There are moments when technology, in seeking to accelerate the world, ends up accelerating its own contradiction. Artificial intelligence is experiencing such a moment. Presented as a lever for decarbonization, in its current form it is becoming an amplifier of energy and environmental issues.
I do not claim to be an expert on the subject, and some who specialize in it speak much better and, above all, with more conviction than I do, but today we can no longer view technology from a purely technical or productivist perspective, denying its political, geopolitical, and societal impacts. Building economic performance at the expense of the ecosystem is only a temporary solution, and my long-standing admiration for Antoine Riboud compels me to remember that the economy, people, and society must move forward together and not be antagonistic. (50 years ago, the Marseille speech. And since then? Not much.) This is all the more true in the age of AI, even if technologies that are already well established in our practices are proving to be environmental disasters (Why AI will never be energy efficient and Digital technology and environment: intangible uses for a real impact).
But a few days ago, the Shift Project published a report that puts precise figures on the situation: the massive deployment of generative AI is propelling global data center electricity consumption on a trajectory that already exceeds any scenario compatible with carbon neutrality ([FR]Artificial intelligence, data, computing: what infrastructure is needed in a carbon-neutral world?)
This seems like a good opportunity to revisit the subject, putting it into perspective while sticking as closely as possible to what the report says. We may or may not agree with the report’s proposals, and my expertise in this area prevents me from going too far on the subject, but I think the findings themselves are clear. Then it’s up to each person to do what they want with them, to be more or less radical in the solutions they propose, or even to say that ultimately it doesn’t matter or that the cause is lost in advance.
In any case, the issue is no longer just technical, but political: who will control digital energy? As AI becomes a strategic priority, it is transforming energy infrastructure into an adjustment variable, reviving fossil fuel projects that were thought to be dead and buried.
In any case, the illusion of a dematerialized digital world is well and truly behind us: it has become industrial, heavy, and deeply physical once again.
In short:
- Artificial intelligence, although presented as a lever for decarbonization, is currently exacerbating energy and environmental issues due to its high consumption of electricity and material resources.
- The massive deployment of AI generates a “usage-infrastructure loop” where anticipation of future needs leads to energy-intensive investments even before concrete uses exist, thus accelerating the growth of electricity consumption.
- Promises of “green” electricity do not compensate for the current boom in the sector, which is leading to the reactivation of fossil fuel power plants, highlighting a growing dependence on energy rather than a real energy transition.
- The energy efficiency of technologies is no longer sufficient to offset the explosion in usage: every gain in performance is offset by growing demand, making it essential to consider functional sobriety.
- The Shift Project report calls for clear governance and carbon assessment criteria to regulate the use of AI, in order to reconcile technological innovation with environmental responsibility.
The promise of decarbonization masks an unprecedented energy rush
We had convinced ourselves that digital technology would help us move away from carbon. It was supposed to make networks smarter, industry more efficient, and society more energy-efficient. And yet, just as artificial intelligence is establishing itself as a symbol of modernity, it is restarting what we thought we had slowed down, namely the energy-consuming machine.
This report therefore serves as a reminder: AI is not suspended in the cloud, it is anchored in the ground, in steel, in copper, in the power plants that feed it. Behind every “prompt” is a server room, and behind every server room is a power line. The illusion of a “virtual” digital world no longer holds.
The AI moment when the engine goes into overdrive
We need to go back to November 2022. ChatGPT appears, and suddenly, the collective imagination is turned upside down.
AI is no longer a little-known industrial tool; it becomes a language reflex, a horizon, and ultimately an injunction. In two years, the world has reacted like a body on dopamine, caught between acceleration and euphoria of adoption. Usage explodes, infrastructure follows. And that’s where the well-known rebound effect kicks in: the more we use, the more we build, and the more we build, the more we use.
The Shift Project refers to a “usage-infrastructure loop””: the computing economy now operates on the promise of future demand, justifying massive investments even before usage exists. It is speculative growth, but “electric” in the negative sense of the term.
Between 2023 and 2030, global data center consumption could triple, rising from 530 to nearly 1,500 terawatt hours. These figures may seem abstract, but they become more meaningful when we consider that this is equivalent to the electricity consumption of a country such as France, Germany, or the United Kingdom, or 10% of global electricity consumption.
At this stage, it is no longer a digital infrastructure, it is a new energy sector.
And the rhetoric of “AI for climate” does not change anything: even the most optimistic scenarios predict up to 920 million tons of CO? emitted by the sector in 2035. Twice the amount of France to support an industry whose growth is paradoxically based on the energy it claims to optimize and feed models that are supposed to help us predict the consequences of what they are exacerbating.
Electricity, the new oil of AI
The major players in the digital world tell a beautiful story: that of green electricity, guaranteed by contracts intended to finance solar or wind farms elsewhere. A parallel accounting system where solar energy in Arizona redeems the conscience of the Irish cloud.
But on the ground, the facts are measured in megawatts, and in the United States, hyperscalers are saturating regional networks. To keep up, gas-fired power plants are being reopened. Meta is planning three, Microsoft is investing in the revival of Three Mile Island (nuclear power is the lesser evil, but it is still the plant that experienced a partial reactor core meltdown, the most serious nuclear accident in the United States), Amazon is dreaming of small nuclear reactors, and Google is drilling for geothermal energy 5 km underground.
