Power Hungry Machines

Modern AI training and inference consume staggering amounts of electricity and freshwater, with GPT-4's training estimated at ~21,660 metric tons of CO₂—a 40-fold increase over GPT-3. Data centres housing thousands of NVIDIA H100 GPUs require massive cooling infrastructure that evaporates millions of litres of freshwater. As generative AI embeds deeper into daily life, the tension between computational capability and environmental sustainability becomes a critical policy and engineering challenge.