Réentraînement d'un modèle, 1 an

Name: Retraining a model, 1 year
Creator: Projet Celsius
License: https://creativecommons.org/licenses/by/4.0/

Projet Celsius

Database entryUsesLevel 5 of 6 organisationOrganisation, level 5 of 6Calculated with 3 entries, level 1Database structure

Retraining a model, 1 year

4 retrainings a year of an open model on the organisation's data, of 24 hours each on 8 H100 GPUs. Projet Celsius estimate.

"H100 hour" accounts for 91% of emissions in this scenario.
100% of the calculation rests on database assumptions; it does not measure actual use.
Out of scope: the model's initial training by its publisher, data preparation and hardware manufacturing.

Compare

Calculated with

Breakdown of emissions by item

350 kgCO2e in total: 760 kWh of electricity × 458 gCO2e/kWh (servers worldwide).

Inputs of the calculation 2

PUE of the data centre, H100 hour

How the calculation is built

A published order of magnitude: a model of 65 billion parameters is fine-tuned in 24 hours on a single GPU.

Comparison with other organisations

See the full list

Automated customer service28 kgCO2e

Document assistant34 kgCO2e

Meeting transcription36 kgCO2e

Licences for everyone130 kgCO2e

Mid-cap company of 500 people150 kgCO2e

Retraining a modelthis entry350 kgCO2e

Large company (5,000)1.4 tCO2e

Coding agents, one team1.8 tCO2e

AI feature of a product3.7 tCO2e

Hosted model4.8 tCO2e

Logarithmic axis: each tick is 10 times the previous one.

Characteristics

Unit	1 service, 1 year : 24 agent hours a day
Default location	world, 458 gCO2e/kWh (Ember 2025, life cycle); 11 other locations to choose from
Counted	the electricity of the GPUs during retraining, and the data centre overhead
Not counted	the model's initial training by its publisher, data preparation and hardware manufacturing
Status	estimated

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

Identifier	IA-0416, projetcelsius.com/en/ai/database/IA-0416/
Version	0.1 of 2 November 2026
Category	Organisations, one organisation, one year
Choice of the selected value	database calculation from the entries this one is made of, at the quantities of the unit
Sources	calculation of the database
Licence	CC BY 4.0

Emissions by server location

Emissions associated with electricity (CO2e)

Poland450 kgCO2e
World350 kgCO2e
United States290 kgCO2e
European Union160 kgCO2e
France32 kgCO2e

× 14 between Poland and France, for the same electricity

Ember 2025, over the life cycle. The same use consumes 760 kWh in each location.

Model retraining, 1 year · result per service per year · estimated · one server location per line.Counted: the electricity of the GPUs during retraining, and the data centre overhead. Out of scope: the model's initial training by its publisher, data preparation and hardware manufacturingFind this case: projetcelsius.com/en/ai/database/organisations/retraining-a-model/Calculation by Projet Celsius, AI Impact Database, version 0.4 of 13 November 2026 · IA-0416, entry in version 0.1 · engine 0.3.1 · CC BY 4.0.

Usage parameters

Advanced settings duration

Duration

Result for the chosen case

Location of the servers

Quantity and duration

Agent hours per dayLimit: beyond it, the case goes outside what the entry describes (1 to 12).

24 agent hours a day, servers worldwide: 350 kgCO2e, 760 kWh.

For a carbon footprint

The row for a carbon footprint

Item	purchased services (GHG Protocol scope 3, category 1); regulatory BEGES, category 4; on rented GPUs; on the organisation's own servers, the electricity goes to scope 2 (BEGES, category 2)
Quantity	1 department over one year
Factor	350 kgCO2e per service over one year, IA-0416 version 0.1, based on the local grid (location-based), servers worldwide
Total	350 kgCO2e · not counted: the model's initial training by its publisher, data preparation and hardware manufacturing

Breakdown of emissions by item

Emissions associated with electricity (CO2e)

H100 hour320 kgCO2e · 91 %
PUE of the data centre32 kgCO2e · 9.1 %

Model retraining, 1 year · result per service per year · estimated · servers worldwide.Counted: the electricity of the GPUs during retraining, and the data centre overhead. Out of scope: the model's initial training by its publisher, data preparation and hardware manufacturingFind this case: projetcelsius.com/en/ai/database/organisations/retraining-a-model/Calculation by Projet Celsius, AI Impact Database, version 0.4 of 13 November 2026 · IA-0416, entry in version 0.1 · engine 0.3.1 · CC BY 4.0.

