AI Energy Use and Carbon Footprint: What One Answer Can Measure

An AI answer can have a useful energy and emissions estimate without having a complete “carbon footprint.” The distinction is the system boundary: what was measured, what was modeled, and what was left outside the calculation.

The units in an AI energy receipt

For conversion, 1 Wh = 1,000 mWh and 1 kWh = 1,000 Wh. Magicdoor can also convert the provider's Wms energy unit: 3,600,000 watt-milliseconds equals 1 Wh.

Inference time is an input to many estimates. If hardware draws a given amount of power while a model runs, runtime helps estimate energy. It is not enough alone: hardware type, utilization, CPUs, memory, batching, and facility overhead can all change the result.

From runtime to energy and CO2e

GreenPT's published alpha methodology describes a two-step process:

1. estimate GPU and CPU energy from measured inference runtime and hardware assumptions, then include data-center overhead through PUE;
2. estimate emissions using carbon-intensity data for the data-center location and time of inference.

GreenPT says it uses one-hour, data-center-level carbon-intensity data rather than a regional annual average. This means the same prompt can receive a different CO2e estimate at another time because the electricity mix changes. It also means the number remains method-dependent: it combines measured runtime with modeled power and infrastructure factors.

Read the provider's carbon-calculation methodology for the current assumptions and alpha notice.

Where PUE and WUE fit

PUE accounts for energy beyond the IT equipment, such as cooling and power distribution. If a calculation estimates only GPU energy, it misses part of the facility load. PUE helps extend the boundary, although it still does not create a full lifecycle assessment.

WUE relates data-center water consumption to computing energy. It is important to AI water-use analysis, but it is not interchangeable with energy or CO2e. A response receipt that reports energy and emissions does not automatically report water consumption. Water effects also depend on cooling method, location, time, and whether indirect water use from electricity generation is included.

Provider attribution matters

Magicdoor displays GreenPT's provider-reported inference time, energy, and emissions estimates after a response on the optional renewable route. That wording matters because the method, infrastructure inputs, and impact object originate with the inference provider.

The receipt is marked alpha and stored only for the current browser session. Its comparison with a 10 W LED bulb or a published median text-prompt value is a way to understand scale—not a direct test against Magicdoor's other route. The cited median prompt context comes from a 2025 preprint on AI inference energy, whose scope and methodology should travel with any comparison.

What one answer does not measure

A per-answer inference estimate is not a lifecycle assessment. Depending on its stated boundary, it may omit:

• model research, data preparation, and training
• failed experiments and model evaluation
• manufacture and end-of-life of servers and accelerators
• embodied carbon in buildings and supporting equipment
• networking, storage, account systems, and other product operations
• the user's device and connection
• water consumption not included in the reported metric

This does not make the estimate useless. It makes the label important: estimated impact of this inference operation is more accurate than the model's complete carbon footprint.

Compare models and methods carefully

Two receipts are reasonably comparable only when their boundaries and conditions are close. Control or disclose:

• input, context, and output length
• response quality and task success
• model version and serving configuration
• hardware, quantization, and batching
• data-center PUE and calculation method
• location, time, and carbon-intensity source
• whether cached tokens or repeated attempts are included

A short answer from one model and a long answer from another are different workloads. A lower number is useful evidence only if the answer still meets the task and the methods align.

Use the number well

Treat a response estimate as an instrument panel, not a moral score. It can help you shorten unnecessary context, avoid repeated generations, choose an appropriately capable model, and select renewable-powered inference where available. It cannot settle every environmental question in one figure.

Learn the broader claim checklist in what is green AI?, then follow the Magicdoor renewable inference guide.

Try renewable inference

Sources and further reading

• GreenPT carbon-calculation methodology
• GreenPT sustainability overview
• Median AI text-prompt context cited by the Magicdoor receipt
• Scaleway Impact Report 2025