OpenAI CEO Sam Altman participated in an event organized by The Indian Express. During the interview made some striking statements, but the greatest of all of them was the one he dedicated to talking about what it costs to train an AI model. In fact, he complained about how many of ChatGPT’s energy consumption discussions they are unfair.
Training humans also consumes a lot. The interviewer asked Altman about ChatGPT’s energy consumption and Sam Altman took a few seconds to answer the question, and then made a peculiar comparison (my bold):
One of the things that is always unfair in this comparison is that it talks about how much energy it takes to train an AI model compared to what it costs a human to perform an inference query.
But it also takes a lot of energy to train a human. It takes about 20 years of life and all the food you eat during that time before you become intelligent. And not only that, it took the widespread evolution of the hundred billion people who have lived and learned not to be eaten by predators and to understand science and so on to create you.
The fair comparison is if you ask ChatGPT, how much energy does it take once their model is trained to answer that question compared to a human? And AI has probably already caught up in terms of energy efficiency if we measure it that way.
A previous Epoch AI study corroborates that energy consumption during inference (when we actually use ChatGPT, for example) is low. Source: Epoch AI.
Training is one thing, inference another.. The answer may be controversial, but to a certain extent it is logical: learning, both in the case of humans and AI, takes time and consumes many resources, but that cost is one thing and the cost of inference, of “applying that training”, is another. Once we have learned, it is not too difficult to answer things. This is what Altman is trying to point out here, who recognizes that AI does indeed consume a lot of energy in training, but that it has then become very efficient in the inference phase, when we actually use ChatGPT. The problem is that although Altman has already spoken that in inference consumption is minimal, does not provide evidence of this.
The water problem is no longer a problem. He also spoke about the controversial water consumption that was theoretically carried out in large AI data centers. Although he acknowledged that this was a problem when “we used to use evaporative cooling in data centers.” Now, however, “we don’t do that,” he recalled, and made it clear that those accusations that “ChatGPT uses 17 gallons per query, or whatever” is totally false, “totally crazy, it has no connection with reality.” But again, there is still no official data from AI companies in this section.
How much does AI really consume? The truth is that at this point we still do not have really clear data on how much the AI consumes both in the training phase and in the inference phase. There are those who have investigated energy and water consumption and have made a mistake. wildly exaggerating the databut for example in the US, where a large number of data centers are concentrated, there is no legislation that forces transparency with those figures.
Increasingly more efficient models and data centers. One of the most interesting studies was the one made by Epoch AI in February 2025, and at that time it was also concluded that AI did not actually consume as much as it was said to consume. In fact, it consumed relatively little and the models have only improved in efficiency. Chips and cooling systems have also improved, and although data centers have certainly require enormous amounts of energywe continue blindly in this section.
In Xataka | Spain has a plan to capture more data centers than anyone else: “shield” them from energy costs
GIPHY App Key not set. Please check settings