Breakdown by item

Quantity, unit value, contribution and nature of the data for each item

Item	Quantity	Unit value	Contribution	Share	Nature of the data
H100 hour	× 768	900 Wh	690 kWh	91 %	assumption · Dettmers et al.
PUE of the data centre	× 1.10 on the rest		69 kWh	9 %	convention

Entries that use it

No other entry in the database is calculated from this one.

Sources of the calculation

3 publications, for this entry and those used to calculate it, with the chosen parameters.

The result is calculated by Projet Celsius. The sources below support its uses, its factors and its conversion; they do not all measure this total.

Uses and assumptions

These publications shed light on AI tasks and uses. They do not measure the carbon footprint of the case calculated.

Dettmers et al. · May 2023
QLoRA: Efficient Finetuning of Quantized LLMs ↗ (new tab)
H100 hour
Role and limits : Dettmers et al.
Author. Dettmers, Pagnoni, Holtzman and Zettlemoyer
What the publication covers. memory and duration of a large-model fine-tuning on one GPU
Reading by the database. An order of magnitude published by researchers
- Model retraining → : H100 hour →
  Four retrainings per year, of 24 hours on 8 GPUs
  Scope of the figure used. GPU and its share of the server; excluding the data centre
  Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.

Energy and factors

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

Watershed · Aug. 2026
Estimating GHG Emissions from AI Use: Framework for Corporate-Level Measurement ↗ (new tab)
PUE of the data centre
Role and limits : Watershed
Author. Bistline, Ulissi, Davis, Glidden, Joyce, Li, Mohsenin and Suh (Watershed)
What the publication covers. factors per 1,000 tokens read and written at the server before PUE (host and standby machines included), PUE separate (1.10 for a large operator, 1.56 for an unknown operator)
What it does not cover. clean measurements, the model and the hardware
Reading by the database. Default factors proposed for review (public data as of June 2026); for a large operator, taken from Google's measurement published in 2025 (Elsworth et al.), which already contains the data centre overhead; constant per token, they underestimate long contexts and agents according to the authors
- Data centre PUE →
  Reference of the selected value
  Scope of the figure used. Default factor for a large operator, from the server to the data centre; 1.56 for an unknown operator
  Limit. A default value, not a measurement

Conversion into carbon

The source of the electricity factor of the selected location.

Worldwide: 458 gCO2e/kWh, life cycle.

Ember · 2026
Global Electricity Review 2026, and Our World in Data based on Ember ↗ (new tab)
Grid electricity
Role and limits : Ember
Author. Ember, Our World in Data
What the publication covers. electricity generation over the life cycle, by country
What it does not cover. guarantees of origin
Reading by the database. The same source for all countries
- Grid electricity →
  Conversion of electricity into carbon: worldwide
  Scope of the figure used. the production of electricity over the life cycle, including power plant construction and fuels
  Limit. An average for the chosen country, not the electricity contract or the data centre of a specific service.

Nature of the data

Published readings, conventions, assumptions

What this figure rests on

Projet Celsius assumptions 100 %

A share counts as an assumption as soon as a quantity set by Projet Celsius (a frequency, a share done with AI, a headcount) enters its calculation, even if the rest rests on published measurements.

Calculation assumptions

Quantities, nature and source of each assumption

H100 hour : × 768
Four retrainings per year, of 24 hours on 8 GPUs
assumption · Dettmers et al. · H100 hour

Methodology conventions

With their alternatives

The full methodology

The data centre	Server electricity is multiplied by 1.10 for the data centre · alternatives: × 1.72 from chips to the full service at Google, never to be combined
The default location	Without other information, the servers are in the world, at the world average electricity · alternatives: the United States, where most of the large assistants run
Electricity, location-based, on a life cycle basis	The kilowatt-hour is that of the servers' country, on a life cycle basis, from a single source for all countries · alternatives: market-based, with purchase contracts, shown alongside when published
Two significant figures	A calculated value is rounded to two significant figures, a published value is reproduced with the precision of its source; the calculation keeps full precision
The formula	the sum of what makes up the entry, then CO2e: electricity (Wh) multiplied by the carbon intensity of the server location (gCO2e/kWh), divided by 1,000
Calculate a case without an entry	The formula, the default values and a step-by-step example

History

Entry in version 0.1 of 2 November 2026, in the database in version 0.4

The changelog

2 November 2026Version 0.1 of the entry, as of the date of its current value.
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, "Model retraining, 1 year", IA-0416 (entry in version 0.1 of 2 November 2026), projetcelsius.com/en/ai/database/IA-0416/ · CC BY 4.0.

Download

JSON Markdown