How estimates are calculated
AI providers don't publish per-query energy data. These estimates are derived from published academic benchmarks and research papers, adjusted for your local grid's carbon intensity.
Key variables
Model architecture
GPT-4 class models use ~3–10× more energy per token than smaller models like Haiku or GPT-4o mini.
Task type
Image generation uses 100–1,000× more energy than a text query. Video generation uses 10–50× more than images.
Grid carbon intensity
The same computation on Norwegian hydro power (26 gCO₂/kWh) emits ~18× less than on the global average grid (475 gCO₂/kWh).
Server load variance
Energy per query can vary 3–5× depending on server utilisation. Estimates use midpoint assumptions.
Ranges explained
Each result shows a min–max range. The midpoint is your headline figure. The spread reflects real-world variance in data centre efficiency, model optimisations, and undisclosed infrastructure differences between providers.
⚠ These are estimates, not measurements. Actual emissions depend on undisclosed server infrastructure. Treat the figures as directionally correct — useful for comparison and awareness, not for formal carbon reporting.
Sources: Patterson et al. (2021) · Luccioni et al. (2023)
IEA Data Centres report (2024) · ElectricityMaps methodology
Strubell et al. (2019) · Goldman et al. (2024)