ChatGPT’s appetite for power no longer reads like trivia, it resembles a layer of infrastructure. Each exchange a user triggers enlarges the AI energy footprint, turning curiosity into measurable strain.
Analysts estimate the service draws about 17 terawatt-hours of electricity each year. That output could keep New York City lit for nearly four months, showing how every reply adds to rising data center load and a growing share of New York City electricity demand, while dense metros electrify transport and heating, and the extra draw intensifies urban grid pressure worldwide.
What the numbers say about ChatGPT’s annual electricity use
BestBrokers data analyst Alan Goldberg uses recent University of Rhode Island AI Lab research to quantify ChatGPT’s power draw. On a global terawatt-hour scale, their latest annual consumption estimate reaches 17.23 TWh devoted purely to answering user prompts, equal to roughly 17.228 billion kilowatt-hours. That load mirrors the yearly grid demand of Puerto Rico or Slovenia and overtakes the total electricity use reported for Latvia or Luxembourg.
Framed against national systems, the same volume of energy would keep New York City running for 113 days and supply the entire United States for about 34 hours. BestBrokers clarify their underlying usage assumptions and cite academic and industry methodology sources to show how they moved from raw model activity to a global footprint.
From a single query to terawatt-hours, how usage scales
University of Rhode Island researchers calculate that a single ChatGPT request consumes about 18.9 watt-hours, or 0.0189 kilowatt-hours, for the most advanced models. In their comparison, this estimated per-query energy use yields an inference cost per request more than fifty times higher than a typical Google search, which BestBrokers peg at roughly 0.3 watt-hours.
Aggregated across the globe, that tiny slice of electricity per prompt turns into around 2.5 billion daily requests and roughly 47.2 million kilowatt-hours consumed every day. BestBrokers attribute this surge to an audience of nearly 810 million weekly active users, each sending about 22 questions over seven days in their working model.
The price tag in dollars and emissions for keeping ChatGPT running
BestBrokers apply an average U.S. commercial electricity cost per kWh of 0.141 dollars, based on September 2025 figures, to ChatGPT’s inferred energy draw. Using that rate, the 17.228 billion kilowatt-hours associated with responses alone translate into around 2.42 billion dollars per year, a sizeable share of OpenAI’s global operational spend before training or hardware are even counted.
On the demand side, BestBrokers calculate that this annual volume could power all U.S. households for more than four and a half days at 29 kilowatt-hours per home and per day, or charge roughly 238 million electric cars with 72.4 kilowatt-hour packs. When typical carbon intensity of grid values are applied, their resulting emissions comparison aligns with those of several mid-sized industrialised nations.
Why transparency claims around power use are under scrutiny
Goldberg at BestBrokers argues that training frontier AI models already consumes “tens of gigawatt-hours” and that ChatGPT’s everyday operation adds another persistent burden to power systems. This rising demand exposes reporting standards gaps, since OpenAI and competitors rarely publish audited, model-by-model figures that break out training, validation and deployment electricity use.
According to the BestBrokers study, investors, regulators and city authorities such as those in New York mostly rely on unverified company statements when assessing grid impact and climate risk. Clearer disclosure practices, sector-wide benchmarks and accessible independent verification would give energy planners far more confidence in claims about efficiency gains and “green” infrastructure choices.