We're all going to be paying AI's Godzilla-sized power bills (www.theregister.com)

AI demand for power is exploding: modern AI datacenters already draw on the order of 100 MW each (some large sites approach or exceed 500 MW today), there are ~746 AI datacenters now, and analysts expect AI-ready capacity to grow at ~33% CAGR to 2030. Training state-of-the-art models requires thousands of high-performance GPUs/TPUs running for weeks or months to tune billions-to-trillions of parameters, and inference isn’t cheap either—chips run at 70–85°C and up to ~20% of a site’s energy can go to cooling. Big projects illustrate the scale: OpenAI has cited needs up to 16 GW (Stargate ~10 GW by 2029), Amazon/Anthropic’s Rainier starts at ~2.2 GW, and planned next-gen sites target multi-GW builds. Even optimistic renewables math is sobering—a terawatt of solar would need roughly 5 million acres (~7,800 sq mi). For the AI/ML community this matters: the industry’s growth creates acute grid strain, rising wholesale prices (Bloomberg found spikes up to 267% near heavy datacenter activity), and long lead times for new generation means timelines and costs will shape who can scale. Practically, expect heavier scrutiny on efficiency, hardware power profiles, on-site generation claims, and policy debates (grid upgrades, siting, and who pays). If utilities and regulators can’t keep pace, projects may be delayed, curtailed, or become far more expensive—costs that will ultimately ripple to users, taxpayers, and training/inference strategy decisions in research and production.