Sustainable AI Computing

Abstract

The computational demands of AI training and inference are growing at a rate that substantially outpaces hardware efficiency improvements, creating an expanding environmental footprint that is increasingly difficult to reconcile with corporate sustainability commitments and regulatory carbon reporting obligations. Training a frontier large language model now consumes estimated 500-1,000 tonnes of CO2-equivalent emissions; global AI inference may already exceed the annual carbon footprint of a small European country. This study evaluates eight sustainable AI computing strategies across three dimensions: energy efficiency improvement, carbon footprint reduction, and capability preservation. Strategies evaluated include model compression (INT8/INT4 quantisation, knowledge distillation, structured pruning), hardware efficiency (H100 vs. A100, TPU v5e), carbon-aware geographic and temporal scheduling, renewable energy procurement, and a proposed Carbon-Aware AI Training Framework (CATF). CATF achieves 42.4% carbon reduction at 1.8pp capability cost combining carbon-aware scheduling with INT8 quantisation and sparse training. H100 hardware achieves 2.8x energy efficiency improvement. Renewable PPA achieves 84% carbon reduction at zero capability cost - the highest-impact single action. Full sustainable scenario (H100 + 100% renewable + CATF + INT8) achieves 96.8% carbon footprint reduction at 0.8pp capability cost. A sustainable AI governance framework aligned with CSRD is proposed.