Artificial intelligence is reaching a new scale. As it moves into production, companies can no longer ask only what AI can do. They must also ask what resources it takes to do it.
Computing power, infrastructure, models, query volumes, energy: every use draws on physical and technological resources that carry both an environmental footprint and a cost.
That is one of the lessons to be drawn from the fact-finding report on the environmental footprint of artificial intelligence, unanimously adopted on September 16, 2026, by the French Senate’s Committee on Regional Planning and Sustainable Development. Christian Cor, CEO of Saaswedo, testified as part of this work in his capacity as a board member of Numeum, the French trade association for the digital industry, and its environmental lead.
Yet the report goes well beyond energy alone. Its assessment covers AI’s full footprint: electricity, with global consumption that could reach about 500 TWh by 2030; emissions, which depend heavily on the electricity mix of the host country; raw materials, with graphics processors that are replaced rapidly and consume large quantities of critical minerals; and finally water and land.
Its 16 recommendations gradually outline what amounts to genuine governance of the resources AI consumes.
Three Levels of Action to Reduce AI’s Footprint
The Senate’s recommendations are organized around three main focus areas.
The first aims to make environmental excellence a competitive advantage in attracting data centers to France.
The Senate is not seeking to slow their development. On the contrary, it recommends accelerating their connection to the power grid and removing administrative hurdles, building on a key asset: electricity that is 95% decarbonized. At the same time, it proposes rewarding the most virtuous facilities with a reduced electricity tax rate, recovering waste heat where feasible, better integrating data centers into local communities, and enlisting them to help smooth out electricity demand.
The stakes are significant. The report stresses that generative AI is changing the scale of infrastructure: while a conventional data center typically has a capacity of around 10 MW, the infrastructure required for AI can reach 100 MW, or even 1 GW.
The second focus area is the development of frugal AI.
The Senate recommends supporting small, specialized models, strengthening the eco-design of digital services, and making greater use of environmental criteria in public procurement. These principles should also be championed at the European level.
The third focus area covers measurement and usage.
The report calls for the publication of data on the environmental footprint of models, the creation of a European assessment standard on which any environmental claim would have to be based, the introduction of an eco-score, and the development of awareness and training initiatives.
These three dimensions are closely linked. To reduce AI’s footprint, you first need to understand what is being used, measure the resources involved, and then act on their consumption.
Measure Before You Optimize
Seen from the enterprise, these recommendations reach beyond public policy. The report places particular emphasis on one difficulty: businesses and public authorities alike still have little comparable data on the footprint of different models.
The life cycle assessments published by some providers use different scopes and methodologies, which currently limits comparisons. The Senate therefore calls for greater transparency and a common assessment standard.
This principle extends well beyond the environmental issue. Within a company, it is just as difficult to control AI costs without visibility into consumption:
- Which model is being used?
- For which use case?
- By which team?
- At what query volume?
- With what token or compute consumption?
- At what cost?
- And for what value delivered?
An overall invoice provides accounting information. It does not necessarily make it possible to manage usage.
Governance therefore starts with visibility.
The Most Powerful Model Is Not Always the One You Need
The Senate’s recommendation in favor of small, specialized models is a particularly good illustration of how environmental sustainability and cost control converge.
Not every use case requires the same capabilities. Classification, information extraction, summarization, translation, complex generation, reasoning, and document analysis do not call for the same models or the same computing resources.
Systematically using a very powerful general-purpose model for simple tasks can mean consuming more resources than necessary. Numeum makes this point in its contribution to the report: for the same result, an ill-suited model could use 10 to 15 times more electricity than another model.
The goal of frugal AI, then, is not to arbitrarily limit the capabilities available to users. It is to size resources to the intended outcome.
In other words, it means applying to AI a principle already familiar in the cloud: right-sizing.
From Environmental Sustainability to Cost Control
This principle has a direct economic consequence. Compute, GPUs, storage, data transfer, and inference all come at a cost.
When usage is limited, a difference in consumption between two models may seem minor. When a process runs hundreds of thousands or millions of times, it becomes structural.
Several levers can then act on both technical consumption and cost at the same time:
- Choosing the right model for the workload
- Controlling prompt and response length
- Cutting unnecessary calls
- Using caching where relevant
- Pooling certain workloads
- Managing usage volumes
That said, environmental footprint and financial cost are not equivalent metrics. The price billed also depends on vendor pricing policies, discounts, contractual commitments, the architecture used, and the location of the infrastructure.
It would therefore be an overstatement to claim that cutting costs automatically reduces the environmental footprint in the same proportion. The two disciplines do, however, share many efficiency levers.
Usage Becomes a Governance Variable
The Senate does not limit its recommendations to infrastructure and models. It also gives significant weight to user awareness and behavioral change.
This is an important shift. As generative AI is rolled out to thousands of employees and embedded directly into business applications, consumption is no longer determined by infrastructure alone. It also depends on usage.
Numeum, cited in the report, considers the employee’s choice of model to be decisive: a poor choice can increase the footprint tenfold.
The question is therefore no longer just “How much does our AI solution cost?” but:
“What is driving this consumption, and is that consumption proportionate to the results achieved?”
This is exactly the kind of question companies have already had to learn to address with the cloud and FinOps.
Toward AI Cost Management
The principles of Technology Expense Management and FinOps can thus be gradually applied to artificial intelligence.
The first step is gaining visibility into the vendors, services, and models in use. The next is tying consumption to teams, applications, and use cases. Then come unit cost analysis, identifying cost overruns, optimizing resources, and putting governance rules in place.
The goal is not simply to spend less. It is to weigh the resources consumed against the value delivered.
That is what fundamentally brings cost control and frugality together.
Governing AI, Not Just Adopting It
The Senate report is not about AI Cost Management. Its recommendations are driven first and foremost by environmental and public policy objectives.
But they highlight a principle that directly concerns businesses: scaling AI requires tighter control over the resources it consumes.
- Infrastructure.
- Models.
- Computing power.
- Volumes.
- Usage.
- Impacts.
- Costs.
As AI becomes a permanent part of enterprise IT, these dimensions can no longer be managed in isolation. Organizations will need to be able to measure, make trade-offs, and optimize.
The challenge is therefore no longer just adopting AI. It is learning to govern it.