Free resource · Word + Excel kit

Frugal AI Kit: The decisions that truly impact your AI project's footprint

What a request, a reasoning method, an agent, or a video actually weighs; what AFNOR, ADEME, and Arcep say; and the ten scoping, purchasing, and management decisions that account for the bulk of the gain, without writing a single line of code. Sourced orders of magnitude, not sensational figures.

📄 Word, 40 pages📊 Excel Workbook, 6 tools🌱 Based on AFNOR Spec 2314🆓 Free

Download the kitSee the orders of magnitude

There are two narratives about AI's footprint, and both are based on real figures. The first says that a query consumes ten times the energy of a Google search and drinks a bottle of water. The second says that a query uses less energy than a second of oven time and that's the end of it. They're simply describing different things: a median query in 2026 versus a query in 2023, on-site cooling water versus total water consumption, a market-based carbon factor after purchasing certificates versus the actual grid mix. Between these two narratives, the sponsor of an AI project can't find the orders of magnitude they need, and, more importantly, what they can actually decide.

This toolkit is based on the French standard definition, AFNOR Spec 2314: the frugality of an AI service depends "on a redefinition of uses or performance requirements." Redefining uses and requirements is a sponsor decision before it's a code optimization. And the data confirms this: the significant differences in footprint (a factor of 59 between a small and a large model, a factor of 13 to 700 for reasoning mode, a factor of 8 to 10 depending on the hosting country, and a factor of 2 when dividing the length of responses) are determined by the scope, the choice of solution, and the purchase. The toolkit brings together these levers, the official standards on which they are based, and the tools to implement them.

1. The true weight of AI, in orders of magnitude

The values ​​below come from the providers themselves (Google, OpenAI, Mistral), research laboratories (Hugging Face, Microsoft Research, universities), and agencies (IEA, ADEME, Arcep), all published between 2024 and 2026. They are expressed in watt-hours per request: 1,000 Wh equals 1 kWh, roughly enough to fully charge one hundred smartphones. For each value, the toolkit provides the source, date, and confidence level, and explains why the CO2 and water emissions for the same request vary by a factor of 30 to 170 depending on the scope considered.

Subject	Order of magnitude	Sources
A classification or extraction using a specialized model	0,002 to 0,05 Wh	Luccioni, Jernite, Strubell (ACM FAccT 2024)
A standard text request to an assistant	0,2 to 0,4 Wh	Google 0,24 Wh (August 2025), OpenAI 0,34 Wh, Epoch AI 0,3 Wh, Microsoft Research 0,31 Wh, IEA 0,10 to 0,40 Wh (April 2026)
A query in reasoning mode	3 to 40 Wh	Microsoft Research (×13), Jegham et al. (o3 12 Wh, GPT-5 up to 33,8 Wh), AI Energy Score v2 (×30 on average, ×150 to ×700 for the same model)
An agent's task or in-depth research	15 to 60 Wh	IEA, “Key Questions on Energy and AI” (April 2026); KAIST (up to 348 Wh in the laboratory)
A generated image	0,5 to 3 Wh	MIT Technology Review and ML.ENERGY (May 2025), Luccioni 2024, IEA 2026
A five-second generated video	100 to 950 Wh	Hugging Face (July 2025), MIT Technology Review, IEA 2026
Training a large model	20,400 tCO2e for Mistral Large 2, including 18 months of operation	Mistral AI, Carbone 4, ADEME (July 2025)
The hosting country	46 gCO2e/kWh in France, 210 in the European Union, 384 in the United States, 458 worldwide	ADEME Base Empreinte 2025, Ember 2026
Data centers around the world	485 TWh in 2025, 950 TWh expected in 2030	IEA, April 2026

Three conclusions emerge, and they structure the toolkit. Inference, meaning usage, now dominates training: your usage decisions carry more weight than the technical choices of a laboratory. Large discrepancies are decisions, not optimizations: none of the levers below require a single line of code. And figures must be interpreted within their scope: a frugal sponsor doesn't calculate down to the gram; they know which questions to ask and reject figures without context.