So-called “clean” technologies are arriving too late to fuel the current boom, so we are burning gas because it is quickly available and easy to use, and what is being sold as an “energy transition” is looking more and more like a detour than a new trajectory.
Digital technology has not escaped carbon and is even re-entering it through another door: that of electrical dependence. And AI is its standard-bearer.
Energy efficiency: an old illusion
For a long time, the narrative was simple: each generation of processors consumes less energy to do more.
Researchers at Lawrence Berkeley Lab even promised a digital energy “ceiling” thanks to exponential efficiency gains. The result: in 2025, the ceiling has become the floor. Efficiency is increasing, yes, but usage is growing even faster. The rebound effect, that great forgotten factor in technological communication, has just regained the upper hand.
The Shift Project puts it bluntly: believing that efficiency is enough is to confuse performance with direction. We are moving faster, but not necessarily in the right direction. Every local optimization is swallowed up by global growth, which makes any gains marginal.
Servers are becoming more powerful, but they are also multiplying. Each GPU gains in thermal power, each data center doubles in size, each AI model requires a hundred times more calculations than the previous one.
Energy efficiency does not compensate for the sector’s energy gluttony.
Functional sobriety: the elephant in the room
Public debate focuses heavily on digital sobriety, which essentially boils down to technical adjustments.
But the real question is not whether AI consumes less per query, but whether the query makes sense. Functional sobriety is about questioning the need before modeling it, and forcing us to ask ourselves, “What is it really for?”
However, neither public policy nor business strategy asks this question. AI has become a strategic imperative, a mandatory part of any roadmap, even when it serves no purpose.
The Shift Project proposes a simple method: evaluate each use of AI in terms of its net energy-carbon benefit. If the impact is greater than the gain, it is abandoned. This may seem radical, but in their minds it is simply rational.
But to apply this principle, we need governance, a framework, and a hierarchy of uses. Today, there is nothing. No ceiling, no compass, no arbitration. Innovation advances in its own name, like an overheated power plant without a circuit breaker, a metaphor not so far from reality.
Regaining control over the course of progress
The Shift Project study does not condemn technology, but simply reminds us that it does not regulate itself.
One thing is certain: the myth of clean digital technology is crumbling. AI does not float above the world, but passes through it, ultimately heating it up, and since it mobilizes physical resources, it must be treated as a physical sector. With ceilings, quotas, and choices.
Putting technology in its place does not mean restricting it, but governing it. Saying that AI does not deserve its kilowatts is not opposing innovation, it is relearning how to choose.
True progress is not measured by the speed of calculation, but by the accuracy of the course. And as long as we confuse power with relevance, artificial intelligence will remain true to its name: brilliant, but artificial.
Bottom Line
Digital technology has never been immaterial, but AI reveals its material scope and the blind spots in its governance. As long as digital management remains outside the scope of energy planning, the trajectory will remain unsustainable: digital technology will advance in the opposite direction to reality.
It is not a question of slowing down technology, but of making it governable, i.e., measurable, capped, and above all, directed.
Reconciling intelligence (human, not artificial) and responsibility means placing innovation back within the chain of human choices, rather than above it.
To answer your questions…
Generative AI consumes enormous amounts of electricity because of the data centers that power it. The Shift Project estimates that this consumption could triple by 2030, reaching the level of entire countries. Each request activates servers, power lines, and cooling systems. Far from being immaterial, AI is becoming a heavy industrial sector, whose energy growth already exceeds trajectories compatible with carbon neutrality.
In theory, yes, because it can optimize certain processes. In practice, no, because the growth of its infrastructure cancels out the gains. By 2035, the sector could emit up to 920 million tons of CO?. For it to become a real climate lever, its use would have to be reserved for applications with proven carbon benefits and its energy consumption would have to be regulated.
Energy efficiency is improving, but usage is skyrocketing. This is the rebound effect: the more efficient a technology becomes, the more it is used, which drives up overall consumption. Servers are multiplying and AI models are becoming increasingly power-hungry. In short, we are moving faster, but not necessarily in the right direction.
It involves questioning the real usefulness of a use before developing it. The Shift Project recommends evaluating each application according to its net energy-carbon impact and abandoning those that exacerbate the problem. This approach requires prioritizing uses and clear governance to put innovation back into a context of meaning and responsibility.
It must be treated as an industrial sector subject to physical constraints. This involves consumption caps, carbon quotas, and appropriate energy planning. Governing AI does not mean slowing it down, but rather choosing its uses according to their real usefulness and energy cost, in order to reconcile technological progress with ecological balance.
Image credit: Image generated by artificial intelligence via ChatGPT (OpenAI)
Related articles:
Why AI will never be energy efficient
Like the climate, does AI deserve its “COP”?
AI in the workplace: avoiding the Wall-E effect
Generative AI in the enterprise: a silo breaker or just another layer in an already complex IT landscape?
Why enterprise AI can’t keep up with consumer AI: beyond ChatGPT, a more complex reality
AI in the digital workplace: a brilliant assistant, but an unreliable colleague
