Which AI models publish their footprint?

A short table of what each vendor discloses, and what is left to independent researchers. Where a number is missing, this page says so rather than filling the gap.

What disclosure actually covers

Vendors publish on-site cooling water because they measure it directly. Energy per prompt is harder: the same model answers short and long questions, and the hardware behind it changes. Independent benchmarks fill the gap by running standardised prompts and measuring power draw, which is why the energy figures on this site come from research rather than from the vendors.

“A single short GPT-4o query consumes 0.43 Wh ... o3 and DeepSeek-R1 emerge as the most energy-intensive models, consuming over 33 Wh per long prompt, more than 70 times the consumption of GPT-4.1 nano.”

Reading the table honestly

A missing number is not a scandal. It usually means the vendor has not published that quantity in that unit. The useful comparison is between figures with the same boundary, which is why every model page below states its scope before its number.

Why energy is missing from most rows

Water at the cooling tower is metered. Energy per prompt is not: it depends on the model, the length of the answer, the batch size and the hardware generation, and the same model changes all four between releases. Vendors publish totals and intensities instead, which cannot be divided into a per-prompt figure without assumptions.

Independent benchmarks solve this by fixing the prompt and measuring the power draw. That produces a comparable number, at the cost of measuring a standardised prompt rather than your actual use.

How to read a row

A published figure is comparable only with another figure using the same boundary. A missing figure is not evidence of a worse footprint; it is evidence of a shorter disclosure. Treating the two as the same thing is the most common error in coverage of this subject.

Model pages

ChatGPT Gemini DeepSeek