A UN-backed study calculates the enormous energy footprint of training frontier AI models and projects that hardware obsolescence will generate millions of metric tons of e-waste by 2030.