Entraînement d'un modèle

Name: Training a model
Creator: Projet Celsius
License: https://creativecommons.org/licenses/by/4.0/

Projet Celsius

Database entryHardwarepublished as isOutside the levels infrastructureInfrastructure, outside the levelsInfrastructure18 entriesFigures published by others, reproduced as they are; they cut across the territories without adding to them.Database structure

Training a model

What it cost to train the models whose provider or researchers published the footprint, one model per row.

Selected reference: GPT-3, 2020, source Patterson et al.
The rows describe the training of different models, not the annual change of a single model.
Out of scope: the trials that precede the final training and the use of the model, unless noted on the row.

Compare

Origin of the value

OpenAI, Anthropic and Google publish nothing for GPT-4, Claude or Gemini; the figures for GPT-3 and GPT-4 are researchers' estimates.

Characteristics

Unit	1 training run : GPT-3, 2020
Location	not applicable: a published total does not depend on the location of the servers
Counted	the electricity of the final training, and hardware manufacturing when the row says so
Not counted	the trials that precede the final training and the use of the model, unless noted on the row
Status	published, 1 reading of the same scope, 2 other published values

All the characteristics identifier, version, rule for the value, sources, licence

Identifier	IA-0104, projetcelsius.com/en/ai/database/IA-0104/
Version	0.3 of 2 November 2026
Category	Infrastructure, published as is
Model	GPT-3, 2020 (552 tCO2e), Llama 3.1, three sizes, 2024 (11.4 ktCO2e), Mistral Large 2, training and 18 months of use (20.4 ktCO2e), GPT-4, 2023, researchers' estimate, manufacturing included (15 ktCO2e), BLOOM, 2022, trained in France, manufacturing included (50.5 tCO2e), Llama 4 Scout and Maverick, 2025 (2 ktCO2e), Google's Gemma 2, three sizes, 2024 (1.25 ktCO2e), The OLMo family, final trainings, 2025 (312 tCO2e)
Choice of the selected value	value taken as it stands from its source, without calculation
Sources	Patterson et al., 2021
Licence	CC BY 4.0

Sources of the calculation

1 publication, for this entry and those used to calculate it, with the chosen parameters.

Energy and factors

The references of the selected values and of the entries used in this calculation.

Patterson et al. · April 2021
Carbon Emissions and Large Neural Network Training ↗ (new tab)
Training a model
Role and limits : Patterson et al.
What the publication covers. electricity and carbon of training
What it does not cover. manufacturing
Reading by the database. External estimate
- Training a model →
  Reference for the variant "GPT-3, 2020"
  Scope of the figure used. GPT-3, 2020, average US electricity; 552 tCO2e
  Limit. An external estimate, OpenAI published nothing

Published values

11 figures for the same quantity, each in its own scope.

Electricity

1.29 GWh selected, between 433 MWh and 32.8 GWh for published values of the same scope

Green: same scope as the selected value (measured in dark, estimated or published in light). Grey: other scope. Empty circle: check by the database or discarded figure. Logarithmic scale.

Source	Date	Electricity	What the figure measures	Scope	Status
Training a modelselected by the database, v0.3	2026	1.29 GWh	the value selected by the database	same scope	published
Morand, Névéol and Ligozat ↗The Environmental Impacts of Language Model Training Keep Rising	Oct. 2025	32.8 GWh	GPT-4, 57 million A100 GPU hours, data centre included (PUE of 1.2) · A researchers' estimate: OpenAI published nothing	same scope	estimated
Luccioni, Viguier and Ligozat ↗Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model	Nov. 2022	433 MWh	BLOOM, 176 billion parameters, trained on the Jean Zay supercomputer, in France · An open model; the same table gives 324 MWh for OPT and 1,066 MWh for Gopher	same scope	measured
Patterson et al. ↗Carbon Emissions and Large Neural Network Training	2021	1.29 GWh	GPT-3, 2020, average US electricity; 552 tCO2e · An external estimate, OpenAI published nothing	same scope	estimated

CO2e

Source	Date	CO2e	What the figure measures	Scope	Status
Allen Institute for AI, Morrison et al. ↗Holistically Evaluating the Environmental Impact of Creating Language Models	March 2025	312 tCO2e	OLMo family, from 20 million to 13 billion parameters, final trainings · Development trials add 159 tCO2e, and manufacturing the chips 22 tCO2e	same scope	measured
Google, Gemma team ↗Gemma 2: Improving Open Language Models at a Practical Size	2024	1.25 ktCO2e	Gemma 2, three sizes (2, 9 and 27 billion parameters), pretraining on Google's chips, data centre overhead included · Without saying whether electricity is counted by grid or by contract	same scope	published
Meta ↗Llama 4 model card	2025	2 ktCO2e	Llama 4 Scout and Maverick, 7.38 million H100 GPU hours, local grid electricity · Zero according to Meta's electricity contracts; "tons" without further detail in the model card	same scope	published
Meta ↗Llama 3.1 model card	2024	11.4 ktCO2e	Llama 3.1, three sizes, 39.3 million GPU hours, location-based · Published by Meta, electricity not published	same scope	published
Luccioni, Viguier and Ligozat ↗Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model	Nov. 2022	50.5 tCO2e	BLOOM, French electricity (57 gCO2e/kWh), hardware manufacturing (11.2 t) and idle consumption (14.6 t) included · 24.7 tCO2e for the electricity of computing alone	other scope	estimated
Morand, Névéol and Ligozat ↗The Environmental Impacts of Language Model Training Keep Rising	Oct. 2025	15 ktCO2e	GPT-4, average US electricity: 10,200 t for server use, 2,000 t for infrastructure, 3,300 t for manufacturing · Hardware manufacturing included	other scope	estimated
Mistral AI, LCA of Large 2 ↗Our contribution to a global environmental standard for AI	Jul. 2025	20.4 ktCO2e	Training and 18 months of use, life cycle · Another scope: use and manufacturing are included	other scope	estimated

Water

Published by its sources, not yet selected by the database: see the convention on water in Methodology and conventions.

Source	Date	Water	What the figure measures	Scope	Status
Mistral AI, LCA of Large 2 ↗Our contribution to a global environmental standard for AI	Jul. 2025	281 Mm³	Mistral Large 2, training and 18 months of use, life cycle assessment: water consumption, server manufacturing included · Another scope: the study does not separate training from use	other scope	estimated

The published figures do not all count the same scope; the scope of each figure is written next to it. Why do the estimates diverge?

Methodology conventions

With their alternatives

The full methodology

The rule	a published total is reproduced as is, with its year and scope; it is not converted with an average factor and does not add to the database's calculations
Calculate a case without an entry	The formula, the default values and a step-by-step example

Uses in the hub

Charts, figures, analyses and articles that cite this entry

ChartTraining emissions, 4 published footprints →
ChartLife cycle assessment of Mistral Large 2 →
Blog articleHow and why does AI pollute? ↗

History

Entry in version 0.3 of 2 November 2026, in the database in version 0.4

The changelog

2 November 2026Designation : The training of a model becomes Training a model.More precise designation: the object first, the specification after the comma, the quantity or period at the end; the former title remains searchable.
1st November 2026Rows : 3 models becomes 8 models.Five published training footprints added, one per row: GPT-4, BLOOM, Llama 4, Gemma 2 and OLMo.
31 October 2026Creation of the database, in its first version (0.1).

Citation

Projet Celsius, AI Impact Database, version 0.4 of 13 November 2026, "Training a model", IA-0104 (entry at version 0.3 of 2 November 2026), projetcelsius.com/en/ai/database/IA-0104/ · CC BY 4.0.

Download

JSON Markdown