Training emissions, 4 published assessments
Training emissions, in tCO2e (logarithmic scale)
Training emissions, 4 published assessments
Sources
Source: Morrison et al., 2025; Patterson et al., 2021; Meta, 2024. projetcelsius.com/en/ai · CC BY 4.0- Allen Institute for AIMorrison et al., Holistically Evaluating the Environmental Impact of Creating Language Models, 2025March 2025 (new tab)
- Google and UC BerkeleyPatterson et al., Carbon Emissions and Large Neural Network Training, 2021April 2021 (new tab)
- MetaLlama 3.1 model cardJuly 2024 (new tab)
The analysis
- The trials before the final model weigh half of the final training at OLMo: a cost that other providers do not publish.
- Llama 3.1 emits 20 times GPT-3: the models are larger and trained for longer.
- Training is only one part: for Mistral Large 2, training and 18 months of use add up to 20,400 tCO2e.
The figures
- OLMo, trial runs (development, 2025): 159 t.
- OLMo, training (final training, 2025): 312 t.
- GPT-3 (estimate, 2020): 552 t.
- Llama 3.1 (3 sizes, 2024): 11,390 t.
Assumptions
- Training electricity, unless stated otherwise; hardware manufacturing not counted.
- Large closed models (GPT-4 and later, Gemini, Claude) do not publish their training assessment.
- Different methods and electricity grids from one assessment to another.
Continue with hardware
The whole Hardware topic →Charts and maps
Chart
HardwareThe footprint of digital devices99%of a smartphone's footprint comes from its manufactureChartHardwareElectronic waste and AI22.3%of 2022 electronic waste collected and recycledChartHardwareElectricity for manufacturing AI chips+351%of electricity to manufacture AI chips, in one yearChartHardwareLife cycle assessment of Mistral Large 261%for materials, the manufacture of servers and chips
HardwareThe footprint of digital devices99%of a smartphone's footprint comes from its manufactureChartHardwareElectronic waste and AI22.3%of 2022 electronic waste collected and recycledChartHardwareElectricity for manufacturing AI chips+351%of electricity to manufacture AI chips, in one yearChartHardwareLife cycle assessment of Mistral Large 261%for materials, the manufacture of servers and chips
Climate
WaterAcross the hub
AI Impact DatabaseAll the figures, their methodology and their sources.
AI Ranking10 assistants rated out of 100: performance, ethics and environment.
AnalysesThe signed pieces: analyses, viewpoints, "Where does this figure come from?".
Charts and mapsYou are hereThe reference visuals, with their data and sources.
ToolsOpen calculations, your assumptions on the left, the result on the right.
A question about a figure?
An error, a source to add, a methodology question, a press request: write to us. Guillaume Pakula, Clément Reynaud, Sébastien Pierfederici and Alexandre Veilly will reply.
Write to us →
Projet Celsius, which runs this hub, also helps companies measure and reduce the footprint of their AI use.
Accounting for AI in your Bilan Carbone® ↗
