Towards a lean and responsible AI: the state framework for purchasing and designing frugal AI

The energy efficiency of artificial intelligence is no longer a pious hope reserved for sustainability reports. Since 2022, the race for giant models has caused data centers to see an explosion in demand for electricity, water, and metals, to the point that the International Energy Agency is now referring to it as a grid issue. In France, the government responded in December 2025 with a practical guide for public procurement officers, which, for the first time, sets precise requirements for AI solution providers: question the necessity, size the model accordingly, measure and publish impacts, limit retraining, manage the data lifecycle, and assess social effects. This framework officially applies only to public procurement, but it currently represents the best available guide for any organization that wants to purchase or design energy-efficient and responsible AI. This article explains its rationale and suggests what a private company can learn from it.

Article published in December 2025 and completely rewritten in September 2026. The first version summarized, without citing it, the "Practical guide for the responsible purchase of artificial intelligence solutions" published on December 5, 2025 by DINUM, the State Procurement Department and the Ecolab of the CGDD; this version explicitly cites it, verifies the figures and adapts it to the context of companies.

Why sobriety is becoming an imperative

The figures used in this report are public and recent. According to the ADEME-Arcep update of November 2024, digital technology represented 4,4% of France's carbon footprint in 2022, and data centers accounted for 46% of this footprint, almost equal to that of devices. Regarding electricity, the International Energy Agency estimates in its "Energy and AI" report of April 2025 that data centers consumed 415 TWh in 2024, or 1,5% of global electricity, and that this consumption is expected to more than double by 2030, reaching approximately 945 TWh, slightly more than Japan's total current electricity consumption. AI is the primary driver of this growth.

In addition to carbon emissions, two less visible pressures come into play. First, water: cooling a large data center can consume the equivalent of 2,500 Olympic-sized swimming pools annually, according to the French Economic, Social and Environmental Council (CESE) estimate cited in the fact sheet. Second, metals: the Bank of France warns of supply risks and price volatility for copper, of which digital infrastructures are major consumers. In other words, the cost of AI is not measured solely in tons of CO₂; it is also measured in cubic meters and tons of metal, and these resources are localized, thus subject to conflicts of use.

Before starting any project: three questions most teams skip

The first contribution of this document is to impose a preliminary questioning, before any specifications are drawn up. First question: Is AI necessary? To meet the defined need? Many needs can be met with a business rule, traditional automation, or a simple reorganization, at a disproportionately low environmental and financial cost. Second question: Does the added value justify the impacts? The guidelines require that an environmental and social impact assessment be integrated into the decision itself, not into an appendix drafted afterward. Third question: what rebound effects What if these factors negate the gains? An assistant who writes ten times faster often produces ten times more documents; a more efficient template per query can be used a hundred times more. Individual gains tell us nothing about the overall impact until the volumes are considered.

This question seems obvious; it is rarely documented. In the organizations we work with, the question "Why AI here?" is generally answered by the enthusiasm of a sponsor or the availability of an already licensed tool. Putting it in writing, with the use case, its limitations, and the alternatives ruled out, is the most economical approach: it stops projects that should never have started.

Size the model as needed, not from the catalog

Once the use of AI is justified, the first lever for efficiency is the choice of model. The type and size must correspond to the level of accuracy actually required: there's no point in using a model with hundreds of billions of parameters to classify support tickets when a compact model, or even a specialized model with a few million parameters, already achieves the necessary quality. The same reasoning applies to the volume of training data, the number of training sessions, and hosting.

Streamlining techniques are now well-established and should be discussed with suppliers: pruning removes unnecessary connections from a network, quantization reduces the numerical accuracy of weights without significantly impacting business performance, and distillation transfers capabilities from a large model to a smaller one. Crucially, reusing pre-trained models, tuned on a smaller dataset, avoids paying the heaviest cost—the initial training cost—multiple times over. For the buyer, the right question is not "What is your best model?" but "What is the smallest model that meets the quality objective, and how do you prove it?"

Measure and publish: transparency as a condition of purchase

You can only manage what you measure. This guide is based on two French frameworks: the General Framework for Eco-design of Digital Services (RGESN), 7 of whose 79 criteria specifically target AI, and AFNOR Spec 2314 on frugal AI, which brings together 31 best practices for assessing and reducing the environmental impact of an AI system. The most advanced providers already publish an eco-design statement inspired by the RGESN, with a score out of 100 that makes their level of requirements comparable from one offering to another.

Three open-source and free tools enable measurement: Green Algorithms estimates the energy consumption and carbon footprint of an algorithm, distinguishing between training and inference; CodeCarbon measures consumption during execution; and Ecologits adopts a multi-criteria approach (energy, carbon, water, abiotic resources) for calls to large models. The key requirement is to drill down to the granularity of usage: the impact of an inference, an hour of service, or a complete user journey. It is at this level that business teams can make informed decisions, with the same rigor as for financial costs.

Retraining and complexity: two clauses that change everything

Training is the most expensive phase of an AI system, and periodic retraining multiplies the cost without the client always seeing any benefit. The guidelines require the owner to limit the number of retrainings to the users' strict needs: documenting the cadence, justifying each iteration, quantifying its consumption, and committing to performing only the necessary updates. A model that responds correctly does not need to be retrained simply because the schedule dictates it.

