The footprint of a prompt
Large language models cost energy, water, and carbon to run. But how much? Pick an everyday activity and compare its carbon footprint with a text AI scenario. Results are approximate and depend on the assumptions below.
Reviewed November 3, 2026 · Default AI measurement: May 2025. This is a scenario comparison, not a measurement of your AI usage.
That's about — of carbon emissions.
Under these assumptions, roughly
—
text requests would match that carbon footprint
≈ — assumed output tokens (at ~500 output tokens per request; not a measured token footprint)
Assumptions & sources (edit the numbers)
Per AI request
Selected activity
Gemini median uses Google's May 2025 Gemini Apps text estimate, published August 2025: 0.24 Wh, 0.26 mL water, 0.03 gCO₂e per request. Carbon is market-based and includes hardware embodied emissions; water is data-center cooling consumption. Training and end-user devices are excluded. This historical median is not an average for all AI or a universal lower bound.
Illustrative heavy request scales the Gemini figures by 40: 9.6 Wh, 10.4 mL, 1.2 gCO₂e. This is an editable scenario, not a measured professional-use footprint or a proven conservative estimate. A June 2026 study models typical queries at 0.31 Wh and an extended reasoning scenario at 3.91 Wh; these energy estimates do not establish matching carbon or water figures.
Token conversion: 500 output tokens is borrowed from Epoch's GPT-4o assumption, not Google's measured Gemini request length. The heavy scenario's 8,000 output tokens is also assumed. Input, cached, and hidden reasoning tokens are not modeled separately. Token totals are illustrative arithmetic only.
The carbon comparison mixes accounting boundaries: driving and gas heating use direct CO₂ emissions; AI uses CO₂-equivalent emissions. The other activity factors are unsourced illustrative estimates. Energy and water below the result describe the AI scenario only; they are not equivalences with the selected activity.
Sources
- Google — AI serving footprint (August 2025; May 2025 measurement): source of the Gemini default.
- Epoch AI — How much energy does ChatGPT use? (February 2025): about 0.3 Wh for a GPT-4o query with an assumed 500 output tokens.
- Oviedo et al. — Energy Use of AI Inference (June 2026 revision): modeled energy varies with workload and deployment.
- US EPA — Typical passenger vehicle: about 400 g CO₂/mile, tailpipe only.
- US EPA — Equivalencies references: 5.3 kg CO₂/therm, combustion only. Five therms per winter day is an assumption.
Read the research and methodology notes for the evidence, accounting boundaries, and remaining unknowns.