Methodology
How this calculator estimates energy, water and carbon per AI operation and where the numbers come from.
What's measured vs. estimated
Measured: Absolute scale comes directly from first-party operator data. Google and OpenAI's 2025 disclosures put a typical text prompt at roughly 0.3 Wh of energy.
Estimated: Relative differences between models (why Model A shows more energy than Model B) are informed assumptions. No AI provider currently publishes per-model breakdowns, so the relative spread reflects general expectations about model size and architecture, not disclosed figures.
The prompt calculator scales with your typed prompt's length, not the model's response length; in practice, output length is the larger driver of real inference cost. Treat all results as order-of-magnitude illustrations, not precise measurements.
Every result carries this disclaimer for a reason: these are estimates for awareness and comparison, not metered actuals or an assurance opinion.
The absolute-scale anchor
Three independent first-party disclosures from 2025 converge on roughly 0.3 Wh for a typical AI text query: Google reports a median Gemini Apps prompt at 0.24 Wh, OpenAI reports an average ChatGPT query at 0.34 Wh, and Microsoft reports a median Copilot query at 0.31 Wh. This calculator uses the average of the three (0.3 Wh) as its energy anchor, replacing a widely cited but outdated 2023 estimate of roughly 2.9 Wh that assumed since-superseded hardware.
Water intensity is derived the same way: Google and OpenAI's disclosures imply roughly 1.0 litre of water per kilowatt-hour of AI compute, covering the full production stack (cooling plus the indirect water used to generate electricity), not just on-site cooling efficiency.
Carbon is calculated from energy using the US average grid intensity of 0.367 kg CO₂e/kWh (EPA eGRID), with an optional toggle to include an amortised estimate of embodied hardware-lifecycle carbon (manufacturing, transport, disposal) at 100 g CO₂e/kWh.
Sources
- Google, Gemini Apps text prompt (median): 0.24 Wh, 0.03 g CO₂e, 0.26 mL water
- OpenAI, average ChatGPT query: 0.34 Wh, ~0.32 mL water (Sam Altman, June 2025)
- Epoch AI, independent analysis of ChatGPT energy use
- EPA eGRID, US average grid carbon intensity
- US EIA, grid and electricity data
- Data centre water usage overview
Need this for organisational or compliance reporting?
If you need a defensible, audit-ready line item for your organisation's AI usage (classified by deployment mode: self-hosted vs. purchased AI services, mapped to NGER / GHG Protocol / AASB S2), see aixhuman for Business. Australian data centre operators can also generate NGER-aligned facility reports at dcreg.online.