@article{4818,
  author = {Koodichimma Chinazom Ibe-Ariwa},
  title = {Drivers of Carbon Emissions in Green AI Hyperscale Data Centers: A Statistical and Multivariate Analysis},
  journal = {Progress in Machines and Systems},
  year = {2026},
  volume = {15},
  number = {2},
  doi = {https://doi.org/10.6025/pms/2026/15/2/68-85},
  url = {https://www.dline.info/pms/fulltext/v15n2/pmsv15n2_2.pdf},
  abstract = {The rapid expansion of artificial intelligence (AI) has drastically increased the electricity demand of
hyperscale data centers, making their carbon footprint a critical sustainability concern. However, the relative
contributions of geographic, infrastructural, and operational factors to these emissions remain poorly
quantified. This study systematically identifies and ranks the primary drivers of carbon emissions in AIoriented
hyperscale data centers. Using a large scale synthetic telemetry dataset comprising over 71,000
complete observations, we employed analysis of variance (ANOVA), bivariate correlations, and multivariate
ordinary least squares (OLS) regression with heteroscedasticity robust standard errors to evaluate the
independent effects of region, energy source, cooling strategy, and workload intensity. Results reveal that
geographic region and primary energy source overwhelmingly dominate the carbon footprint, collectively
explaining nearly 90% of the adjusted variance. Transitioning from conventional grid power to renewables
reduced mean emissions by approximately 84%, while siting in the EU versus the APAC region yielded a
similar magnitude of reduction. Cooling strategy (liquid versus air) and hardware architecture (GPU versus
standard compute servers) emerged as significant secondary levers. Surprisingly, computational workload
intensity exhibited only a marginal association with total emissions once infrastructural variables were
controlled. These findings demonstrate that decarbonizing AI infrastructure requires structural interventions
such as strategic geographic siting, 24/7 renewable power purchase agreements, and liquid cooling mandates
rather than relying solely on computational efficiency or workload optimization.},
}