“People often Openai CEO Sam Altman wrote in a long blog post last week.
For a company with 800 million active users per week (as well as growth), the question of the energy used by all these searches is becoming increasingly urgent. But experts say that without Openai’s more public context, Altman’s numbers don’t make much sense, i.e. how it reaches that calculation, including the definition of “average” queries, whether it includes image generation, and whether Altman includes additional energy use, such as training AI models and cooling OpenAi’s servers.
As a result, Sasha Luccioni, the head of climate at AI companies embracing faces, didn’t invest much in Altman’s figures. “He could have pulled it out of his butt,” she said. (Openai did not respond to a request for more information on how the number reached the number.)
As AI takes over our lives, this also promises to change our energy systems, and when we try to resist climate change, carbon emissions can be placed on recharge. Now, a new research institute is trying to make tough numbers on how much carbon actually emitted by all our AI.
The fact that major players like OpenAI have little environmental information complicates this effort. Luccioni and three other authors submitted a peer-reviewed analysis this week aimed at the need for more environmental transparency in AI models. In Luccioni’s new analysis, she and her colleagues used data from the Large Language Model (LLM) traffic ranking OpenRouter, and found that 84% of LLM usage in May 2025 were used for models with zero environmental disclosure. This means that the vast majority of consumers choose models with completely unknown environmental impacts.
“What shocked me is that you can buy a car and know how many miles per gallon it consumes, but we use all of these AI tools every day, and we absolutely have no efficiency metrics, emissions factors, nothing,” Luccioni said. “It’s not mandatory, not regulation. Given our place in the climate crisis, it should be the agenda of regulators everywhere.”
Lucionice said that due to the lack of transparency, the public is being exposed to situations where estimates are meaningless, but is seen as a boon. For example, you may have heard that the average CHATGPT request requires 10 times more energy than a regular Google search. Luccioni and her colleagues trace the claim back to public remarks that John Hennessy, chairman of Google’s parent company Alphabet, was founded in 2023.
A company (Google) board member’s product of products that he has no relationship with (OpenAI) is trivial at best – but Luccioni’s analysis found that this number is repeated time and time again in news and policy reports. (When I wrote this post, I got this exact statistic.)
