How do you measure the environmental impact of AI?
Written by SOMA AI, which makes one of the tools named below, so check each claim against the linked sources.
Page updated
Part of the SOMA AI Emission Factors cluster.
How do you measure the environmental impact of AI?
Measure four quantities: energy in kWh, carbon in kg CO2e, water in litres and the lifecycle impact of the hardware. Carbon is energy multiplied by the carbon intensity of the grid, plus a share of the emissions embodied in the hardware. Water is what data centres evaporate to cool servers plus what power stations use to generate the electricity. The method depends on whose computers run the AI. For AI you run yourself you meter it, and for AI you call through an API or a chat product you estimate it from token counts.
What are the four quantities?
- Energy. Electricity used by accelerators, host servers, networking and cooling, in kWh. The ratio of total facility energy to IT energy is the power usage effectiveness, or PUE.
- Carbon. Energy multiplied by the grid carbon intensity of the place and time it was used, in kg CO2e, plus the embodied emissions of the hardware.
- Water. Direct water used for cooling, plus indirect water used to generate the electricity, in litres. The water usage effectiveness, or WUE, is litres of on-site water per kWh of IT energy.
- Hardware and lifecycle. Manufacturing, transport and disposal of chips and servers, spread over their working life. It is rarely visible in a power meter, so it has to be modelled.
How do you measure the AI you run?
Training, fine-tuning and self-hosted inference run on hardware you control, so you can meter them. Record GPU and CPU power draw and runtime, then multiply by the grid carbon intensity. CodeCarbon does this in Python by reading hardware counters where they exist and estimating the rest. A power meter or smart plug at the wall catches what software misses, such as fans and power supplies. If the machines sit in a data centre, multiply by its PUE to include cooling and other overhead. Keep the runtime of idle machines in the count, because a reserved GPU draws power while it waits.
How do you estimate the AI you call?
When you call a model through an API or use a chat product, the provider's machines do the work and no meter is available to you. You estimate instead. Take token counts from the usage field in API responses or from billing exports, apply an energy figure per token for the model class, and apply the carbon intensity of the region that serves the request. Two open tools do this.
- EcoLogits is a Python library that wraps provider SDKs and estimates the impact of each call.
- SAEF, the SOMA AI Emission Factors, is an open dataset under CC BY 4.0 with per-token factors by model class and cloud region. The method is published in Sustainable Production and Consumption 70 (2026) 320-347, and the factor table is on the factors page.
The arithmetic is spelled out in the token CO2e formula, and the free SOMA AI calculator applies it to a usage export. SOMA AI estimates these figures and does not measure them. For chat seats, Copilot and the AI that a company uses as a customer, see how to measure the carbon footprint of the AI you use.
Training versus inference
Training is a one-off cost paid before a model is released, and it is shared across everyone who later uses the model. Inference is paid on every request, so it scales with use and with the length of the output. Output tokens cost more energy than input tokens, because each one is generated in sequence while a prompt is read in parallel. A model that is trained once and used by millions of people can have a larger lifetime footprint in inference than in training. For your own training runs, meter them. For the training share of a model you only call, an emission factor has to allocate it across expected use, as SAEF does.
Which standards and frameworks should you align with?
- GHG Protocol. The corporate standard for inventories. AI bought as a service is usually Scope 3 category 1, purchased goods and services.
- Google's August 2025 methodology for Gemini prompts. It counts full-system power, including idle machines, host CPU and memory, data-centre overhead through PUE and water, and reports 0.24 Wh for a median Gemini Apps text prompt.
- OECD, 2022. The report "Measuring the environmental impacts of artificial intelligence compute and applications" sets out what a measurement framework should cover.
- ITU. The International Telecommunication Union has published work on measuring the environmental impact of AI, including a recommendation on AI lifecycle assessment.
- UNE 0086:2025. A Spanish specification on the sustainability of AI systems. We name it here and do not summarise its content.
What are the common blind spots?
- Closed models. Providers do not disclose hardware, batching or serving region, so any estimate carries a range, and a wrong model class is the largest single error.
- Idle and overhead power. Reserved capacity, host machines and cooling use energy that a per-chip reading leaves out.
- Embodied hardware. Manufacturing emissions do not show on any meter.
- Where the data centre is. The grid that counts is the one at the data centre, not the one at your office.
- Water stress. A litre matters more in a dry basin than in a wet one, so volume alone says little. See water by region.
Published figures also differ because their boundaries differ. The statistics page lists each one with what it counts.
Short answers
- Can you measure the carbon footprint of ChatGPT exactly?
- No. OpenAI does not publish energy per customer or per request, so the footprint of ChatGPT use can only be estimated from your usage, an energy factor for the model class and the grid where it runs. Label the result as an estimate and show a range.
- What tool measures AI energy use?
- CodeCarbon measures energy on hardware you control, for training, fine-tuning and self-hosted inference. For AI you call through an API, EcoLogits and the open SAEF factors estimate it from token counts, because the provider's machines are not yours to meter.
- Does AI use a lot of water?
- It depends on the data centre, its cooling and the grid that powers it, and published figures count different things. The AI water footprint guide on this site explains the three kinds of water and how to read a per-prompt figure. Read the AI water footprint guide.
- Which standard should a company report AI emissions under?
- For AI bought as a service, the GHG Protocol, usually in Scope 3 category 1, purchased goods and services. If you report under the CSRD, the figure feeds ESRS E1, and the ESRS guide on this site covers how to present it. Read the ESRS E1 guide.
Keep reading
- How can I measure the environmental impact of the AI I use? The AI a company uses
- What is the formula to estimate CO2e from LLM API token usage?
- SOMA AI calculator. Free, no sign-up
- SAEF factor table and download
- AI water footprint guide
- How should AI and cloud services be reported under ESRS E1 for CSRD?
