environmental impact

Training and running large models burns real electricity. Thousands of specialized chips run for weeks to train a frontier model, drawing power that, depending on the grid, comes with a carbon footprint, and the data centers that host them consume water for cooling. Once trained, a popular model answers millions of prompts a day, so the steady cost of inference can, over time, dwarf the one-time cost of training. The numbers are large enough that the field has started reporting them.

The picture is not all bleak, and it is easy to overstate in either direction. Efficiency has improved fast — smaller distilled models, better hardware, and cleaner energy all cut the cost per answer — and a single query is modest compared with many everyday activities. But the aggregate keeps climbing as usage explodes, so the honest stance is neither panic nor dismissal: the impact is real, worth measuring, and worth weighing against what the tool actually delivers.