24 August 2025, updated 25 August 2025
The big AI platforms are starting to disclose some energy and resource use and emissions. When comparing different platforms, treat these as indications rather than absolute measurements because every platform is going to measure things with different methods. Anyway:
| Energy Wh | Emissions gCO2e | Water mL | Resources mg of Sb eq | |
|---|---|---|---|---|
| Google Gemini “median Gemini Apps text prompt” | 0.24 down “33x” in 1y “watching tv for < 9 seconds” | 0.03 down “44x” in 1y | 0.26 | not disclosed |
| OpenAI ChatGPT “average query” | 0.34 | not disclosed | 0.32 “0.000085 gallons of water” | not disclosed |
| Mistral Large 2/Le Chat “400-token response” | not disclosed | 1.14 “watching online streaming for 10 seconds” | 45 “growing a small pink radish” | 0.16 “producing a 2 euro cent coin” |
This is progress, but as the Google debate shows (see links below), we need a standard, independently-verifiable (and verified) methodology for measuring impact. Until then, both good and bad numbers can be tainted by PR and what-iffing.
Sources and methodologies:
- Google: Our approach to energy innovation and AI’s environmental footprint and Measuring the environmental impact of AI inference and Measuring the environmental impact of delivering AI at Google Scale (pdf). The Google paper has a good overview of previous disclosures and independent emissions modelling efforts. In section 3.2 Energy Measurement Methodology they say their method is more “comprehensive” because it includes idle machines, campus PUE etc, but that (I paraphrase) if they demeaned themselves by using everyone else’s half-hearted approach their numbers would be better at 0.10 Wh energy, 0.02 gCO2e emissions, and 0.12 mL water per prompt. OTOH some wonder if Google is under-reporting water use or unhelpfully using median numbers and market-based emissions metrics, so there’s plenty of debate about it.
- OpenAI: haven’t disclosed a report, this is an aside in a terrifyingly breezy June 2025 blog post from Sam Altman The Gentle Singularity.
- Mistral: A path toward a global environmental standard for AI, which also says training and 18 months of usage had these impacts: 20.4 ktCO2e emissions, 281,000 m3 water, 660 kg SB eq. Scope: “use of our AI assistant Le Chat for a 400-token response”, excluding users’ terminals, 18 months of use.
- Elsewhere: I need to see if Sasha Luccioni or this point at any other disclosures.
Previously:
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.