Methodology
This page explains how MyAIUsage turns your prompt count into estimates for water, electricity, and carbon. The goal isn’t lab precision; it’s a clear sense of scale that helps people make informed choices. We base our numbers on public research, industry talks, and reasonable simplifications.
Assumptions & Formula (short version)
- Water: ~0.0166 litres per text prompt.
- Energy: ~0.003 kWh per text prompt.
- Images: Image prompts are weighted much higher (≈25× a text prompt), reflecting their higher compute demand.
- Carbon: CO₂e is calculated from electricity using a region-specific grid factor (see below).
These are simplified averages. Actual impact varies with model size, hardware efficiency, data-centre cooling, and local energy mix.
What’s included (and not)
- Included: Per-prompt inference energy and water.
- Excluded: Model training, persistent storage, upstream network transit, and device-side energy.
- Focus: Everyday usage rather than one-off, large-scale training runs.
Carbon calculation & regions
Carbon is derived from electricity using a regional grid intensity:
kg CO₂e = kWh × grid factor.
In the calculator, you can pick a region to reflect different electricity mixes.
| Region | Grid factor (kg CO₂e / kWh) | Notes |
|---|---|---|
| UK | 0.233 | Representative recent average; varies by year and time of day. |
| EU | 0.275 | Illustrative EU-wide average; national grids differ. |
| US | 0.386 | Varies strongly by state and utility mix. |
| Global | 0.475 | Broad global average used when region is unknown. |
These are working figures for awareness. For audits, use official, time-resolved intensity for your specific grid/provider.
Everyday comparisons
To make the numbers relatable, we compare results with common activities (e.g., boiling a kettle, brewing coffee, running a dishwasher), and for carbon, an approximate km driven using ~0.171 kg CO₂e per km for a typical small/medium petrol car. These are ballpark comparisons to help with context — not strict equivalences.
Sources we look at
- Academic work on AI/data-centre water and energy (e.g., commonly cited studies on cooling and AI workloads).
- Public talks, blog posts, and whitepapers from cloud and hardware vendors discussing inference energy use.
- Government or utility conversion factors for household activities and grid intensities.
Limitations & updates
- Figures are indicative; local hardware, cooling, and grid mix matter.
- Numbers will evolve as providers publish better data or adopt new infrastructure.
- If you have newer, credible data, please share it — we’ll review and update.
Questions or sources to suggest? Email contact@myaiusage.com.