Guide · methodology v2026.06 · updated 2026-06-12
How much CO₂ does AI emit?
Short answer: one AI chat answer from a standard model (about 500 output tokens) uses roughly 0.24–0.48 Wh of electricity and emits 0.089–0.18 g CO₂e in a typical US East data center. The same answer emits 0.0060–0.012 g CO₂e in Europe (Stockholm) and 0.15–0.30 g CO₂e in Asia (Mumbai): where the model runs changes the footprint by about 25×, and choosing a reasoning model multiplies it by 4–8×.
Energy and CO₂ per AI query, by model class
Per query of 500 output tokens, at a data-center PUE of 1.2. Figures are ranges because the energy used per token is uncertain by a factor of three to five; emissions use location-based annual grid averages.
| Model class | Energy (Wh) | g CO₂e · Mumbai | g CO₂e · N. Virginia | g CO₂e · Frankfurt | g CO₂e · Paris | g CO₂e · Stockholm |
|---|---|---|---|---|---|---|
Light Haiku, GPT-4o-mini, Gemini Flash | 0.072–0.15 | 0.045–0.094 | 0.027–0.055 | 0.025–0.052 | 0.0040–0.0082 | 0.0018–0.0037 |
Standard GPT-4o, Sonnet, Gemini Pro | 0.24–0.48 | 0.15–0.30 | 0.089–0.18 | 0.084–0.17 | 0.013–0.026 | 0.0060–0.012 |
Frontier / reasoning Opus, o-series, extended thinking | 0.90–3.60 | 0.57–2.27 | 0.33–1.33 | 0.31–1.26 | 0.049–0.20 | 0.023–0.090 |
What that adds up to
A product serving one million standard-model queries a month from US East (N. Virginia) uses about 240–480 kWh and emits 88.8–178 kg CO₂e per month — comparable to driving 522–1,045 km in an average car. Served from Europe (Stockholm), the same traffic emits 6.00–12.0 kg CO₂e, a reduction of about 93%.
Carbon intensity of cloud regions
Grams of CO₂ emitted per kWh of electricity (location-based annual average, Ember), cleanest first. Emissions per query scale in direct proportion.
| Region | Grid | g CO₂/kWh | Provider regions |
|---|---|---|---|
| Europe (Stockholm) | Sweden (hydro + nuclear) | 25 | AWS eu-north-1 · Azure Sweden Central |
| Canada (Montréal) | Canada — Québec (hydro) | 30 | AWS ca-central-1 · GCP northamerica-northeast1 |
| Europe (Paris) | France (nuclear) | 55 | AWS eu-west-3 · Azure France Central |
| Europe (London) | United Kingdom | 220 | AWS eu-west-2 · Azure UK South |
| Europe (Amsterdam) | Netherlands | 270 | Azure West Europe · GCP europe-west4 |
| US West (Oregon) | United States — Oregon | 285 | AWS us-west-2 · GCP us-west1 |
| Europe (Ireland) | Ireland | 290 | AWS eu-west-1 · Azure North Europe |
| US Central (Iowa) | United States — Iowa | 340 | GCP us-central1 · Azure Central US |
| Europe (Frankfurt) | Germany | 350 | AWS eu-central-1 · GCP europe-west3 |
| US East (N. Virginia) | United States — Virginia | 370 | AWS us-east-1 · Azure East US · GCP us-east4 |
| Asia (Tokyo) | Japan | 450 | AWS ap-northeast-1 · Azure Japan East |
| Asia (Singapore) | Singapore | 470 | AWS ap-southeast-1 · GCP asia-southeast1 |
| Australia (Sydney) | Australia — NSW | 510 | AWS ap-southeast-2 · Azure Australia East |
| Asia (Mumbai) | India | 630 | AWS ap-south-1 · GCP asia-south1 |
Four ways to cut the footprint of AI
- Choose a cleaner region. Moving a workload from Asia (Mumbai) to Europe (Stockholm) cuts emissions by about 96% with no change to the model.
