How much energy and water does AI use, and how much CO₂ does it emit?
Honest, research-backed estimates for the AI in your business, shown as ranges instead of headlines. The real answer spans 100×, and it depends on choices you control.
What the research actually says
0.2–2 Wh
per typical text request
Somewhere between a Google search and a few seconds of microwave time, plus a fraction of a millilitre of cooling water.
~100×
spread from model choice
Reasoning-heavy models use 10 to 30 times more energy than standard text. Efficient small models use far less.
2–1,000×
spread from what you count
Including the whole data center, training, or the water behind the electricity multiplies any figure. That’s why we show ranges, not point estimates.
Sources: Google’s production measurement, Mistral’s lifecycle audit, and independent benchmarks. Anthropic had published no per-query disclosure as of mid-2026.
Estimate your AI footprint
Pick the closest match for how your business uses AI and tweak the numbers. The calculator shows ranges, because that’s what honest looks like.
Tweak the numbers:
That’s about 95 per working day.
Response length
Your estimated monthly footprint:
Energy
1.9 kWh
range: 0.77 – 3.8 kWh
≈ under 1 hour of home central A/C
CO₂e
749 g
range: 300 – 1,498 g
≈ 2 miles in an average gasoline car
Water
1.9 L
range: 0.77 – 3.8 L
data-center cooling only
≈ under 1 minute of showering
Equivalents use EPA and EIA conversion factors and are communication aids, not inventory figures.
How we calculate this: assumptions and sources
This is a directional estimate for education and planning, not a certified carbon measurement. Not suitable for ESG or regulatory reporting.
Monthly energy = requests × per-request energy × 1.2 facility uplift. CO₂e = energy × your grid’s carbon intensity. Water = energy × 1.0 L/kWh for onsite cooling.
Per workload class
Efficient text models
0.1–0.3–0.6 Wh per 1,000 output tokens
Sources: arxiv.org, huggingface.github.io
Grid carbon intensities
- United States (eGRID average): 0.39 kg CO₂e/kWh
- European Union average: 0.24 kg CO₂e/kWh
- World average: 0.47 kg CO₂e/kWh
Source: www.epa.gov
Assumptions last reviewed: 2026-06-10
The footprint and the bill are the same problem
Everything that cuts AI’s environmental footprint also cuts your API bill: right-sizing models to tasks, caching repeated answers, keeping prompts tight, and processing documents once instead of every time. We build AI workflows for small businesses with exactly those habits. Efficiency isn’t an add-on, it’s just good engineering.
Run an engineering or environmental firm? See how we work with firms like yours.
Frequently asked questions
How much energy does a single AI request use?
For a typical text request to a modern model, credible 2025 and 2026 disclosures land between roughly 0.2 and 2 Wh, about the range between a Google search and running a microwave for a few seconds. Reasoning-heavy models can use 10 to 30 times more. Image generation runs a few Wh per image.
Does AI use a lot of water?
Per request, onsite cooling water is small, from fractions of a millilitre to tens of millilitres. Headlines like "a bottle of water per email" also count the water used to generate the electricity, a much larger scope that takes in the whole upstream supply chain and goes beyond the scope of our estimate. Both framings are legitimate; we label ours clearly as onsite cooling only.
Why does your calculator give ranges instead of exact numbers?
Because published estimates for the same workload span up to 100×, a single precise-looking number would be false confidence. Ranges reflect the genuine uncertainty in model choice, hardware, and measurement boundary.
Can I use these numbers for ESG or sustainability reporting?
No. This is a directional, educational estimate, not a GHG Protocol Scope 3 measurement. For formal carbon accounting, start from your AI providers' own disclosures or work with a carbon accounting platform like Greenly. If you need real numbers, we can instrument your actual token usage.
What actually reduces AI's environmental impact?
The same things that reduce cost: use the smallest model that does the job, cache answers to repeated questions, keep prompts short, avoid reasoning models for routine tasks, and process documents once. Model right-sizing alone can cut energy 10× or more.
Tools and sources we trust
Our methodology follows EcoLogits, an open-source standard for estimating LLM impacts. To compare specific models, see the AI Energy Score leaderboard. For formal carbon accounting and ESG reporting, we refer clients to Greenly. Household conversion factors come from the EPA and EIA.