How to Measure Your AI Carbon Emissions
On this page
You cannot manage what you cannot see, and AI emissions are easy to ignore because they happen inside a data center far away. The good news is that the math is approachable, and once you learn it you can measure AI emissions for a single request or a whole month.
This guide walks through the tokens to energy to carbon method, shows how to use a calculator, and explains how per-request tracking turns a fuzzy worry into a number you can act on. The figures here are estimates that vary by model and region, but they are close enough to guide real decisions.
The basic chain: tokens to energy to carbon
Every AI request can be traced through a short chain. The model processes tokens, tokens consume energy, energy produces carbon based on the local grid, and that energy also carries a water cost for cooling. Once you see the chain, measurement is mostly arithmetic.
Per-token energy depends on model size. A small model uses roughly 0.0008 Wh per token, a mid model around 0.0015 Wh, a frontier model about 0.0038 Wh, and a reasoning model near 0.0040 Wh. Image generation is a separate bucket at roughly 2.0 to 3.5 Wh per image. For a deeper look at the electricity side, see how much electricity AI uses.
Do the math step by step
Start by estimating tokens. A normal answer might involve a few hundred to a couple thousand tokens across your prompt and the reply. Multiply that by the per-token energy for your model to get watt-hours, then apply a data center overhead factor, since real facilities use extra energy for cooling and power delivery at a PUE of about 1.56.
Next convert energy to carbon and water. The US grid averages around 0.395 kg CO2 per kWh, and water runs about 3.4 L per kWh. Do the multiplication and a typical text answer lands under a gram to a few grams of CO2. It sounds tiny, but it adds up across thousands of requests.
- Estimate total tokens for the prompt and response
- Multiply tokens by per-token energy for your model
- Apply a PUE overhead of about 1.56
- Multiply kWh by 0.395 for carbon and 3.4 for water
Use a calculator to skip the arithmetic
Doing the math by hand is a useful way to build intuition, but you do not want to repeat it for every request. A calculator handles the conversions and lets you compare models instantly, so you can see how switching from a frontier model to a smaller one changes the result.
The AI carbon footprint calculator takes your inputs and returns energy, carbon, and water in one view. It is the fastest way to answer questions like how much a long document summary costs compared with a quick factual query.
Track emissions per request
Manual estimates are fine for spot checks, but continuous tracking is where the value lives. When each request is measured as it happens, you get an accurate running total instead of a rough guess, and you can spot which habits drive the most impact.
Ecoia measures every request and records energy, carbon, and water automatically, which removes the guesswork entirely. If you build software, the carbon-tracked AI API returns the same data with each call so you can log it. Teams reporting under Scope 3 AI emissions can feed those numbers straight into their accounting.
Turn the numbers into action
Measurement is only useful if it changes something. Once you can see per-request impact, the obvious moves appear: pick a smaller model when accuracy allows, write tighter prompts, and cut needless regenerations. Each of those shows up immediately in the numbers.
The final step is to offset what remains. Measuring first means you offset a real figure rather than a guess, which is exactly the order recommended in how to offset AI carbon emissions.
- Switch to a smaller model where accuracy allows
- Tighten prompts to cut tokens on both sides
- Avoid regenerating answers out of reflex
- Offset the remaining measured impact
The headline: Measure AI emissions by tracing tokens to energy to carbon, use a calculator to skip the arithmetic, and track per request so your totals reflect reality.
FAQ
How much CO2 does a single AI text answer produce?
A typical text answer produces somewhere under a gram to a few grams of CO2, depending on the model, the length, and the grid. Larger reasoning models and longer responses sit at the higher end. The number is small per request but meaningful across thousands of them.
Why does the data center overhead factor matter?
Servers are not the only thing drawing power in a data center. Cooling, power conversion, and other systems add overhead, captured by a PUE of about 1.56. Multiplying raw compute energy by that factor gives a more honest total than counting the chips alone.
Do I need to count water separately?
Water tracks closely with energy because it is mostly used for cooling. A common estimate is about 3.4 L per kWh, so once you know energy use you can derive water with a single multiplication. It is worth reporting alongside carbon for a full picture.
Can I measure emissions automatically instead of by hand?
Yes. Platforms that record impact per request remove the need for manual math, and a carbon-tracked API can return energy, carbon, and water with every call. Automatic tracking is more accurate than periodic estimates because it reflects your actual usage.