Sam Altman Says It Takes 20 Years of Food to Train a Human, So AI Energy Use Is Fine

Sam Altman sat down with Anant Goenka at Express Adda in New Delhi on Friday, on the sidelines of the AI Impact Summit 2026. The conversation covered India’s AI ambitions, OpenAI’s growth, and Elon Musk being Elon Musk. But one moment caught fire online: Altman’s argument that training a human takes more energy than training an AI model.

Here’s the clip:

What Altman Actually Said

Altman was responding to criticism about AI’s growing energy and water consumption. He made three distinct claims:

On per-query energy: He pushed back on a figure attributed to Bill Gates suggesting a single ChatGPT query uses the equivalent of one iPhone battery charge. “There’s no way it’s anything close to that much. It’s way, way, way less.”

On water usage: He called the viral claim that ChatGPT uses 17 gallons of water per query “completely untrue, totally insane, no connection to reality.” He acknowledged older data centers used evaporative cooling but said that practice has been discontinued.

On training costs, the big one: “People talk about how much energy it takes to train an AI model relative to how much it takes a human to do one inference inquiry.” His rebuttal: “It also takes a lot of energy to train a human. It takes like 20 years of life and all of the food you eat during that time before you get smart.”

He went further, noting that human intelligence is the product of “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.” His conclusion: AI has “probably caught up” on an energy efficiency basis when measured per-task after training.

The Part Where He’s Right

There’s a kernel of a real argument here. When people compare the cost of a single ChatGPT query to something tangible (a glass of water, a light bulb running for an hour), they’re comparing inference cost to human inference cost. But they’re ignoring the massive upfront investment in the human: two decades of caloric intake, education, shelter, healthcare. If you amortize all of that across every “query” a human ever answers, the per-query cost is… not zero.

Altman is also correct that the viral stats about water usage per query are wildly inflated. Most of those figures trace back to misreadings of a 2023 University of California Riverside paper that measured total facility water use, not per-query consumption. The numbers got laundered through enough headlines that “one glass of water per query” became accepted wisdom.

The Part Where He’s Not

The comparison falls apart once you look at scale. A human, once trained, runs on roughly 2,000 calories a day and handles a huge variety of tasks without additional infrastructure. An AI data center runs on megawatts of electricity, requires constant cooling, and needs re-training or fine-tuning as models evolve.

MIT Technology Review projects that by 2028, over half of all data center electricity will power AI systems, potentially matching 22% of total U.S. household electricity consumption. That’s not a per-query problem. That’s a civilization-scale energy allocation question.

Altman actually acknowledged this himself in the same conversation: “What is fair, though, is the energy consumption, not per query, but in total, because the world is now using so much AI, which is real, and we need to move towards nuclear or wind and solar very quickly.”

So he made the strongest counterargument to his own point. Total energy consumption is the real issue, and “but humans eat food” doesn’t address it.

The Missing Disclosure Problem

There’s another layer here that Altman didn’t touch. No legal requirement currently mandates that tech companies disclose their energy and water consumption. OpenAI doesn’t publish per-model training costs. Neither does Google, Meta, or Anthropic. We’re debating numbers in a vacuum because the companies generating those numbers won’t share them.

Altman asking people to trust that the per-query costs are low while simultaneously not releasing the data is a familiar pattern. The ask is always: trust us, the numbers aren’t that bad. The response from researchers is always: then show us the numbers.

Why the Food Line Went Viral

The reason this quote spread isn’t because it’s wrong in the narrow sense. It’s because it sounds like a CEO of a company burning through billions in compute costs telling regular people that they’re the real energy problem. The framing landed badly.

It’s the same dynamic as when airline executives point out that private jets are a tiny fraction of total emissions. Technically true. Rhetorically tone-deaf.

Altman’s broader point, that we should compare like to like and not cherry-pick misleading per-query stats, is reasonable. But wrapping it in “it takes 20 years of food to train a human” gave the internet exactly the kind of quote that writes its own punchline.

The Actual Energy Question

Strip away the meme-able quote and there’s a real conversation underneath: how much energy should AI consume, who pays for it, and what energy sources should power it?

Altman said the right thing about needing to move toward nuclear, wind, and solar quickly. OpenAI has signed deals with nuclear energy providers. So have Google and Microsoft. The industry is putting money behind the claim that AI growth and clean energy can coexist.

Whether that’s happening fast enough is a different question. Data center construction is outpacing renewable energy buildouts in most regions. Natural gas is filling the gap. The IEA projects global data center electricity consumption will double by 2028.

Altman comparing AI training to human biology is a fun thought experiment. It just doesn’t change any of those numbers.

---

Sources: TechCrunch, Tom’s Hardware, Complex, AI Chief, BusinessToday