ChatGPT: 0.34 Wh and no methodology

The number

In "The Gentle Singularity", published in June 2025, Sam Altman wrote that the average ChatGPT query uses about 0.34 watt-hours and about 0.000085 gallons of water. The oven comparison that followed, "a little over one second", has been repeated in most coverage since. The index records both figures as disclosed because the company's CEO stated them; it also records that nothing else was stated: no model, no date, no boundary for the water, no definition of average and no method.

The independent check

In February 2025 Epoch AI published a bottom-up estimate for GPT-4o: a typical query, around 500 output tokens on H100 hardware, at about 0.3 Wh, lower than the 3 Wh figure that had circulated for years from older models. The estimate's value is the curve, not the point: about 2.5 Wh for a 10,000-token input and roughly 40 Wh for 100,000 tokens. Altman's average is plausible for short prompts and tells you nothing about the tail.

Water and carbon

0.000085 US gallons is 0.32 mL, close to Google's on-site 0.26 mL, which suggests it is an on-site cooling figure; that is an inference, since no scope was given. The index estimates ChatGPT's water at about 0.1 mL on site for 0.34 Wh at Azure's 0.30 L/kWh, and notes the company's own figure alongside. Carbon per prompt is derived at about 0.13 g on the average US grid, location-based; OpenAI publishes no carbon figure at all.

Training

Nothing has been disclosed for any GPT model. The index's estimate for a GPT-4-class run, about 25,000 A100s for 90–100 days, is roughly 37 GWh, or about 15,000 tonnes on the US grid; for a GPT-4o-class run on H100s about 10,000 tonnes. Both are labelled estimates with ±3× uncertainty and a method note.

Why the distinction matters

A number is a point; a disclosure is a number with enough around it to be checked. Google published a 26-page method with its 0.24 Wh. OpenAI published a sentence with its 0.34 Wh. The index gives OpenAI's efficiency pillar 70 because the sentence is plausible, and its transparency pillar 25 because a sentence is what it is.