See how much energy, carbon and water your team's AI use adds up to, and how much the model you pick changes it.
Each model has its own energy use per input and output token, from Greenpixie’s GPU benchmarks. It includes data centre overhead and a share of training.
Research, documents, coding and agents send and receive very different amounts of text. Each task uses a typical prompt size for that kind of work, and your prompts per day count for each use you pick. Totals assume 250 working days a year.
Energy is turned into carbon using the latest full month of grid data, averaged across cloud regions, plus a share of hardware manufacturing. Water covers data centre cooling and power generation.
The methodology has been independently verified in accordance with ISO 14064-3, prepared with reference to ISO 14064-1, and designed to be in line with the GHG Protocol Corporate Standard. Read the full AI methodology.
To get a reliable picture of AI energy use, we ran our own benchmarks across a wide range of models and workloads. The AI methodology docs set out every step in full.
32 open-weights LLMs, from 1.7 billion to 1 trillion parameters, run quantised (FP8, INT8, NVFP4) on two generations of current GPUs, the way hyperscalers deploy them.
Standard prompts from 14 public benchmark datasets, from coding to summarisation, with prompts up to about 14,000 tokens and 16 to 2,048 requests running in parallel.
GPU energy measured directly and the rest of the server estimated from inference time, with the prefill, overlap and decode phases separated.
A share of the energy used to train each model, embodied emissions from manufacturing the hardware, and a scaling factor where prompts are cached.
Regression models predict energy and inference time from a small set of model characteristics, so one method covers small and very large models.
Closed models are placed in a size band based on the tier in their name, and the details that can’t be observed are sampled across realistic ranges.
A typical chat prompt to a large frontier model such as GPT-6.1 Sol uses about 1.6 Wh of electricity, about 8 mL of water and about 0.7 g CO₂e. It sounds small, but it adds up fast across a team, and some models and tasks use many times more. Our guide to AI energy consumption explains what drives it.
| One chat prompt to | Electricity | Carbon | Water |
|---|---|---|---|
| GPT-6 Astra | 7.0 Wh | 3.3 g CO₂e | 34 mL |
| GPT-6.1 Sol | 1.6 Wh | 0.74 g CO₂e | 8.1 mL |
| GPT-6 Luna | 0.19 Wh | 0.08 g CO₂e | 0.95 mL |
Greenpixie is an enterprise data company that helps organisations understand and cut the environmental impact of their cloud use. We combine cost, carbon, energy and water into one consistent method that teams can use across regions and providers. AI is the next frontier of cloud use, and we hold it to the same standard as the rest of the cloud estate.
Having been involved in the Sustainable AI Working Group and seen the methodology up close, it’s clear Greenpixie are setting the gold standard for how enterprises can measure the energy, carbon and water impact of AI workloads.
Mark BradleySenior Manager, Product, FlexeraThe methodology strikes the right balance between scientific rigour and enterprise usability. It gives organisations a credible way to understand and optimise the sustainability impact of AI, including consumption of water, kWh and carbon.
Andy WestbrookDirector, Cloud Transformation & FinOps, NTT DataQuick answers on AI energy, water and how this calculator works.
Beyond the estimate
Every major model, your actual usage, and the region and hour each request runs in. One method across providers, SaaS tools and agents.