Version 2026-07.1
Methodology
How we estimate the environmental cost of AI inference, and how we translate it into a donation amount. Every constant is cited; every formula is auditable.
The formula
kWh = (output_tokens + input_tokens × 0.1) × Wh/token × PUE / 1000
kg CO₂e = kWh × grid_intensity
liters H₂O = kWh × water_intensity
damage_$ = (kg CO₂e / 1000) × social_cost_of_carbon
donation = damage_$ × multiplier
Constants
| Constant | Value | Source |
|---|---|---|
| PUE (data center overhead) | 1.2 | Uptime Institute Global Survey 2023 |
| Grid intensity | 0.38 kg CO₂e/kWh | US EPA eGRID national average 2022 |
| Water intensity | 0.55 L/kWh | Li et al. 2023, "Making AI Less Thirsty" |
| Social cost of carbon | $190/t CO₂e | US EPA 2023 SC-GHG report, central estimate |
| Input token energy fraction | 0.1× | Prefill ≪ decode; Luccioni et al. 2024 |
Energy per token by model class
| Class | Wh / output token | Blended $/MTok (for spend fallback) |
|---|---|---|
| small | 0.0002 | $0.6 |
| medium | 0.0006 | $5 |
| large | 0.0025 | $15 |
| frontier | 0.006 | $40 |
Limitations
- These are estimates, not measurements. Actual energy use varies by data center location, cooling method, utilization, and hardware generation.
- We use a US-average grid intensity. Many large providers use significant renewable energy, which would reduce real-world emissions.
- Subscription tier estimates are derived from published average-usage studies and are inherently approximate.
- The social cost of carbon is contested. $190/t is the EPA 2023 central estimate; ranges from ~$50 to $400+ depending on discount rate and model.
- The default 2× multiplier is designed to overshoot under any reasonable uncertainty range in the underlying estimates. We suggest the donation amount; you choose whether to pay it.
Citations
- Luccioni, A., Jernite, Y., & Strubell, E. (2024). Power Hungry Processing: Watts Driving the Cost of AI Deployment? ACM FAccT 2024.
- Epoch AI (2025). How much energy does ChatGPT use?
- US EPA (2023). Report on the Social Cost of Greenhouse Gases.
- US EPA eGRID (2022). Emissions & Generation Resource Integrated Database.
- Uptime Institute (2023). Global Data Center Survey.
- Li, P. et al. (2023). Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models.