UsesLevel 2 of 6 machine operationMachine operation, level 2 of 6Used to calculate 2 entries, levels 2 and 3Database structure 
6Territories
5Organisations
4Occupations and activities
3Uses
1Calculation factorsReason, multiplier by effort
A model that reasons first writes a draft, then its answer.
- This setting multiplies the electricity of the query by 1.92.
- The factor covers the GPU electricity for the same queries, with and without reasoning.
- Out of scope: the rest of the server and data centre cooling.
Sentence to citeReasoning, multiplier by effort: × 1.92.
- Used to calculate
Origin of the value
It produces more tokens. The database distinguishes 3 levels of effort: none (× 1), medium (× 1.92, the average effect that PEReN derives from its measurements on 22 open models, with a 95% interval from - 21% to + 348%) and high (× 10, based on measurements and estimates published since 2025).
Characteristics
| Unit | factor |
|---|---|
| Counted | the GPU electricity for the same queries, with and without reasoning |
| Not counted | the rest of the server and data centre cooling |
| Status | measured, 4 readings with the same scope, 4 other published values |
All the characteristics
| Identifier | IA-0015, projetcelsius.com/en/ai/database/IA-0015/ |
|---|---|
| Version | 0.3 of 2 November 2026 |
| Category | Machine operations, what the server computes |
| Reasoning | Without reasoning (× 1.00), With reasoning, medium effort (× 1.92), With reasoning, high effort (× 10) |
| Choice of the selected value | value taken as it stands from its source, without calculation |
| Sources | PEReN, Hugging Face, IEA, Epoch AI, 2025 |
| Licence | CC BY 4.0 |
Variants
This block is made of no other: its value is read from its sources.
Entries that use it (2)
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.
Role and limits : PEReN
Author. PEReN, published by Arcep
What the publication covers. GPU energy at inference, 22 open models, 6 question sets
What it does not cover. cooling, manufacturing, training
Reading by the database. Open models
Reference for the variant "With reasoning, medium effort"
Scope of the figure used. + 92% on average with reasoning, effect estimated by regression on 22 open models and 6 question sets; 95% interval from - 21% to + 348%
Limit. Open models, varied tasks
Published values
8 figures for the same quantity, each in its own scope.
× 1.92 selected, between × 1.92 and × 30 for published values of the same scope
| Source | Date | Multiplier | What the figure measures | Scope | Status |
|---|---|---|---|---|---|
| Reasoningselected by the database, v0.3 | 2026 | × 1.92 | the value selected by the database | same scope | measured |
| PEReN, 22 open models ↗Étude IA et environnement, rapport technique | 2025 | × 1.92 | + 92% on average with reasoning, effect estimated by regression on 22 open models and 6 question sets; 95% interval from - 21% to + 348% | same scope | measured |
| IEA, Energy and AI ↗Energy and AI | Apr. 2025 | × 2.00 | DeepSeek-R1 compared with a model of comparable size without reasoning, simple text task, GPU only | same scope | estimated |
| Epoch AI ↗How much energy does ChatGPT use? | Feb. 2025 | × 2.57 | o1 versus GPT-4o, on a small sample of conversations: 1,390 tokens written versus 540 | same scope | computation |
| Jegham et al. ↗How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference | 2025 | × 7.36 | GPT-5, average question (1,000 tokens read, 1,000 written): 17.15 Wh with high reasoning versus 2.33 Wh with minimal reasoning | same scope | computation |
| Hugging Face, AI Energy Score v2 ↗AI Energy Score v2 | Dec. 2025 | × 30 | 30 times more on average between two groups of models | same scope | measured |
| Microsoft Research ↗Estimation de l'énergie des requêtes en production (Joule) | Apr. 2026 | × 10 | Long reasoning and agent queries, compared with the median query, in production | other scope | estimated |
| ML.ENERGY, benchmark v3 ↗The ML.ENERGY Leaderboard v3, données par tâche | Feb. 2026 | × 10.2 | Qwen 3 235B: 8,955 tokens written on average by its reasoning version (GPQA science questions), versus 874 for its version that answers directly (conversations) | other scope | computation |
| Hugging Face, AI Energy Score v2 ↗AI Energy Score v2 | Dec. 2025 | × 150 to × 700 | Three models tested with then without reasoning | other scope | measured |
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
| 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
History
Entry in version 0.3 of 2 November 2026, in the database in version 0.4
- 13 November 2026Description and reading : 23 open models becomes 22 open models.Table 3 of the PEReN report (p. 21) lists 22 distinct models, as does Arcep's report of May 2026; the conclusion counts 23 because Mistral-Small-3.2-24B appears twice. The factor (× 1.92) does not change.
- 11 November 2026Description : medium (× 1.92, the average measured by PEReN on 22 open models) becomes medium (× 1.92, the average effect estimated by PEReN on the 23 open models tested, with a 95% interval from - 21% to + 348%).PEReN states that it tested 23 models; the + 92% is the average effect of a regression, with its 95% interval (table 4); the "up to" wording is not in the source. Address of the report added to the source. The factor does not change.
- 2 November 2026Designation : Reasoning, as a multiple of one question becomes Reasoning, multiplier of one query.More precise designation: the object first, the specification after the comma, the quantity or period at the end; the former title remains searchable.
- 2 November 2026Designation : Reasoning, multiplier of one query becomes Reasoning, multiplier by effort.In people's own words; the effort becomes a setting.
- 2 November 2026Address : raisonnement-multiplicateur becomes raisonner-multiplicateur.Address in people's own words; the old one redirected.
- 2 November 2026Selected value (1.92).The value 1.92 moves to the medium effort row, the default row of the block, unchanged.
- 31 October 2026Creation of the database, in its first version (0.1).