2. What the framework says, and what actually obliges you

In two years, France has produced a unique body of work on frugal AI: AFNOR Spec 2314 and its thirty-one best practices (June 2024), the Engagement Kit with fifteen best practices from Ecolab, Hub France IA, and AFNOR (June 2025), the practical guide for responsible procurement of AI solutions from DINUM and the State Procurement Directorate (November 2025), the ADEME opinion (July 22, 2026), the Arcep report (May 21, 2026), and the Senate report (September 16, 2026). The kit extracts what a sponsor can use directly: the three cumulative conditions for a frugal service, the six-step process, sourcing questions and standard clauses, and the seven algorithmic criteria of the RGESN.

It must also be said without mincing words: No legislation currently imposes a direct environmental obligation on a company that deploys an AI assistant or agent.The obligations fall on model providers (energy documentation for general-purpose models, Annex XI of the AI Act), data centers (European reporting, sustainability label A to G proposed by the Commission on September 21, 2026, for 2027), cloud providers (data collection by Arcep with AI segmentation, publication in 2027), and public purchasers. They indirectly concern you: because they produce data that you can demand, because the purchased AI falls within your Scope 3 obligations, and because your environmental claims have been regulated since September 27, 2026. The sponsor's leverage is therefore contractual and organizational.

Date	What happens	What you can do with it
August 2, 2026	Commission enforcement powers on energy documentation of general purpose models (AI Act, Annex XI).	Request documented energy consumption data for the model in your tenders.
September 27, 2026	Application of Directive 2024/825 on environmental claims.	Review all "responsible AI" or "green AI" communications with chapter 4 of AFNOR Spec 2314.
March 31, 2027	Second Arcep collection of cloud providers; publication of 2025 data in the 2027 "Pour un numérique soutenable" survey.	Compare your hosting providers using public data.
August 2, 2027	Compliance of general purpose models prior to August 2025.	All your suppliers then have energy documentation: no more excuses.
August 15, 2027	First sustainability labels A to G for data centers, if the delegated act is adopted as proposed.	Require the data center rating in your hosting contracts.
2028	Commission report on the standardisation of AI energy efficiency (article 40); parliamentary follow-up (eco-score of models).	Prepare indicators compatible with a future eco-score: gCO2e per unit of work, location-based.

3. The ten levers of the sponsor

Each lever is documented by at least one source published between 2024 and 2026 and can be decided without writing a single line of code. The toolkit provides, for each lever, the decision-making process, what the data reveals, the question to ask your team or supplier, and the pitfall to avoid. The first four levers account for the majority of the gains.

Lever	Documented Gain	Source
1. The smallest model capable of the task, by default; the larger model upon justification.	Factor 59 between Llama 3.1 8B and 405B; up to 90% savings with small, specialized models	MIT TR and ML.ENERGY 2025; UNESCO and UCL 2025; Arcep 2026
2. Reasoning, advanced search, and agents are disabled by default.	Factor 13 to 700 for reasoning; 15 to 60 Wh per agent task	Microsoft Research 2025; AI Energy Score v2; IEA 2026; ADEME 2026
3. Classify rather than generate, text rather than image, image rather than video	Factor of up to 1,450 between classification and image generation	Luccioni 2024; IEA 2026
4. Short answers and a controlled context	Halving the length of responses: 54% less energy	UNESCO and UCL 2025
5. Low-carbon hosting, proven on a location-based basis	46 gCO2e/kWh in France compared to 458 worldwide: a factor of 8 to 10	ADEME, Ember 2026, Arcep 2026
6. Reuse rather than retrain: prompt, framed RAG, light fine-tuning	RAG: 20 to 51% less energy with an adjusted perimeter	Guo, Gao, Bogner 2026; AFNOR Spec 2314; RGESN
7. Caching, request grouping, and routing	"One complex prompt is better than five simple prompts"	Razorfish and Green IT 2024; Mistral 2025
8. Require compressed models and an efficiency rating	Quantization: 22 to 44% less energy; up to 73% with a set of optimizations	UNESCO and UCL 2025; Fernandez, Luccioni et al. 2025; Hugging Face
9. Postpone non-urgent treatments to off-peak hours.	From 9 to 58 gCO2e/kWh per year in France: factor 6 (secondary lever)	RTE 2025; MiNumEco fact sheet
10. Manage, raise awareness, stop: five indicators, two thresholds, one lead, one charter	Only 12% of organizations measure the footprint of their generative AI.	Capgemini 2025; Ecolab 2025 Commitment Kit

