What is the carbon footprint of AI?
An AI carbon footprint is the greenhouse gas emitted by running and training AI models, from the electricity used and from making the hardware.
Why it matters
The footprint of an AI workload has three main parts. Electricity at run time is multiplied by the carbon intensity of the local grid at the hour it runs. Embodied emissions come from making the GPUs and servers, spread over their working life, and count as Scope 3. A share of the energy used to train the model is added on top.
Organisations use AI through cloud services, direct APIs and AI tools built into SaaS products. Each reports usage as tokens or spend in its own way, and few report energy or carbon at all. Comparing or adding up those footprints needs one method applied at model and token level.
An example
The same request can have a very different footprint depending on where it runs. On a grid powered by wind and nuclear, the electricity part is small. Run the identical request in a region that relies on gas or coal and the carbon rises, even though the energy used is the same. Model choice also changes the result, because a large model uses more energy for each token than a smaller one.
How Greenpixie measures it
GPX AI estimates energy, carbon and water per token for over 500 models. Greenpixie benchmarked 32 open-weights LLMs on NVIDIA H100 and B200 GPUs, then extended the results to other models with predictive models. Each carbon figure includes:
- data centre overhead, through PUE
- a sampled range of training energy, spread across the tokens a model serves
- cradle-to-grave embodied emissions of the server hardware
- grid carbon intensity for the region and hour of the request
Every value is a median with an interquartile range. Any usage report that shows tokens by model and token type can be used, including cloud billing data, API logs and per-seat AI tools.