, Jilin, and Heilongjiang, the contribution of location-related factors declines significantly, supporting the effectiveness of China's “Eastern Data, Western Computing” initiative. As the strategy continues to advance, China is expected to partially mitigate the potential “spatial relocation–scale expansion” trade-off. • Proposes the drivers of CPI-related CO₂ emissions and their transmission mechanisms. • Integrates Random Forest model to capture nonlinear and heterogeneous CPI emission drivers. • Computing scale (S) dominates emissions, while CEI, PUE, EF, and ES gain rising importance. • By 2030, CPI emissions reach 485 Mt. under BAU but fall to 219 Mt. under policy interventions. • Evidence shows “Eastern Data, Western Computing” redistributes CPI emissions spatially.