4. The case of off-the-shelf assistants

Most large organizations don't build AI systems from scratch; they purchase licenses for Copilot, ChatGPT Enterprise, Gemini, Claude, or Vibe (formerly Le Chat) from Mistral. The prevailing view is that the environmental footprint is the vendor's responsibility. Their contribution is real; yours remains, and the toolkit details it across seven key areas: default settings (model and modes), disabling unnecessary "augmented" features (a recommendation made by ADEME since July 2026), the number of licenses and their usage rate, the inference location and carbon footprint of your tenant, usage reports as measurement proxies, user awareness training, and the transparency clause in the framework agreement. It lists what a sponsor can expect from a vendor in 2026: three of these items correspond to what the vendor already provides to regulators.

5. Buying: questions, indicators, clauses

Since no text mandates it, your contract will bind your supplier. The toolkit adapts the sourcing questions and standard clauses from the government's practical guide to private contracts: separate questions for the model publisher, the integrator, and the hosting provider; nine indicators to require, each with the correct answer and a red flag (country and inference data center, PUE and WUE, footprint per unit of work and method, location-based agreement, training energy, efficiency score, model size, bounding options, usage report); six pre-written clauses (transparency, reporting, proportionality and non-complexity, retraining and versions, data lifecycle, skills); and the weighting rule, a minimum of 10% of the overall score, recommended by the French Digital Affairs Directorate (DINUM) and the State Procurement Department. The workbook's grid scores fifteen criteria in five categories and ranks your candidates out of 100.

6. Management: five indicators, two thresholds, one lead

The minimum sustainable metrics include: the number of requests, estimated electricity consumption, estimated location-based emissions, the proportion of requests in high-cost mode, and the monthly cost, with the license utilization rate serving as a contextual indicator. Two thresholds, taken from the governance rule published by SNCF (the French National Railway Company) with its Impact'IA calculator, are also relevant: beyond one tonne of CO2e per year, a documented analysis is required; beyond ten tonnes, a mandatory reduction plan is necessary. A designated frugal AI lead, as requested in the Ecolab Commitment Kit, maintains the use case register and prepares the quarterly review for the steering committee. The kit also addresses what all guides overlook: the criteria for stopping, retraining, and decommissioning, and how to discuss these publicly without violating the Environmental Claims Directive.

7. What the kit contains

The Word document has 40 pages, in eight parts and six annexes: the user manual and the three conditions of frugality; what AI really weighs (orders of magnitude, why the figures diverge, five misconceptions, three conclusions); what the framework says (references, real obligations, 2026-2028 timetable); the self-diagnosis in twelve questions; the framing (necessity test, "AI or no AI" decision tree, sufficient performance level, task family, usage volume and AFNOR formula, rebound effects, impact sheet); the ten levers and the case of off-the-shelf assistants; the purchase (sourcing questions, indicators, six clauses, weighting); the management (indicators, thresholds, reference person, charter, decommissioning, communication); the 90-day and 12-month roadmap. The appendices provide the one-page impact sheet, the nine-rule responsible use charter, the note to the management committee, the thirty-point checklist, the fifteen-term glossary, and dated sources.

