Data centers have become the back office of the modern economy. They host professional applications, paperless public services, video platforms, e-commerce, social networks, banking services and artificial intelligence tools. Their utility is in no way debated. However, their energy appetite is a different matter.

According to the International Energy Agency, data centers consumed approximately 415 TWh of electricity worldwide in 2024, representing nearly 1.5% of global electricity demand. This volume has grown by about 12% p.a. over the last five years. The road ahead looks even steeper: in its central scenario, the IEA forecasts that this consumption will nearly double by 2030, to approximately 945 TWh, or just under 3% of global electricity demand. This implies an average growth of about 15% p.a. between 2024 and 2030, more than quadruple the rest of electricity consumption. Accelerated servers, those driving the rise of AI, are expected to see their consumption climb by about 30% p.a., representing nearly half of the net increase in data center consumption over the period.

These figures must be handled with caution, as not all data centers are created equal. A small regional site bears no resemblance to a hyperscale campus designed to train or run AI models. However, the trend is clear: digital technology is becoming a major electricity consumer.

AI changes the orders of magnitude

The growth of data centers is not new. What has changed is the intensity. Generative AI models rely on very dense, very expensive and very hot GPU clusters. In older centers, power density per rack remained compatible with traditional cooling architectures. In new AI factories, that world is already a thing of the past.

Requirements per rack are soaring. Hyperscalers no longer think only in square meters, but also in available megawatts, connection lead times, cooling capacity and access to stable electricity. The global map of data centers is thus becoming an energy map. Where power is abundant, reliable and relatively cheap, projects move forward. Where it is lacking, ambitions slow down.

This is why the United States, China, the Nordic countries, certain Gulf states, and a few better-equipped European zones are attracting attention. The issue is no longer just fiscal, regulatory, or real estate-related, it is becoming physical. A region may have engineers, capital and available land; but without electrical power, it remains sidelined.

The power grid becomes the ultimate arbiter

Public debate often focuses on total consumption. This is useful but insufficient. An electrical system can produce enough energy over the year and yet be unable to quickly absorb several giant projects in the same area. The problem then stems less from national volume than from local bottlenecks: saturated lines, insufficient transformers, connection delays - and the need to reinforce infrastructure.

The United States offers a glimpse of this tension. Data centers are concentrated in a few highly sought-after states, notably Virginia, Texas, and N&S Carolina. In these areas, the growth of digital projects can create pressure on wholesale prices, grid reliability, and the investments needed to keep pace. Several studies suggest that the national impact on prices may be moderate, while local effects become much more visible.

This is the data center paradox: on a national scale, it may seem absorbable, while on a local scale, it can become overwhelming. A handful of very large sites is enough to alter a region's electrical balance. And when grids must be reinforced, the bill is never entirely private. It eventually trickles down to end customers, transmission tariffs, or public finances.

Electricity, AI's new hidden cost

AI is often marketed as software. In income statements, however, it also appears as an energy expense. For data center operators, electricity is one of the primary cost items. For customers, it is reflected in cloud pricing. For households and businesses, it may return indirectly via higher grid costs.

In the US, economists have estimated that the rise of data centers has already contributed to increasing electricity production costs by 5% to 15%, depending on usage assumptions. If the announced construction wave fully materializes, the impact could become significantly more pronounced by the end of the decade. The most extreme scenario will not necessarily come to pass: not all announced projects will be built, and the euphoria surrounding AI will proably eventually separate profitable uses from costly demonstrations.

But even with such a sorting process, pressure will remain high. Models are being integrated into search engines, enterprise software, programming tools, cybersecurity, healthcare, finance, and industry. AI will not just be used by a few platforms; it will be disseminated everywhere. This is precisely what makes it energy-sensitive.

The winners: power producers, equipment suppliers, capacity owners

This new reality is reshuffling the pack. Electricity producers, gas suppliers, renewables developers, electrical equipment manufacturers, cooling specialists, data center REITs, and power component suppliers find themselves at the center of the game. AI creates additional demand for assets that the market sometimes considered boring: power lines, transformers, generators, cooling systems, pumps, heat exchangers, cables, and substations. Glamour lies with generative models; steady monetization can also be found amongst those selling the energy "picks and shovels". Major cloud operators have understood this well. They are signing long-term supply contracts, investing in renewable capacities, looking closely at nuclear power, and seeking sites capable of offering power quickly. The choice of a data center location increasingly resembles that of an electro-intensive factory.

The losers: saturated grids and passive consumers

The other side of the coin concerns territories that host too much, too fast. A giant data center can create activity, tax revenue, and a few skilled jobs. But it can also consume a significant portion of available local capacity, complicate grid access for other industries, and accelerate investments that will be mutualized. Households do not need to use AI massively to pay for part of it. If electrical infrastructure is reinforced to absorb new loads, the cost can end up in grid tariffs. If demand increases in areas supplied by marginal fossil fuel plants, wholesale prices and emissions follow. If operators capture the best available capacities, other industrial projects may be delayed. This is where AI ceases to be a matter for geeks or financial markets. Instead, it becomes a subject of energy planning, regional development and cost-sharing.

Climate in the equation

The carbon impact depends heavily on location. A data center connected to a low-carbon grid does not have the same footprint as a site powered by gas or coal. This is obvious, although often forgotten in global discourse on digital technology. The same AI query may have a different footprint depending on the time, the region, and the electricity mix used. This is why geography becomes essential. Deploying data centers in regions where electricity is abundant and low-carbon can reduce climate impact. Installing them in already stressed areas, where marginal demand is met by fossil fuel plants, can instead worsen emissions. AI is not inherently brown or green. It takes on the color of the grid that powers it. This reality makes international comparisons more interesting. Countries with a low-carbon mix, a robust grid, and competitive energy have a card to play. Others risk seeing projects go elsewhere, or hosting them at the cost of rising emissions and bills.

Digital sovereignty goes through the power outlet

Digital sovereignty is often approached through the lens of data, software and semiconductors. This is insufficient. A country that wants to develop its own AI capabilities must have computing power, therefore data centers and therefore electricity. Sovereignty is not decreed in a press release. It is connected to the grid. This is a delicate trade-off. Depending on foreign data centers poses problems of security, control, confidentiality and imported carbon footprint. Building locally requires reserving a portion of the electrical system for these uses. Between external dependence and internal saturation, there exists a zone of intelligent planning, although it must be occupied. Public authorities will therefore have to prioritize. Not all AI uses are equal. Accelerating medical research, optimizing a power grid, or improving industrial productivity does not have the same collective interest as infinitely generating promotional content or valueless synthetic videos. Electricity is a strategic resource; it would be strange to treat it as a help-yourself, all-you-can-eat, buffet.

The meter is running

The AI revolution is often told as a story of models, chips, and stockmarket valuations. It is also a story of electrical transformers, liquid cooling, connections, megawatts, and heat dissipation. It is less romantic, although much more decisive. Digital technology long promised to dematerialize the economy. AI reminds us that nothing truly disappears. Servers heat up, grids take the strain, power plants produce and bills circulate. And behind every instantaneous response from a generative model, there is now a very concrete question: who provides the power, at what price and for what purpose?