On 19 September, as part of Nantes Digital Week, Anaëlle Monti and Boris Bailly represented 🌵 Aguaro at the round table discussion ‘Le grand déballage de l'IA’ (The Great AI Reveal), organised by Marie Bernard from Nantes University and Thibaud Menanteau, Green IT Manager at Nantes Métropole, alongside Juliette Plaire, IA Regional Coordination Officer at Nantes Métropole, and Walter Bonomo, Head of Educational Innovation at INRAE (National Research Institute for Agriculture, Food and the Environment) and coordinator of the INRAE Training Lab.

The event, which brought together more than forty participants, offered a fresh perspective on Artificial Intelligence — not as an abstract concept, but as a material, social, and political system that can be measured and framed.

AI is not immaterial

Firstly, it is essential to remember that behind every query and every prompt there are data centres, specialised servers, energy-intensive GPUs*, communication networks, and global supply chains that consume water, electricity, and mineral resources. Training a large model can mobilise thousands of processors for weeks; above all, the cost doesn’t end there: each individual query also consumes energy. It is a question of balancing the value produced and the resources mobilised.

As part of the life cycle assessment (LCA) of its Mistral Large 2 generative AI model, the French start-up Mistral AI reported that model training and inference accounted for 85.5% of its greenhouse gas (GHG) emissions and 91% of its water consumption. These phases also represented 29% of mineral resource use, while the majority of resources (61%) were consumed during the manufacturing, transport, and end-of-life stages of the servers. (Source: Mistral AI)

Beyond energy, the footprint of AI is also social and geopolitical: material extraction in fragile contexts, the concentration of power in the hands of a few players, and outsourcing of AI model training tasks to click workers. Without a shared framework or reliable measures, these dimensions remain difficult to assess.

From fascination to responsibility

The question posed to the people of Nantes is simple: how can we avoid repeating the ‘'limitless’ digital era of the 2000s? It is no longer a question of opposing progress and sobriety, but of inventing a framework where innovation is guided by measurement.

According to Arcep*** and ADEME

This is where Aguaro's proposal comes in: providing objective data, reliable indicators, and decision-support tools to help balance performance, costs, and impacts. This approach enables proportionate, useful and justified decisions to be made, rather than promises that are disconnected from material constraints.

Measuring to decide: from data to scenarios

You can only improve what you measure. Measuring doesn’t stop at kilowatt-hours — it means integrating multiple indicators (GHG, water, pressure on resources, etc.), as well as usage parameters and operational costs, in order to compare scenarios: should you host locally or in a remote cloud?

What is the trade-off between model size and efficiency? At what threshold does automation become counterproductive? The choice is not polarised between “pro” and “anti” AI, but rather oriented towards “ ’AI with a purpose’”.

In concrete terms, the Aguaro R&D team has identified a need to support its customers in controlled and responsible AI, initiating work to create a scoring tool and an AI action catalogue to translate complexity into operational levers: quantifying the footprint, comparing infrastructure, location and model size alternatives, and guiding choices towards more sober configurations without sacrificing performance.

Three practical and actionable steps

Drawing on our prior R&D work, we presented three key practices for more mindful AI use:

1. Question the necessity of using AI: consider alternatives, the time actually saved by using this AI, the purpose and the real need.

2. Favour local Artificial Intelligence: reduce data transfers, dependence on large operators, and bring computation and usage closer together.

3. Formulate precise queries and request short, grouped responses: limit unnecessary iterations, which multiply server calls.

These levers do not require technological disruption; above all, they require clarity in design and use — a change in method rather than tool.

Making the invisible measurable: orders of magnitude and current limits

Energy estimates per query vary greatly depending on the models and execution context. Published ranges show orders of magnitude ranging from fractions of Wh to tens of Wh per prompt; they must be interpreted with caution, as the calculation method and environment (hardware, PUE****, location) cause the results to vary. In other words, we still lack a consistent framework for analyzing the unit impact of the impact of a query, hence the need for a rigorous measurement framework. The ‘greenest’ option isn’t necessarily the most transparent noor energy-efficient— sometimes, it’s simply the least accountable…

Towards a common AI standard

Today, everyone evaluates in their own way.

As a result, the incomparability of data is causing confusion. A promising path forward lies in co-developing a shared framework that combines Bilan Carbone®, Life Cycle Assessment (LCA) and local contextual data (electricity mix, PUE, usage profiles), not to force standardisation, but to make choices comparable and therefore open to collective discussion. This is essential for the consistent management of public policies, corporate strategies and research projects.

Nantes as a testing ground for sustainable digital practices

The Nantes ecosystem illustrates a pragmatic approach: a committed university, active local authorities and a network of companies that already integrate environmental measures into their practices. Making the Nantes metropolitan area a laboratory for responsible digital technology means combining technological power with balance and showing that innovation can go hand in hand with digital sobriety. It is in this context that Nantes Métropole shared their own experience.

A more lucid innovation

Artificial Intelligence can help reduce emissions, optimise processes and support research, but it also carries material impacts from manufacturing to end-of-life. TThe aim isn’t a ‘perfect’ AI, but a collective intelligence capable of managing its effects: measuring, comparing, arbitrating and adjusting. This is at the heart of innovation that is more conscious of its consequences and compatible with planetary boundaries and social floors, as modelled in Kate Raworth's Doughnut Theory*****.

* Graphics processing unit
** the stage when a trained model applies what it has learned to new data to generate a response
*** the French Regulatory Authority for Electronic Communications, Postal Services, and Press Distribution*
***** Power Usage Effectiveness*
***** This concept illustrates the key challenges of the 21st century through two concentric circles: an outer ring that represents the ecological limits we must not exceed to preserve our environment — what Kate Raworth calls the environmental ceiling — and an inner ring that represents the essential foundation for every human being — the social floor. The space between these two rings defines the area where humanity can thrive in balance with the planet — what she describes as the safe and just space for humanity.