The Excel workbook includes the tools in six tabs, plus a reference tab where each quantified hypothesis has its source and date: the self-assessment that calculates your profile and the phase to prioritize; the calculator that transforms your volumes by task type, your hosting country, and your PUE into an annual range of kWh, kgCO2e, and liters, with the level relative to thresholds and the savings from hosting in France; the decision matrix for each option, with the impact index of each; the supplier evaluation grid with fifteen weighted criteria; the twelve-month dashboard; and the use case register with estimated footprint, level, and retraining and shutdown criteria. For more detailed calculations including training and manufacturing, the kit refers to the SNCF's Impact'IA calculator.

The kit, part by part: what each part contains and what it's for. The templates to fill out are in the Word document and the Excel workbook.

Part 1. The true weight of AI

Why the figures don't agree with each other

Each source chooses its scope. Google includes on-site cooling water (0,26 mL) but not electricity generation water, and converts energy using a market-based factor of 94 gCO2e/kWh, compared to 345 gCO2e/kWh in 2024 based on location data. Meta reports 0 tCO2e for Llama 3.1 training based on market data, and 11,390 tCO2e based on location data. Hence the rule for reading the toolkit: before any figure, ask about the scope, the carbon convention, the water usage, and the query.

Five misconceptions, and what the data says.

“The problem is training”: according to the IEA, energy consumption “has already decisively shifted from training to inference.” “Each query is negligible”: Google reduced the median energy of a Gemini query by a factor of 33 in twelve months, but its emissions increased by 18% in 2025, while Microsoft’s increased by 25%: this is the Jevons paradox. “Our supplier is carbon neutral”: the manufacturing of the hardware, 130 to 270 kgCO2e per GPU, is excluded from almost all per-query figures.

"Every conversation swallows a bottle of water."

The "500 mL for 10 to 50 responses" (Ren et al., 2023) concerned GPT-3 and included water used for power generation; Google reports 0,26 mL when only on-site cooling is considered; Mistral finds 45 mL per response in a full analysis. The 170-fold discrepancy stems from the scope. Most of the water is indirect and location-dependent: require WUE and location data.

Part 2. What the framework says: standards, obligations, timetable

AFNOR Spec 2314, "General reference framework for frugal AI" (June 2024)

Free of charge and developed at the initiative of Ecolab, it establishes an evaluation method based on life cycle analysis and thirty-one best practice sheets categorized by benefits and effort. Its six-step process serves as a project plan: define the need, identify all solutions, with or without AI, evaluate each one throughout its life cycle, integrate the evaluation into decision-making, implement tools for monitoring actual usage, and adapt roles.

RGESN, General Eco-design Framework for Digital Services, 2024 version

Of its 78 criteria, seven relate to algorithms and are presented as such to a provider: need for training (9.1), complexity proportionate to use (9.2), limited training and data (9.3 and 9.4), optimized retraining (9.5), compression (9.6), inference strategy (9.7). Criterion 1.1 requires that the service has "been favorably evaluated in terms of utility, taking into account its environmental impacts."

The standards and metrics that are coming

ITU-T Recommendation L.1801, approved on February 26, 2026, is the first comprehensive international methodology that compares a service with and without AI. SCI for AI (Green Software Foundation, December 17, 2025) measures carbon intensity per functional unit: per token, per image, per page. The AI ​​Energy Score (version 2, December 4, 2025) rates models from one to five stars across ten tasks: the simplest free comparison tool for a buyer.

Part 3. Self-assessment in twelve questions

Lecture du résultat

Twelve questions, each scored 0, 1, or 2, for a total of 24 points: framing (out of 6), solution selection (out of 6), purchase (out of 4), and management (out of 8). These questions assess whether the footprint has become a factor in your decisions. Scores from 0 to 8 indicate a "Discovery" profile: start with the default model and expensive options. Scores from 9 to 16 indicate a "On the Way" approach: focus on the lowest-scored phase. Scores from 17 to 24 indicate a "Frugal" approach: document your decisions so you can discuss them publicly without risk.

