Google and Mistral report vastly different carbon footprints for AI prompts, with accounting methods explaining most of the 40x difference. The gap highlights the need for standardized carbon accounting in AI.
TL;DR
- Google and Mistral report vastly different carbon footprints for AI prompts, with accounting methods explaining most of the 40x difference.
- The gap highlights the need for standardized carbon accounting in AI.
- AI tasks like image and video generation have significantly higher carbon footprints than text prompts.
What happened
Google reports a median carbon footprint of 0.03 grams of CO2e per Gemini prompt, while Mistral reports 1.14 grams of CO2e for a Le Chat reply. The discrepancy arises from different accounting methods and what each company includes in their measurements.
Google's measurement is market-based, crediting the company for clean energy purchases, while Mistral's measurement is location-based, using the average emissions of the grid where the electricity is consumed. This alone can triple a figure.
Other factors contributing to the gap include model size, reply length, and the boundary of the measurement. For example, Google's measurement includes only active AI chips, while Mistral's includes hardware manufacturing and the entire supply chain.
Why it matters
The 40x difference in carbon footprint measurements highlights the need for standardized carbon accounting in AI. Without a consistent method, it's difficult to compare the environmental impact of different AI models and services.
For developers and startups, understanding the carbon footprint of AI tasks is crucial for making informed decisions about sustainability. Investors can also use this information to evaluate the environmental impact of their portfolio companies.
The carbon footprint of AI tasks varies significantly. For instance, a short AI video can emit anywhere from under a tenth of a gram to more than 140 grams of CO2e, while a text prompt emits about 0.1 grams. This information can help users make more sustainable choices when using AI services.
Key facts
- Google's median Gemini prompt emits 0.03 grams of CO2e (market-based accounting).
- Mistral's Le Chat reply emits 1.14 grams of CO2e (location-based accounting).
- Google's 2025 Scope 2 emissions: 2,815,000 tonnes market-based and 15,148,700 tonnes location-based.
- A long reasoning answer emits about 13 times more than a normal prompt.
- One AI image emits about 1 gram of CO2e on average.
- AI video generation has the widest carbon range, from under a tenth of a gram to more than 140 grams for one short clip.
- Training one large AI model emits anywhere from 25 tonnes to an estimated 150,000 tonnes of CO2e.
Context
The carbon footprint of AI is a growing concern as the technology becomes more widespread. AI models require significant amounts of electricity to train and run, contributing to greenhouse gas emissions.
Different companies use different methods to measure and report their carbon footprints, making it difficult to compare the environmental impact of different AI services. Standardized carbon accounting could help address this issue.
As AI becomes more integrated into our daily lives, understanding its environmental impact is crucial. Developers, startups, and investors all have a role to play in promoting sustainability in AI.