- Use the smallest model that does the job. Light models use about a third of the energy of standard ones; reasoning models use 4–8× more.
- Generate fewer tokens. Energy scales with output length — shorter answers, caching and avoiding unnecessary reasoning all help.
- Run batch jobs when the grid is clean. Grid carbon intensity changes hour by hour; work that can wait can run in the cleanest window.
What these figures do not include
They cover inference only. Model training is not amortized into per-query numbers, hardware manufacturing (embodied carbon) is an optional +30% in the estimator, and water use is out of scope. They are estimates for comparison and planning, not audited emissions data. The full method, every factor and its source are on the methodology page; what each provider has published is in the AI energy disclosures tracker.
Estimate your own AI footprint
Enter your monthly queries or tokens, model classes and cloud region to get kWh and kg CO₂e, a cleaner-region alternative and a vendor comparison. Free, no account needed.
Open the AI Footprint Estimator →Frequently asked questions
- How much energy does one ChatGPT query use?
- About 0.3 Wh for an average query, according to Epoch AI's 2025 analysis and OpenAI's own statement; Google reports a median of 0.24 Wh per Gemini text prompt. CO2Radar's estimate for a standard-class model producing about 500 output tokens is 0.24–0.48 Wh including data-center overhead. The older figure of 3 Wh per query is outdated.
- How much CO₂ does one AI query emit?
- Between 0.0060–0.012 g CO₂e and 0.15–0.30 g CO₂e for a standard-class model, depending on the electricity grid behind the data center. In US East (N. Virginia), where much US AI capacity runs, it is 0.089–0.18 g CO₂e.
- Which cloud region has the lowest carbon footprint for AI?
- Of the regions CO2Radar compares, Europe (Stockholm) is the cleanest at about 25 g CO₂ per kWh, followed by Canada (Montréal) (30) and Europe (Paris) (55). The most carbon-intensive is Asia (Mumbai) at 630 g CO₂ per kWh — about 25 times higher for the same workload.
- Do reasoning models use more energy than standard models?
- Yes. Frontier and reasoning models (such as Claude Opus or OpenAI's o-series) use roughly 4–8 times more energy per 1,000 output tokens than standard models, partly because they generate hidden reasoning tokens. Light models (Haiku, GPT-4o-mini, Gemini Flash) use about a third of a standard model.
- How do I calculate the carbon footprint of my AI usage?
- Multiply your output tokens by the energy per token for your model class, multiply by the data center's PUE to get total energy, then multiply by the carbon intensity of the local grid: energy (Wh) = tokens ÷ 1000 × Wh per 1k tokens × PUE, and emissions = energy (kWh) × grid intensity (g CO₂/kWh). CO2Radar's free estimator does this for any mix of workloads and regions.
- What is the difference between location-based and market-based emissions?
- Location-based emissions use the actual mix of power plants on the grid where the data center draws electricity. Market-based emissions subtract renewable energy the provider has bought through certificates or power-purchase agreements, so they are usually much lower. CO2Radar reports location-based figures because they describe the physical grid your workload runs on.
- Does the footprint of a query include training the model?
- No. Per-query figures cover inference only: the chips answering your request, idle capacity, cooling and other data-center overhead. Training is a large one-off cost that cannot be attributed to a single query without knowing a model's total lifetime usage, so CO2Radar reports it separately rather than amortizing it.
Sources
- Google (2025) — median Gemini text prompt: 0.24 Wh / 0.03 gCO2e, incl. idle, CPU/RAM and datacenter overhead
- Epoch AI / Sam Altman (2025) — ~0.3 Wh per average ChatGPT query
- Mistral AI (2025) — LCA: ~1.14 gCO2e per 400-token answer
- Ember — Yearly electricity data (grid intensity per country, location-based)
- ENTSO-E Transparency Platform — live EU generation mix (planned live layer)
- Uptime Institute (2024) — global average PUE ~1.55