Part 4. Framing: Decide before you build

The necessity test

Five questions, in a meeting with the business and the IT department: what problem, for whom, and how will we know it's solved? Does a solution without AI solve 80% of it? What is the simplest form of AI that is sufficient? What level of performance is sufficient? What usage volume and what rebound effects? The acceptance threshold is defined before testing; we first test the smallest plausible model on fifty real-world cases, and we only scale up based on measured failures.

The "AI or no AI" decision tree

Eight questions, each with a default orientation and an impact index (standard text query = 1): AI-free solution if an existing rule or tool is sufficient 80% of the time (index close to 0); small specialized model for classifying or extracting (0,03 to 0,3); bounded response or framed RAG for writing short (1 to 3); cutting-edge model and justification-based reasoning (10 to 130); agent for rare tasks (50 to 200); image if text is not sufficient (2 to 10); exceptional video (300 to 3,000).

Estimating the usage volume: the formula a·x + b + c·y

AFNOR bases its calculations on a year of service: a, the cost of an inference; x, the number of queries; b, the amortized training; c·y, the retraining. For a purchased model, the essential element is a·x. Example from the kit: 500 users with 10 queries per working day, i.e., 1,1 million per year, consume approximately 400 kWh and 180 kgCO2e in standard text; 13,000 kWh and 6 tonnes in continuous reasoning; more than 15 tonnes with ten agent tasks per day per person (at a global emission factor of 458 gCO2e/kWh).

The one-page impact sheet

Intended for the committee that decides on the launch, it brings together the need, the AI-free alternative studied, the solution chosen, the estimated volume, the annual order of magnitude in range, the level in relation to the thresholds, the levers activated, the supplier's responses, the indicators and the stopping criterion, as well as three probable rebound effects (ten versions of a text, one summary per meeting) with their control measure.

Part 5. Choosing the solution: the ten levers of the sponsor

Lever 5: low-carbon, location-based hosting

The inference process runs in France or another European country with low-carbon electricity, with supporting evidence available on a location-based basis; market-based processing is not accepted. In SNCF's calculation tool, the same project's carbon footprint increases from 41 to 180 kgCO2e when the inference is moved from France to the United States. The pitfall: confusing data location with computing location. Ask for the inference location, not the storage location.

Lever 6: Reuse rather than retrain

In order: a well-constructed prompt, a well-defined RAG, and light fine-tuning; retraining remains the exception, triggered by measured degradation rather than a schedule. Adjusting the similarity threshold of a RAG reduces energy by 20 to 51% without loss of accuracy (Guo, Gao, and Bogner, 2026). The pitfall: the "all-SharePoint" RAG, which is more expensive for each question and less relevant.

What a sponsor can expect from an assistant publisher in 2026

The inference location of its tenant and the associated location-based factor; the PUE and WUE of the data centers; the median and reasoning mode energy per request, including the perimeter; the documented energy consumption of the model (Annex XI of the AI Act); an efficiency rating; the group deactivation of expensive modes; and a monthly usage report. Three of these pieces of information are already produced for regulators.

Part 6. Purchase and contract: demanding, weighing, writing

Sourcing issues, by supplier type

To be submitted in writing, with answers attached to the contract. To the software publisher: What is the smallest model in your catalog capable of handling our use cases? To the systems integrator: What AI-free or lighter alternative have you considered, and how do you limit token usage? To the hosting provider: What are the PUE, WUE, and CUE values, and a footprint report per tenant on a location-based basis, including server manufacturing costs?

The indicators to require, and the correct answer

For each indicator, the toolkit provides the correct answer and a red flag. Location of inference: a named country and site, not "our data is in Europe." PUE: less than 1,3, with WUE. Footprint per unit of work: a median value and a reasoning-mode value, with scope, method, and date. Carbon convention: market-based alone is a red flag. For training energy, the answer "confidential" after August 2, 2027, is another.

