Methodology
Estimating the environmental impact of artificial intelligence is an inexact science. AI companies rarely disclose the exact energy consumption or water usage of their individual data centers.
This calculator uses industry-validated heuristics and peer-reviewed proxies to translate abstract digital actions into tangible environmental footprints. It provides a reliable baseline for consumer awareness rather than an exact server audit.
Built with AI, leanly
We built this tool using Gemini; we supplied the wireframes, the logic, and the raw data. Gemini made sense of it all and coded it up in Astro. We tried to follow efficient prompting rules by working out as much of the specification in advance and putting it into a single prompt, using “medium” reasoning models.
The cost of speccing and building the site was around 5kg of CO2 and 6 litres of water.
Excludes training power
The figures presented do not account for the massive energy required to train the AI models in the first place. For context, GPT-4 was reportedly responsible for more than 6,000 tons of CO2 during its training phase alone.
From prompts to power
We base our calculations on the volume of tokens processed. A token is roughly equivalent to three-quarters of a word. We use a blended average to account for both the prompt you send and the output the model generates.
Standard models For everyday models like Gemini 1.5 Pro or GPT-4o we allocate 0.002340 Watt-hours (Wh) per token. This baseline assumes standard inference workloads on modern GPU architectures.
Reasoning models Models designed for complex problem-solving like DeepSeek-R1 or OpenAI’s o-series use hidden chain-of-thought processing. They evaluate multiple paths before showing you an answer. To account for this invisible compute tax, we apply a heavier multiplier of 0.006240 Wh per token.
Media generation Images and video bypass token counting entirely. Generating media is vastly more energy-intensive than text. We apply a flat rate of 2.50 Wh for a standard image and 15.00 Wh for a short video clip, based on studies measuring the power draw of latent diffusion models.
Data center overhead
Servers do not operate in a vacuum. They require massive amounts of infrastructure to run, primarily for cooling the server racks and managing power distribution.
We apply a Power Usage Effectiveness (PUE) multiplier of 1.56 to all base energy calculations. PUE is the standard metric for data center efficiency. While some tech giants claim PUEs closer to 1.1 for their flagship facilities, 1.56 represents a realistic global average across all cloud infrastructure.
Calculating carbon emissions
Electricity does not have a universal carbon footprint. Its impact depends entirely on how the local grid generates power at that specific moment.
The calculator provides a dropdown to select grid intensity. If you choose the global average, we multiply your total kilowatt-hours (kWh) by 430 grams of CO2 equivalent per kWh.
The water footprint
Data centers require millions of gallons of fresh water to prevent servers from melting, typically used in evaporative cooling towers.
We estimate water consumption using an industry standard proxy of 1.8 liters of fresh water consumed per kWh of server energy. This figure represents the direct operational cooling cost and excludes the embedded water footprint of manufacturing the server hardware.
Primary sources and further reading
Frugal AI The mack daddy. Frugal AI switchboard
Power Hungry Processing A foundational study on the energy footprint of generative AI by Sasha Luccioni et al. 2024. Read the paper on arXiv
Making AI Less Thirsty Research detailing the spatial footprint and water consumption of large language models by Pengfei Li et al. 2025. Read the paper on arXiv
Uptime Institute Global Data Center Survey The source for global average PUE metrics across enterprise and hyperscale facilities. View the Uptime Institute research
W3C Web Sustainability Guidelines The international framework for reducing the environmental impact of digital products and services. Review the WSG drafts
Epoch AI Loads of AI / power data here. Epoch.AI
Carbon equivalencies
To make abstract figures like kilowatt-hours understandable, the calculator maps your estimated emissions against familiar physical activities.
The engine evaluates your total footprint and selects the closest real-world proxy where the multiplier sits between 1 and 100 (targeting a human-friendly scale near 10). The table below lists the baseline emissions proxies used across each tier:
| Equivalent Activity | Base CO2e | Format & Rate | Proxy Context & Assumptions |
|---|---|---|---|
| 50g lump of coal | 0.12 g | Discrete count | Embodied emissions of domestic bituminous coal. |
| 100W incandescent bulb | 1.0 g | Rate (1.4 mins/g) | Based on drawing 100W on an average European/global grid mix. |
| Human breathing | 1.0 g | Rate (90 secs/g) | Average human exhalation (~1 kg CO2/day at rest). |
| Streaming mobile video | 1.0 g | Rate (8 secs/g) | Cellular data transmission and data center stream decoding. |
| Standard email | 10.0 g | Discrete count | Average blended text email across client and server delivery. |
| Driving a car (short distance) | 10.0 g | Rate (50 m/10g) | Average petrol combustion engine emissions (~200 g/km). |
| Streaming video (HD) | 50.0 g | Rate (15 mins/50g) | 1080p stream over fixed-line broadband to a smart television. |
| Watching television | 50.0 g | Rate (2 hours/50g) | Modern 50-inch LED television operating at standard brightness (~50W). |
| Electric train journey | 50.0 g | Rate (3.5 miles/50g) | National rail average passenger-kilometre emissions under mixed traction. |
| Cloud storage (1GB) | 100.0 g | Discrete count | Storing 1GB of data in active multi-region cloud storage for one year. |
| Large dairy latte | 200.0 g | Discrete count | Farm-to-cup lifecycle footprint, driven primarily by dairy milk production. |
| Washing machine cycle | 200.0 g | Discrete count | Standard 45-minute cycle at 30°C on a modern energy-efficient appliance. |
| Driving a petrol car (1 mile) | 400.0 g | Rate (1 mile/400g) | Tailpipe and fuel lifecycle emissions for a medium petrol passenger vehicle. |
| Hot shower (5 minutes) | 1,000.0 g (1 kg) | Discrete count | Domestic gas or electric boiler heating ~35 litres of water. |
| Glass of wine | 1,000.0 g (1 kg) | Scaled (3 glasses/kg) | Viticulture, bottling, transportation, and glass manufacturing overhead. |
| New paperback book | 1,000.0 g (1 kg) | Discrete count | Paper sourcing, ink production, binding, and distribution logistics. |
| Beef burger | 2,500.0 g (2.5 kg) | Discrete count | Livestock agriculture, land use, feed production, and processing emissions. |
| Driving an average car (300 km) | 100,000.0 g (100 kg) | Dual distance (300 km / 186 mi) | Long-distance passenger car journey lifecycle emissions. |
| Transatlantic flight | 1,000,000.0 g (1 tonne) | Discrete count | Economy passenger one-way flight between London and New York, including radiative forcing. |
| Powering a home | 1,000,000.0 g (1 tonne) | Rate (1.5 years/tonne) | Typical renewable-leaning household electricity consumption profile. |
Lifestyle baseline context
To give these numbers context, the calculator compares your total against a 20-tonne annual lifestyle footprint (20,000 kg CO2e per person per year). This reflects the typical per-capita lifestyle carbon intensity of high-consumption Western economies.
For users tracking monthly usage, the target is scaled down to its equivalent monthly share (~1,666.7 kg/month) so comparisons remain strictly proportional.