The complexity adaptation clause is the counterpart to this: the supplier must not increase the model's complexity without consultation, and must offer the option to remain on a previous version if it meets the needs. This clause directly addresses the now common practice among software vendors of automatically migrating all clients to the latest, larger, and more resource-intensive model, even though the previous version perfectly served their purpose. For a company, this translates into a simple requirement in the contract: the choice of model version belongs to the client, not the vendor.

Data: lifecycle, retention, location

The third section focuses on data, the storage and movement of which significantly contribute to the material and energy footprint. The guidelines require defining the complete data lifecycle, from collection to deletion: distinguishing between "hot" storage, which is constantly accessible, and "cold" storage, which consumes less energy; setting minimum retention periods; and deleting obsolete data. This discipline serves two purposes simultaneously, since data minimization is also a requirement of the GDPR.

The location of data centers must be made public. It determines the carbon intensity of the electricity consumed (a kilowatt-hour in France emits several times less than in Poland or Virginia), the pressure on local water resources, and the legal framework applicable to sensitive information. A provider that cannot disclose where its models are located cannot address environmental concerns or questions of sovereignty.

Responsibility does not stop at the environment.

AI incorporates a social dimension that this document addresses on an equal footing with the environmental aspect. It highlights the growing number of unsung heroes, from miners to annotators, moderators, and other data workers upon whom the models depend. It posits that organizations using AI remain responsible for the fairness of the results, which implies demanding explainability, bias assessment, and guarantees of individual rights. Finally, it calls for anticipating changes in working conditions and professions, and for respecting the protection of personal data in accordance with the GDPR.

For a company, this aspect aligns with the obligations of the AI Act regarding high-risk systems and social dialogue on the introduction of new technologies. It is easier to address upstream, in the specifications and a usage charter, than after the first incident of discrimination or the first conflict with employee representatives.

Translating requirements into clauses: from sourcing to consultation rules

It all starts with solid sourcing. Identifying developers, integrators, and hosting providers, and asking them consistent questions about their eco-design practices, their approach to inclusion, and their due diligence plans, helps avoid blind spots and calibrate realistic requirements—neither superficial nor impossible to meet. The document then organizes the contractual translation into three documents.

Le CCAP It carries the general commitments: to train the contractor's teams in responsible digital practices, including AI; to measure and communicate energy consumption, CO₂ emissions, and water usage; to precisely define the added value of AI compared to the previous solution; and to comply with applicable environmental standards. CCTP sets the technical rules: declare impact indicators (CO₂eq emissions, water consumption, total energy in kWh), limit retraining, preserve the possibility of remaining on a previous version of the model, define the data lifecycle. consultation rulesFinally, it gives weight to all this: the State Procurement Directorate recommends an environmental criterion weighted at at least 10% of the total score, and a social criterion also at 10%, so that these requirements really weigh on the choice of the offer instead of remaining a wish.

Hosting providers called upon to contribute

Infrastructure is the final link in the chain, and this document doesn't overlook it. Hosting providers must publish their efficiency indicators: PUE (Power Usage Effectiveness), WUE (Water Use Effectiveness), and CUE (Carbon Impact). They must offer the option of scheduling training sessions during periods of low carbon intensity on the electricity grid, provide tools that facilitate the design of more responsible AI, and allow users to leverage pre-trained models rather than recalculating everything. This enhanced transparency prevents impact shifting, where the problem is simply moved from a provider to a subcontractor, and encourages the use of underutilized capacity and off-peak hours for the most energy-intensive calculations.

What a private company can learn from this

A company is under no obligation to apply a framework designed for public procurement, but it's not prevented from doing so either, and the exercise is inexpensive. Five simple steps are all it takes to transpose the essentials. For each project, document the answers to the three preliminary questions: necessity, added value in relation to the impacts, and rebound effects. Ask each supplier for the smallest model that meets the quality objective, and provide proof. Require impact indicators by use case, control over retraining, and the right to remain on a previous version in the contract. Establish a data retention policy and know the location of data processing. And include these rules in the AI ​​usage charter so that they also apply to tools that teams adopt without going through a tendering process.

The result isn't just environmental. AI scaled to meet specific needs is less expensive to run, less vulnerable to price increases from software vendors, more resistant to regulatory audits, and inspires greater trust in both users and customers. Tomorrow's AI won't just be more powerful; above all, it will need to be more relevant, more transparent, and more human-centered. Organizations that set these rules now will be ready; others will discover them in a set of specifications they didn't create.

Sources. DINUM, the State Procurement Department and Ecolab of the CGDD, Practical guide for the responsible purchase of artificial intelligence solutionsDecember 5, 2025, for the preliminary questioning, clauses and reference documents cited; ADEME and Arcep, updating the environmental footprint of digital technology in France (2022 data), November 2024; International Energy Agency, Energy and AI, April 2025.

To learn more: our analysis on the environmental transparency of generative AI and our article on Energy as a competitive advantage in the AI ​​race.