Sam Altman states that Chatgpt’s water and energy consumption is tiny. The problem is that it does not give evidence of it

An email of 100 words generated by GPT-4 Consume 519 milliliters of water. That was the conclusion to which researchers at the University of California arrived a few months ago after analyzing this OpenAi model. Sam Altman, CEO of the company, has just yielded its own estimate on the consumption of water and energy of each consultation of Chatgpt. And it is very different.

1,000 times less than what was said. According to Altman, an average consultation in Chatgpt consumes much less than what had been indicated in previous studies. Your data are strikingand to understand them makes interesting analogies:

“As production automated in data centers automates, the cost of intelligence should approach electricity. (People are usually curious to know how much energy consumes a chatgpt consultation; the average consult (0.32 ml);

A previous study of Epoch ai corroborates the data that Sam Altman has now wielded. Source: Epoch AI.

And the tests? Those figures mentioned by the OpenAi CEO have a problem: they have no visible support. He throws them without citing sources or explaining where he has taken them out, something that makes it difficult to believe. A Meta executive answered the question of How much consumes the inference AI A year and a half ago, responding that “only two nuclear reactors would be needed to cover it.”

But previous studies coincide with Altman. Although he does not mention any evidence, in February, Epoch AI researchers precisely They published a study trying to estimate the energy consumption of chatgpt. In their conclusions they indicated that on average a chatgpt consult Previous report of the researcher Alex de Vries. Since then, of course, many things have happened.

Too pessimistic. And as they commented on the study of Epoch AI, the difference comes from the fact that the models are today much more efficient than in 2023, when VRies conducted their study. So is the hardware in which these models are executed, and that estimate was also used a “especially pessimistic” approach. In Openai’s study they also threw an especially pessimistic estimate and pointed out that “most of the requests (A chatgpt) are much cheaper (energetically).”

More studies. Another independent study published by Andy Masey in January 2025 reached a similar conclusion and claimed that “using Chatgpt is not bad for the environment.” It was based on EPRI data May 2024 that also estimated a high consumption of 2.9 Wh by chatgpt consultation. Estimated water consumption In data centersfrom A SUNBIRD studyit was also very modest compared to other online activities.

Water consumption in data centers for various online activities. Source: Andy Masley.

Fifte. Precisely the data of water consumption was another striking in that estimate of Sam Altman. According to him, a chatgpt consultation barely consumed 0.32 ml of water, “a quinceava part of a teaspoon.” The figure suggests that the water needed to refrigerate data centers that process these requests is much less than what was thought only one year ago.

And training, what? These estimates focus on the AI ​​inference section, that is, our use of chatgpt that receives a consultation and processes it inferring (generating) a text result. Although Altman does not clarify it, he does not seem to include here the energy and water cost of training AI models, which is very high and makes thousands of Gpus They work at full power For months, with the consequent water expense in data centers to refrigerate all those components that dissipate high heat amounts. As I pointed out The researcher Ethan Mollick, GPT-4 probably used more than 50 GW to be trained, enough to give energy to 5,500 homes in a year.

We continue without definitive data. Altman’s claims are as always striking, but the lack of clear evidence makes it difficult to believe these data. Other recent studies are more useful when it comes to reflecting this increasingly lower cost both in energy and water from the use of AI, but there are no accepted standards or a consensus on the true impact of energy and water consumption when using chatgpt or other AI models.

Image | Lukáš Lehotský | Village Global

In Xataka | The light price is again negative: it is a sign that the system needs a redesign