OpenAI CEO Sam Altman rejected an online claim that ChatGPT uses 17 gallons of water per query during a February 20, 2026, interview in New Delhi, while acknowledging concerns about AI’s total electricity consumption.
Speaking on the sidelines of the India AI Impact Summit, Altman described warnings about the purported water requirement and said: “This is completely untrue, totally insane, no connection to reality.”
Altman called concerns about AI’s water use “totally fake” and linked his position to a move away from evaporative cooling. He said “we used to do evaporative cooling in data centres” and that this was no longer the case.
Cooling methods and water use
Data centers can use water directly for cooling and indirectly through the generation of the electricity they consume. Systems that avoid evaporation can reduce direct water consumption, while facilities may still require water for other purposes.
In December 2024, Microsoft announced a data center design that it said would circulate cooling water through a closed loop without evaporation. The company said the system would require an initial fill during construction and that facilities would still use water for restrooms and kitchens.
In that announcement, Microsoft said its existing fleet would continue to use a mix of cooling systems. It scheduled pilots of the new design for 2026 and said new sites would begin coming online in late 2027.
Electricity use and efficiency claims
Altman said concern about AI’s total energy consumption was “fair” because of its growing use. He called for a rapid move toward nuclear, wind and solar power.
Interviewer Anant Goenka also asked whether a single ChatGPT query used electricity equivalent to 1.5 iPhone battery charges. “There’s no way it’s anything close to that much,” Altman replied.
Altman challenged comparisons between the energy needed to train an AI model and the energy a person uses to answer one question. “But it also takes a lot of energy to train a human,” he said, referring to years of development and food consumption.
He argued that the comparison should consider the energy a trained model uses to answer a question alongside the energy required for a human response. Altman suggested AI had probably already caught up with humans on energy efficiency under that comparison.
The International Energy Agency estimated that data centers worldwide consumed around 415 terawatt-hours (TWh) of electricity in 2024, about 1.5% of global electricity consumption. Its base case projected demand would roughly double to around 945 TWh by 2030, just under 3% of global consumption, with AI a major driver of growth. The agency identified substantial uncertainty around future demand, including the pace of AI adoption and improvements in efficiency.
A 2024 Lawrence Berkeley National Laboratory report produced for the U.S. Department of Energy projected U.S. data center electricity consumption of 325–580 TWh in 2028, equivalent to approximately 6.7%–12% of national electricity consumption. Both the U.S. and global projections cover data centers across workloads.
