AI’s Hidden Environmental Costs

Abstract

Artificial Intelligence (AI) has become a pivotal driver of the digital economy, advancing green economy objectives in renewable energy, precision agriculture, and resource-efficient operations (Morand and Ligozat and Névéol, 2024; Xu et al., 2023). Yet, this narrative obscures a paradox: the infrastructure underpinning large-scale AI models incurs significant environmental costs. Training and inference require vast electricity, producing substantial CO₂ emissions (Liu and Yin, 2024), while workloads also generate notable water consumption, directly via data-centre cooling and indirectly through water-intensive electricity generation (Jegham et al., 2025; Campbell, 2025; MIT News, 2025).
This study employs a secondary research methodology, reviewing literature, technical reports, and sustainability disclosures to evaluate AI’s operational footprint. The analysis prioritises CO₂ and energy while integrating water as a secondary dimension (Jegham et al., 2025; Murray, B. and Difelice, M., 2025). Data harmonisation produces comparable metrics (kWh/workload and CO₂e, PUE, WUE), while scenario modelling estimates potential reductions through algorithmic optimisation, renewable integration, and advanced cooling (Liu and Yin, 2024).