Electricity under tension: the dark side of AI

Every day, artificial intelligence is gaining ground: in our search engines, our creativity tools, our administrative and healthcare services. Behind this veneer of algorithms, we often forget the immense physical back office that powers these calculations—thousands of data centers, GPU farms, networks that never sleep. However, this hardware layer is becoming one of the critical nodes of the French and European energy transition, as its appetite for electricity is growing faster than that of any other industrial sector.

Global data center energy consumption reached 415 TWh in 2024, according to the International Energy Agency (IEA), with a growth rate that has accelerated since the late 2010s. It could reach approximately 945 TWh in 2030 according to the IEA's central scenario, more than double current demand. The share attributable to AI applications, still a minority in 2025 (≈ 15%), is expected to exceed one-third before the end of the decade. This growth is not linear: it is self-sustaining. Each new data center lowers the marginal cost of a digital service, unlocking new applications—such as the integration of generative models into office suites, embedded chatbots, and synthetic videos—which, in turn, justify the expansion of infrastructure.

This supply-use dialectic is already visible in Europe. In Ireland, for example, data centers absorb more than 20% of available electricity, exceeding the consumption of urban residential areas. Across the continent, demand could double between 2023 and 2030 and then quadruple by 2035, without national planning scenarios having truly internalized this shock. However, the European Union does not have a completely decarbonized electricity mix; achieving climate objectives therefore requires reserving low-carbon electrons for high-leverage sectors (mobility, heavy industry, heating), not letting them flow into digital services whose utility is sometimes gimmicky.

Across the Atlantic, the tension is being addressed differently: the preferred response to the capacity shortage is to connect new gas-fired power plants, viewing energy scarcity as a mere logistical obstacle. Such a headlong rush shifts the climate burden onto the energy systems themselves and fuels a trajectory incompatible with carbon neutrality: emissions could thus peak at 920 million tonnes of CO₂ equivalent per year, twice the annual budget of a major European country.

France is not immune. Connections already approved will reach full capacity around 2035, at which point they risk finding themselves in direct competition with the electrification of transport or industrial heating. If recent announcements materialize, the share of data centers in national electricity consumption would jump from around 2% today to 7,5% in 2035. When compared to industrial electricity alone, it could represent a third of total demand. Remaining on this trajectory would jeopardize the country's decarbonization objectives before they have even produced their first effects.

Faced with this horizon, three structuring levers are essential. The first is transparency: mapping all sites, monitoring their consumption in real time, and publishing this data in a format that can be used by both network managers and civil society. Without this observatory, any energy-carbon planning will remain theoretical. The second lever is the establishment of an electricity consumption cap for the sector, rolled out at the national level and revised periodically in light of actual progress on the carbon intensity of the mix and the manufacturing footprint of equipment. The third concerns the allocation of uses: prioritizing AI applications that present a net collective benefit and questioning those that simply stimulate overconsumption.

At the scale of each service, the key lies in understanding the "two stages" of a model's life cycle. Training, which is one-off but extremely resource-intensive, has seen its carbon footprint grow exponentially over the past ten years. Inference, on the other hand, is proportional to the volume of users; when deployed massively, it exceeds the impact of training in a few weeks. The technical determinants are known: number of parameters, type of task (image generation is more energy-intensive than text classification), pipeline depth, hardware performance. Therefore, action must begin in the design phase: compress networks, favor low-precision architectures, and split features to activate only those that meet a clearly articulated need.

A robust decision matrix should systematically compare each AI feature to a reference energy-carbon budget. When design levers (lighter models, optimized hardware) and deployment levers (restricted access, quotas, incentive pricing) are not sufficient to respect this budget, the default option should be to abandon the feature or replace it with a non-AI solution. This approach, far from restricting innovation, requires us to distinguish between innovation for convenience and innovation for necessity.

The success of this transformation is not solely the responsibility of engineers; it requires broader governance. Host companies must report on their capacity plans and commit to trajectories compatible with national climate scenarios. Public authorities, for their part, have the responsibility to direct investments toward efficiency (adiabatic cooling, waste heat recovery, local data centers powered by renewable energy) and to condition any new connections on a broader cost-benefit analysis. Finally, public debate cannot be captured by the technological narrative alone: ​​developing skills for digital sobriety also means reallocating cognitive resources currently mobilized by the media hype surrounding AI.

It is tempting to consider digital growth as inevitable, "natural." However, energy history shows that trajectories are always the result of political and industrial choices. It is precisely in this 2025-2030 window that the balance is being struck: either the digital sector proves that it can grow while aligning with planetary boundaries, or it will become the weak link in the transition, rendering the gains made elsewhere obsolete. Approaching the subject from the angle of sobriety does not mean giving up on AI; it means using it where it creates societal value without threatening the common carbon budget. This requirement for efficiency, long confined to laboratories, must now permeate the strategic roadmaps of operators and governments alike.

In the short term, the priority is therefore twofold: adopt a public and binding measurement system, then set an indicative consumption cap that is broken down by type of service and by use. In the medium term, the challenge will be to couple this cap with market mechanisms and eco-design obligations, so that the cost of the virtual kilowatt-hour finally reflects its real impact. In the long term, we cannot avoid the question of the social relevance of the calculations performed: carbon neutrality requires as much the avoidance of superfluous calculations as the quest for ever more precise yields.

Digital technology is often touted for its ability to "dematerialize" the economy. However, it is instead materializing a growing portion of our energy and resource consumption. The next frontier of AI will not be technical but political: it will be about showing that a connected society can remain within a global framework, by lucidly balancing computing power with the power of social transformation.

To go further: this energy issue continues with our analysis of energy as a competitive advantage et our article on cognitive dependence and the ecological cost of AI.