AI's Massive Energy Toll

­We’ve spoken a lot about AI, lately, and there’s a good reason for that (and it’s not because we’ve already been taken over by AI ‘round these parts). Artificial intelligence is set to upend nearly every facet of our lives – not unlike the IoT, for better or worse. But what about AI’s monumental energy drain? And how do we estimate it?

Whatever your methodology, it’s certainly not straightforward. The Verge poses the rhetorical question “How much electricity does AI consume?”, and while there’s no official answer, they agree that “machine learning consumes a lot of energy.”

And training AI programs is dramatically more energy-intensive than common, everyday tasks like George Washington singing the Macarena. One study estimates that training a language model like GPT-3 uses about 1,300 MWh of electricity, the equivalent consumption of 130 US homes annually or 1,625,000 hours on Netflix.

Of course, there’s the inevitable parallels with other digital technologies like cryptocurrency, a true energy hog which consumes about 138 Terawatt-hours annually, or around 773.61 kilowatt-hours per transaction (what the average US household uses in 26 days).

AI companies, like OpenAI (of ChatGPT fame), have been less-than-forthcoming about their creations’ energy usage, likely to offset criticism which is already in no short supply – if there’s a singular futuristic technology that’s been a lightning rod for controversy (and doom and gloom prognostications), it’s artificial intelligence, and environmental concerns don’t help matters.

Sasha Luccioni, a researcher at French-American AI firm Hugging Face, noted that “With ChatGPT we don’t know how big it is, we don’t know how many parameters the underlying model has, we don’t know where it’s running … It could be three raccoons in a trench coat because you just don’t know what’s under the hood.”

Here’s hoping AI’s – and other digital technologies’ – massive power consumption kickstarts the development of more sustainable energy sources.