Six clauses for a private contract

Environmental transparency, quarterly reporting, proportionality and simplicity, retraining and versions (keeping a previous version for twelve months), data lifecycle, skills. The rubric scores fifteen criteria in five categories (transparency 30%, efficiency 25%, hosting 20%, contract 15%, governance 10%), and the environment accounts for at least 10% of the overall score, as recommended by the MiNumEco guide.

Part 7. Steering and governance: five indicators, two thresholds, one lead

The five indicators

Each case has its own trigger: a 30% increase in requests in a single month warrants an explanation; estimated emissions are compared to the thresholds of one and ten tons per year; when more than 10% of requests are processed in high-cost mode, settings are reviewed; and when license utilization falls below 40%, licenses are reassigned, training is provided, or the system is shut down. The kgCO2e per case processed is the maturity indicator.

The charter for sobriety in use

Nine rules, each justified by an order of magnitude: ask yourself if AI is necessary, ask a complete question the first time, ask for the short version, reserve reasoning and agents for tasks that deserve it, no image when text is enough, do not regenerate "just to see", keep the default model, reuse validated prompts, report unnecessary uses to the referent.

Stop, retrain, decommission

Three rules were written from the outset. Discontinuation: utility rate below 20% for two quarters, business objective not achieved, or redundancy with a standard function. Retraining: only upon measured degradation below the acceptance threshold, never by schedule. Decommissioning: data deleted or archived, licenses released, indexes destroyed, and a line item recorded in the registry.

Part 8. 90-day and 12-month roadmap

The first 90 days

Days 1 to 30, gain clarity: appoint the lead user, inventory use cases and licenses, complete the self-assessment, and adopt the charter. Days 31 to 60, quantify and resolve: run the five largest use cases through the calculator, draft the impact assessment for the largest use case, configure default settings, and consult with the three main suppliers. Days 61 to 90, manage: dashboard, initial committee review, and transparency and reporting clauses for the first renewal.

8. Frequently Asked Questions

Does a ChatGPT query really consume ten times the resources of a Google search?

This figure is from 2023, based on an assumption of 2,000 tokens per response. Measurements and estimates for 2025-2026 converge on 0,2 to 0,4 Wh for a median text query (Google, OpenAI, Epoch AI, Microsoft Research, IEA). But the median masks the tail of the distribution: a reasoning mode, a long context, or an agent costs ten to a thousand times more. The real question is no longer "how much per query" but "which model, which mode, how many tokens?"

Are we obligated to measure the footprint of our AI?

No, not directly. The AI Act mandates energy documentation for general-purpose model providers, not for the companies that use them; the REEN law and Arcep rulings target local authorities, data centers, and cloud providers. Indirectly, purchased AI falls under Scope 3 of your reporting, and your environmental claims are subject to regulations. The toolkit explains what you are and aren't obligated to do, and what you can require from your suppliers based on their own obligations.

Our tools are Copilot or ChatGPT Enterprise: what can we do?

Many things to consider: choosing the default model and modes, disabling unnecessary "enhanced" features as recommended by ADEME, measuring license usage, obtaining the locus of inference and carbon footprint of your tenant, using usage reports as a metric, training users, and introducing a transparency clause for renewals. The toolkit dedicates a section to this and lists what a sponsor can expect from a software vendor in 2026.

Does the kit calculate our exact footprint?

No, and no one can do it today: 84% of models have no environmental information available (Arcep, May 2026). The calculator provides a low-to-high range based on published work units, your volumes, your hosting country, and your PUE, accurate enough to classify your use cases against thresholds and make informed decisions. For a life cycle assessment, it refers users to the AFNOR method and to a consulting firm.

Why are there thresholds of one and ten tonnes?

