AI Carbon Calculator
Estimate the carbon of AI prompts to a hosted model, or of GPUs you run, in your own region.
Before you start: what to have readyShow
- Prompts a day, from your provider’s usage page, or your GPUs and how long they run.
- The region it runs in, if you know it.
Each one fills in the fields below for a typical workload of that kind. You can change any of them.
- Electricity a year
- 876 kWh
- Per prompt
- 0.082 g CO₂e
- In ca-central-1 instead
- 97% less carbon
Location-based, at 343 g CO₂e per kWh. Same grid as Google's Northern Virginia region (us-east4), whose published figure it uses (2025).
Where the energy per prompt comes from
- Google Gemini, median text prompt: 0.24 WhThe chips, the host CPU and memory, idle capacity and data center overhead. Google gives 0.03 g CO2e, market-based.Google, Measuring the environmental impact of AI inference (arXiv 2508.15734), August 2025
- OpenAI ChatGPT, stated average query: 0.34 WhNot stated: OpenAI published the figure without a method.Sam Altman, The Gentle Singularity, June 2025
- Long reasoning answer, open models: 7.63 to 12.79 WhThe GPU only, measured, for three open models with reasoning on; about 30 times the same models with it off. No server or data center overhead.Hugging Face, AI Energy Score v2, December 2025
Anthropic, Microsoft and most model makers publish no figure per prompt. The figure you choose, highlighted above, is the biggest assumption on this page.
What this model assumes
Show
- Location-based: the grid where the work runs, before any renewable energy the provider buys. Market-based figures, like Google’s 0.03 g per Gemini prompt, are lower.
- Running only. Making the chips and servers, and training the hosted model in the first place, are not counted.
- The method for GPUs is Green Algorithms’ (Lannelongue and others, 2021), with NVIDIA’s maximum power for each GPU.
- The cost side of the same prompts: the LLM Token Cost Calculator.
Want this done on your real bill?
Cloud Cost Audit: We go through your real AWS or Google Cloud bill and hand you a ranked list of what to cut first. From $1,500, fixed. Nothing paid upfront.
Questions
How much carbon does one AI prompt cause?
For a short chat reply, a fraction of a gram. Google measured 0.24 Wh for its median Gemini text prompt, which is about 0.1 g CO2e on an average US grid. Long reasoning answers are the exception: open models measured by Hugging Face use around 30 times more energy with reasoning on, and that is where AI carbon stops being small.
Why is there no figure for Claude, Copilot or my model?
Because their makers have not published one. Choose the closest published figure, or enter your own under My own figure if your vendor gives you one. The page shows what each published figure includes.
How is the GPU figure worked out?
By the Green Algorithms method: GPUs times their power, how busy they are, the hours and the data center overhead (PUE), times the grid carbon of the region. The low end counts the GPUs alone; the high end adds the rest of the server and idle capacity, at the ratio Google measured in its own AI fleet.
Does it count training the model?
For a hosted model, no: training is shared across every user and its makers rarely say what it cost. For your own training or fine-tuning, yes: choose GPUs you run and One run.
Why is my figure higher than the provider’s?
This page is location-based: the grid where the work runs, before the renewable energy the provider buys. Providers often quote market-based figures, which are lower. Google’s 0.03 g per Gemini prompt is market-based.
Related tools
LLM Token Cost Calculator
Compare monthly AI spend on Claude, GPT, Gemini and more before the invoice arrives.
OpenCloud Workload Carbon Calculator
Where you run a workload changes its carbon far more than which cloud you pick.
OpenLLM Provisioned Throughput and Service Tier Planner
PTUs, GSUs or pay as you go: when reserved LLM capacity pays, and what Flex or Batch saves.
OpenMore tools for: Carbon and sustainability, AI and model usage
AI spend grows fastest. Finitizer tracks its cost and carbon together.
Finitizer follows your AI token and GPU spend across clouds and model providers, with the carbon beside it, so the teams using it see both before the bill arrives.
See how Finitizer works