AI is being unfairly targeted over its energy use, OpenAI CEO Sam Altman claims, as the naysayers ignore the vast amount of resources humans have consumed over millennia – not least to avoid being eating by predators.
Altman, speaking during an interview at the AI Summit in India, said it was reasonable to be concerned about AI’s resource consumption: “We need to move toward nuclear or wind and solar very quickly.”
But he suggested some concerns were overdone. “Water is totally fake,” he said, given that AI datacenters often rely on liquid cooling with closed systems, rather than traditional evaporative cooling.
The interviewer referenced claims by Bill Gates that a single ChatGPT query uses the equivalent of 1.5 iPhone battery charges. Altman rubbished those figures, saying: “There’s no way it’s anything close to that.” He claimed such complaints ignore the total amount of energy it takes to create and train an actual human.
He said it was unreasonable to focus on “how much energy it takes to train an AI model, relative to how much it costs a human to do one inference query.”
“It takes like 20 years of life and all of the food you eat during that time before you get smart,” he said. “And not only that, it took the very widespread evolution of the 100 billion people that have ever lived and learned not to get eaten by predators and learned how to figure out science and whatever, to produce you.”
AI’s biggest energy burn comes in the training phase, Altman said. “If you ask ChatGPT a question, how much energy does it take once its model is trained to answer that question versus a human? And probably AI has already caught up on an energy efficiency basis, measured that way.”
Needless to say, working out these numbers is tricky. So we asked Gemini to tell us the total energy consumption needed to create all the humans today, and it came up with 10,800,000 TWh.
By comparison, according to Gemini, the total energy invested in the global AI ecosystem stands at 850 to 1,100 TWh. Which would be minimal in comparison if we ignore the fact this has all occurred since the Second World War, with the vast bulk consumed in just the last four years – and that new models are being trained all the time.
Neither does it take into account the vast corpus of material those LLMs were trained on. Material such as “science and whatever” produced by… humans. Or at least the humans that had managed not to be eaten by major predators.
Apart from declaring that humans were terribly inefficient when it comes to energy consumption compared to a datacenter stuffed with GPUs, with each chip consuming the equivalent of a domestic dwelling, Altman touched on multiple other topics.
He played down the impact of AI on jobs, suggesting it would create many other things for people to do, and pointing out that previous waves of innovation have yet to deliver the “leisure” humans had hoped.
Space-based datacenters were highly unlikely to be a thing, at least this decade, he added, thanks to both prohibitive launch costs and the difficulty of replacing defective GPUs once they’re in orbit. “And they do break a lot, still.”
Not taking equity in OpenAI had been “a dumb thing” on his part, he admitted.
Pushed to identify one thing he admired about Elon Musk, Altman said: “He’s extremely good at physical engineering and also extremely good at getting people to perform incredibly well.”
However, he said it was more likely TSMC would lose its monopoly on chip manufacturing than Altman and Musk becoming friends again. ®