Because they are simple, public, and proven: these are the ones that SNCF applies to its digital projects and which it made public with its Impact'IA calculator in 2026. One ton of CO2e is roughly equivalent to a round trip from Paris to New York per passenger; ten tons, to slightly more than the annual carbon footprint of a French resident. You can adapt them to your own needs, but keep two: one that requires reflection, one that requires action.

What to start with ?

Through the two most cost-effective and quickest decisions: the lightest default model capable of handling the task, and the expensive modes (reasoning, agents, image, video) disabled by default. Then, through the twelve-question self-assessment, which tells you which phase to work on next. The 90-day roadmap in the kit is written for an organization starting from scratch.

9. Going further

This kit complements the four other kits from the sponsor: the AI Act kit covers what applies to a deployer and the AI ​​literacy program, the GDPR kit covers the qualification of projects and the file that the DPO expects, the AI and Social Dialogue Kit prepares the consultation of the CSE and the Security and data kit addresses the project's security risks. The AI Systems Governance Kit provides the register and qualification grid, the AI charter template sets the rules of use, within which the responsible use charter of this kit fits. Organizations that want to quickly pass the necessity test on their use cases and configure their assistants to be more energy-efficient do so within the framework of the AI training in companies and the mission Governing your agents.

Read also : AI Project Sponsor Security and Data Kit, Sponsor's AI Act Kit: The 2026-2027 Roadmap, GDPR Kit for AI Project Sponsors, AI systems governance: the 2026 toolkit et AI Charter in Business: The 2026 Model (Downloadable).

The entire series of free kits. To drive adoption: AI charter template, AI Ambassadors Kit, Manager's toolkit for dealing with AI, Adoption Roadmap Kit, AI maturity grid, AI Measurement and ROI Kit et AI Governance KitFor the AI ​​project sponsor: AI Act kit, GDPR kit, Security and data kit, AI and Social Dialogue Kit, AI Sovereignty Kit et responsible AI kitBy profession: AI HR kit, AI marketing kit, AI finance kit, AI kit for lawyers, AI kit for salespeople, AI toolkit for project managers, AI innovator kit et communicator's AI kit.

10. Sources

The figures and rules cited on this page and in the kit are based on the following dated sources. All are free. The kit lists approximately forty more in the appendix.

• AFNOR, AFNOR Spec 2314, General reference for frugal AI, June 2024.
• Ecolab (CGDD), Hub France IA, AFNOR, Engagement Kit for Frugal AI: 15 Best Practices, June 2025.
• MiNumEco, Ecolab, State Procurement Department, Practical guide for the responsible purchase of AI solutions, November 2025.
• ADEME, Opinion piece: "Generative artificial intelligence: significant environmental impacts", July 22, 2026.
• Arcep and PEReN, Generative artificial intelligence: what are the environmental challenges?, May 21, 2026
• Senate, The environmental footprint of AI: a challenge for the planet, an opportunity for France and Europe, report no. 949, September 16, 2026.
• Mistral AI, Carbone 4, ADEME, Life cycle analysis of Mistral Large 2July 22, 2025; Google, Measuring the environmental impact of AI inference, August 21, 2025.
• International Energy Agency, Key Questions on Energy and AI, April 2026; UNESCO and UCL, “Smarter, Smaller, Stronger: Resource-Efficient Generative AI”, 2025.
• S. Luccioni, Y. Jernite, E. Strubell, Power Hungry Processing: Watts Driving the Cost of AI Deployment?, ACM FAccT 2024; Hugging Face et al., AI Energy Score, version 2December 4, 2025; Microsoft Research, Energy Use of AI Inference, Efficiency Pathways, and Test-Time Scaling, 2025.
• SNCF, Responsible Digital Department and CSR Department, Impact'IA calculatorJuly 2026, CC BY-NC-SA license; GenAI Impact, EcoLogits Ministry of Culture and DINUM, Compar:IA.