When you ask ChatGPT a question, the energy cost of that single interaction is trivial — roughly the equivalent of powering an LED light bulb for a few seconds. But the training run that made that interaction possible? That is a different story entirely, and one the AI industry has been reluctant to discuss in detail.
The training of GPT-3, as estimated by researchers at the University of Copenhagen, consumed approximately 1,287 megawatt-hours of electricity and emitted 552 tonnes of CO2 equivalent — roughly the same as 120 petrol-powered cars driven for a year. GPT-4, with its estimated 10x parameter count, likely consumed proportionally more, though OpenAI has not published specific figures. Google’s PaLM model consumed an estimated 3.4 million litres of water during training for cooling, according to the company’s own environmental report. Google’s total water consumption rose 20% between 2021 and 2023, a period that coincided with a significant expansion of AI training infrastructure.
The carbon intensity of AI training depends heavily on where the training happens. A training run powered by Hydro-Quebec’s hydroelectric grid in Canada has a dramatically lower carbon footprint than one powered by a coal-heavy grid in parts of the United States or China. Location matters enormously, and some AI companies are beginning to site their training infrastructure strategically: Microsoft has committed to being carbon-negative by 2030, and Google aims to run on 24/7 carbon-free energy by the same year.
Inference — the ongoing cost of running trained AI models to answer user queries — may actually exceed training costs over a model’s lifetime. A 2023 analysis by SemiAnalysis estimated that ChatGPT’s inference costs alone run to approximately $700,000 per day. Multiplied across the industry, the long-term energy footprint of AI could be substantial. The counterargument, made by AI optimists, is that AI itself can be used to optimise energy grids, design more efficient chips and accelerate the development of clean energy technologies — offsetting its own environmental costs. Whether that equation nets out positive remains to be seen.