Artificial intelligence is advancing at a spectacular pace, with each new generation of models proving more powerful, faster, and more optimized than the last. There is talk of “lightweight” models, reduced consumption, “frugal” computing, and the idea that AI could become energy efficient is gaining traction among industry players.
But this sobriety is largely illusory. Not because the technologies are inefficient, but because they are, on the contrary, too efficient for us to remain reasonable in our use. Like other sectors before it, AI is entering a cycle where energy efficiency gains fuel a massive intensification of use, resulting in the opposite of lower consumption.
In short :
- The increasing efficiency of AI models reduces their unit cost but encourages a massive increase in their use, leading to higher overall energy consumption.
- Technical efficiency gains at the individual level do not translate into energy efficiency at the system level due to the rebound effect.
- The proliferation of everyday uses of AI, made possible by falling marginal costs, is contributing to the constant expansion of its energy footprint.
- Lightweight and local models do not guarantee lower consumption, as they facilitate uncontrolled diffusion of uses without overall supervision.
- True sobriety requires explicit collective limits on digital uses, through political choices and regulations, beyond technical innovations alone.
Is AI becoming more energy efficient? An appealing but illusory idea
Technical progress is undeniable: thanks to model compression, distributed computing, quantification, and framework optimization, it is possible to reduce the carbon footprint of a model or query. On a unit scale, this is concrete and measurable.
But these gains do not mean that AI consumes less overall—in fact, the opposite is true. In fact, the more energy-efficient models are, the lower their unit economic and ecological cost, the fewer barriers there are to using them, and therefore the more we use them. This dynamic is well known: it is called the rebound effect and has been documented since the 19th century in the coal industry, but it can be found in all technologies whose use is based on a balance between cost and utility.
Doing more, all the time
In the automotive industry, more fuel-efficient engines have made it possible to produce heavier, more powerful vehicles that are used longer and travel faster. In aviation, improvements in passenger efficiency have led to an explosion in global traffic. The unit consumes less, but the system consumes more.
AI is following the same trajectory. As the marginal cost of a query decreases, the number of use cases multiplies. Conversational AI is becoming an integrated feature of everyday tools, accessible in office suites, browsers, customer service tools, and business software. In fact, what was once occasional is becoming frequent, and what was once reserved for complex tasks is now being used for the most mundane tasks.
We are not more energy efficient with constant use, but the decline in unit energy costs is leading to an explosion in the scope of uses and overall energy impact.
“Small models”: a false good idea
Some argue that the solution will come from lighter, locally embedded, less power-hungry models. This is indeed an interesting idea, but it does not solve the only question that matters, that of the global consumption limit.
If every device is capable of running its own AI assistant, consumption becomes diffuse and much more difficult to control. As we have seen, lower consumption per unit does not mean lower total consumption, but simply allows for a natural and spontaneous increase in the number of units and uses, without any questions being asked.
This is where the concept of collective limits, proposed by some specialists, comes into play. It is based on a simple observation: as long as no overall energy constraints are set, efficiency gains automatically translate into increased intensity. A more energy-efficient AI will not consume less if it is used a thousand times more.
A collective limit is an explicit limit, set for a given perimeter (business, sector, country), which frames all digital uses, including AI (because it is far from being the only one involved (Digital technology and environment: intangible uses for a real impact and Sustainable digital: no more hypocrisy)) based on a non-extendable energy budget.
This can take several forms:
- a consumption ceiling assigned to a service or business,
- a quota of computing or AI requests allocated according to defined criteria,
- differentiated regulation of use cases according to their usefulness,
- or taxation integrated into carbon footprints.
This can take several forms:
- a consumption ceiling assigned to a service or business,
- a quota of computing power or AI requests allocated according to defined criteria,
- differentiated regulation of use cases according to their usefulness,
- or taxation integrated into carbon footprints.
The issue is no longer just technical, but political. It is not optimization that produces sobriety, but the trade-offs made about what we allow to be consumed and why.
Bottom line
Thinking that AI will become more energy-efficient solely through technological progress is to confuse local efficiency with systemic control. As long as there are no collective limits on its expansion, every technical advance, however virtuous, fuels a dynamic of acceleration.
Sobriety will not result from smaller models or greener data centers, but from our ability to structure uses, prioritize useful functions, and make collective choices about what deserves to be addressed by AI, of course, but also by digital technologies in general.
Image credit: Image generated by artificial intelligence via ChatGPT (OpenAI)
Related articles:
Like the climate, does AI deserve its “COP”?
AI in the workplace: avoiding the Wall-E effect
AI Summit: a real success or a smokescreen for Europe?
AI: who will impose their law on the geopolitical scene?
Digital technology and environment: intangible uses for a real impact
What if China made AI free of charge: chronicle of a global shift
