Three big things we still don’t know about AI’s energy burden

The problem with finding that number, as we explain in our piece published in May, was that AI companies are the only ones who have it. We pestered Google, OpenAI, and Microsoft, but each company refused to provide its figure. Researchers we spoke to who study AI’s impact on energy grids compared it to trying to measure the fuel efficiency of a car without ever being able to drive it, making guesses based on rumors of its engine size and what it sounds like going down the highway.

But then this summer, after we published, a strange thing started to happen. In June, OpenAI’s Sam Altman wrote that an average ChatGPT query uses 0.34 watt-hours of energy. In July, the French AI startup Mistral didn’t publish a number directly but released an estimate of the emissions generated. In August, Google revealed that answering a question to Gemini uses about 0.24 watt-hours of energy. The figures from Google and OpenAI were similar to what Casey and I estimated for medium-size AI models. 

So with this newfound transparency, is our job complete? Did we finally harpoon our white whale, and if so, what happens next for people studying the climate impact of AI? I reached out to some of our old sources, and some new ones, to find out.

The numbers are vague and chat-only

The first thing they told me is that there’s a lot missing from the figures tech companies published this summer. 

OpenAI’s number, for example, did not appear in a detailed technical paper but rather in a blog post by Altman that leaves lots of unanswered questions, such as which model he was referring to, how the energy use was measured, and how much it varies. Google’s figure, as Crownhart points out, refers to the median amount of energy per query, which doesn’t give us a sense of the more energy-demanding Gemini responses, like when it uses a reasoning model to “think” through a hard problem or generates a really long response. 

The numbers also refer only to interactions with chatbots, not the other ways that people are becoming increasingly reliant on generative AI. 

“As video and image becomes more prominent and used by more and more people, we need the numbers from different modalities and how they measure up,” says Sasha Luccioni, AI and climate lead at the AI platform Hugging Face.