Estimate the energy, water and carbon behind your AI use — and what it means in everyday terms.
Inference only (training and data-centre overhead excluded); water is on-site (scope-1); the historical Google-search figure is from 2009. All figures are estimates.
These are order-of-magnitude estimates from public measurements of comparable models — inference-only, with on-site (scope-1) water and adjustable carbon and water factors. Closed frontier models aren’t directly measured, so treat the result as a ballpark, not a precise audit.
Each AI task carries a per-use energy figure in watt-hours, drawn from current measured sources. Your monthly counts are multiplied out to a year, summed, and converted to kilowatt-hours.
Water is the on-site cooling water: energy in kWh multiplied by a water-use-effectiveness factor (litres per kWh). Carbon is energy multiplied by your electricity grid's carbon intensity (grams of CO₂ per kWh).
A typical short text prompt to a current frontier model uses roughly 0.3 Wh — the convergence point of Google's 0.24 Wh, OpenAI's 0.34 Wh and Epoch's 0.3 Wh estimates. Reasoning ('thinking') prompts and image generation use far more.
Water shown here is on-site cooling water (scope-1): kWh × a water-use-effectiveness factor. It does not include the off-site water used to generate the electricity (scope-2), so the true total can be higher. The factor varies a lot by data centre, from about 0.26 to 1.9 L/kWh.
No. They are central estimates from public measurements of comparable models, labelled inference-only. The closed frontier models are not directly measured, and real-world figures include data-centre overhead the raw measurements exclude.
No. Every figure is inference-only — the cost of using the model, not the one-time cost of training it.
This calculator turns your AI use into its real-world footprint: energy in kilowatt-hours, water in litres, and carbon in kilograms of CO₂. Pick "My usage" to estimate a year of your own prompts, or "Per task" to see what a single chat, reasoning request, or image generation costs. Every per-task figure is drawn from current, cited measurements and is inference-only — the cost of using a model, not of training it.
The headline number is your carbon footprint (annual in "My usage" mode, per the chosen volume in "Per task" mode). The three metric cards break that down into energy, water and carbon, each paired with an everyday equivalent — a phone charge, a glass of water, a kilometre driven — so the abstract units become tangible. Treat the result as an order-of-magnitude estimate: the underlying per-task energy figures span a wide range depending on the model, prompt length and data centre, and the water and carbon factors are adjustable assumptions, not fixed facts.
You send about 100 short AI chats a month, with the water-use factor at the 1.8 L/kWh industry default and the world-average grid (471 gCO₂/kWh).
100 chats a month is 1,200 a year. At 0.3 Wh each that is 360 Wh, or 0.36 kWh — roughly 30 phone charges. The water works out to about 0.65 L (under three glasses) and the carbon to about 170 g of CO₂ — well under a kilometre in a typical car. A single year of light chat use is a very small footprint; reasoning prompts and image generation are where it grows.
A typical short text prompt to a current frontier model uses roughly 0.3 Wh of electricity — the convergence point of Google's measured 0.24 Wh for a median Gemini prompt, OpenAI's stated 0.34 Wh per ChatGPT query, and Epoch AI's bottom-up 0.3 Wh estimate for GPT-4o. Long documents (~2.5 Wh), reasoning models (~7.6 Wh) and image generation (~2.9 Wh) cost far more per use.
Scope-1 water is the on-site water a data centre evaporates to cool its servers, calculated as energy (kWh) multiplied by a water-use-effectiveness (WUE) factor in litres per kWh. It excludes scope-2 water — the water used off-site at the power plant generating the electricity — so the full water cost is higher than what is shown. WUE varies widely between operators (about 0.26 to 1.9 L/kWh), which is why it is adjustable here.
No. They are central estimates from public measurements of comparable models, and they are deliberately labelled inference-only and scope-stated. Closed frontier models are not directly measured, and measurement boundaries differ (full-system energy is roughly 2.4× GPU-only energy), so two "per-query" numbers from different studies are not interchangeable.
No. Every figure here is inference-only: the cost of running a prompt, not the one-time cost of training the model. Training is a large but separate, amortised cost that this calculator does not attempt to estimate.
Open the water and grid assumptions section. The grid selector swaps in the lifecycle carbon intensity for the world average, the US, the EU, India or China (2024 OWID data), and the WUE selector lets you model a Google, Microsoft, Meta, industry-default or fleet-average data centre. Carbon and water both scale linearly with these factors.
Each AI task has a cited per-use energy value in watt-hours: short chat 0.30 Wh (Epoch/Google/OpenAI convergence), long document 2.5 Wh (Epoch, ~10k-token input), reasoning 7.6 Wh (Hugging Face AI Energy Score v2, DeepSeek-R1-70B with reasoning on), and image generation 2.907 Wh (Luccioni "Power Hungry Processing", Table 2 mean). In "My usage" mode the monthly counts are multiplied by 12, summed, and divided by 1,000 to give annual kilowatt-hours; "Per task" mode uses the counts as entered. Water in litres = kWh × WUE (default 1.8 L/kWh, scope-1 on-site). Carbon in grams = kWh × grid intensity (default 471 gCO₂/kWh, world 2024, OWID lifecycle), then divided by 1,000 for kilograms. Everyday equivalents divide by cited anchors: 12 Wh per phone charge, 0.3 Wh per historical (2009) Google search, 120 Wh per 1-litre kettle boil, 244 gCO₂/km for a typical US car and 106.4 gCO₂/km for an average new EU car, 60 L per shower, 0.25 L per glass and 0.5 L per bottle.
Reviewed by the YouCalc Team · Last reviewed
Spot a translation issue, a calculation issue, or have a suggestion? Let us know.