Energy: who will pay for the gigawatts?

10.1AI's energy bill

In context

How much does a query cost?

Reduced to a single query, the cost looks tiny. In 2025, Google published a detailed estimate: a median text query to its Gemini assistant consumes about 0.24 watt-hour of electricity (the equivalent of a few seconds of television) and 0.26 milliliter of water (about five drops), with efficiency rising sharply (a more than thirtyfold reduction in a single year). Independent estimates for the other assistants converge around 0.3 watt-hour per text query; the older estimate of 3 Wh is now considered too high. Two important caveats: this is a median, and a long query or a reasoning model (Chapter 4) can consume ten to a hundred times more; above all, generating an image or a video costs far more (up to nearly a kilowatt-hour for a few seconds of video). The real issue, then, is not the isolated query but its multiplication by the billions: a large service can handle several billion queries a day.

10.2Water, electricity, carbon footprint

In plain terms

Beyond electricity, two impacts are less visible.

Water. Cooling tens of thousands of scorching chips consumes enormous volumes of water (a very large data center can use several million liters a day), creating local tensions, especially in dry regions. The case of the Mexican state of Querétaro is emblematic: while enduring one of its worst droughts in a century, it has become a major data-center hub (several billion dollars invested, dozens of sites planned), to the point that residents subject to rationing denounce a logic that "puts servers ahead of citizens." Comparable conflicts have erupted in Uruguay (2023) and Arizona. Operators respond with air cooling (which saves water but consumes more electricity) or "water positive" pledges, but the lack of transparency about actual volumes fuels distrust.
Carbon. When electricity is not enough, operators hastily install gas turbines (as seen with the Colossus 2 site in Chapter 8), or even convert old aircraft engines into generators. The real carbon footprint of some sites may be far higher than the figures reported.

To these impacts must be added an effect on grids and bills. In the United States, the rise of data centers could push the average electricity bill up by around 8% by 2030, with much steeper increases in saturated areas (northern Virginia, for example). Hence a question of fairness: will local residents pay, on their bills, for the energy appetite of the AI giants?

To these effects two blind spots must be added. The first is grid stability: to avoid depending on a saturated grid, some operators build their own on-site generation (gas, solar, batteries, even dedicated nuclear), a trend known as "behind the meter" that in turn raises questions of emissions and fairness. The second is the material footprint: manufacturing millions of chips and servers consumes critical minerals (copper, rare earths) and generates, as hardware is refreshed, a growing mass of electronic waste. AI's footprint, then, cannot be reduced to the electricity it burns today.

The bottleneck. The most structural problem is a timing mismatch, about which analysts (such as the Uptime Institute) are warning as early as 2026:

Diagram10.1. The energy "scissors." A data center is built far faster than the power plant meant to supply it. This timing gap is, according to several experts, the main physical brake on AI's expansion in the years ahead.

Note: not all countries are in the same boat. In France, for instance, the grid operator (RTE) reckons it can absorb the expected needs of data-center projects (on the order of 4 GW), thanks to an already largely decarbonized electricity mix.

Under the hood

Cooling the chips

A GPU computing at full tilt gives off enormous heat, and a data center packs in tens of thousands of them: removing that heat is a challenge in its own right, and it explains much of the water and electricity consumption mentioned above. Traditional air cooling (fans, air conditioning) is reaching its limits against AI chips. So the shift is increasingly toward liquid cooling: water or a fluid circulates as close as possible to the chips to carry away the heat, far more efficiently than air. The most advanced form, immersion cooling, plunges the servers outright into a non-conductive liquid. These techniques reduce the energy spent on cooling, but can increase water consumption, hence delicate trade-offs, particularly in water-stressed regions. Cooling has thus become a field of innovation as strategic as the chips themselves.

In context

A data center's efficiency (PUE) and waste heat

To measure a data center's energy efficiency, a standard indicator is used, the PUE (Power Usage Effectiveness): the ratio between the total energy consumed by the facility and the energy that actually goes to computing. A PUE of 1.0 would be ideal (all the energy goes to the chips); in practice, the best centers approach 1.1, whereas older facilities exceed 1.5, the surplus going mostly to cooling (previous point). Reducing PUE is therefore a direct lever, through more efficient cooling and careful design. A complementary path is to make use of waste heat: rather than expelling it, it can be recovered to heat buildings or district heating networks, as some Nordic data centers already do. A data center then ceases to be a pure energy dissipator and becomes a useful heat source, turning a nuisance into a resource.

10.3Solutions: efficiency, nuclear, renewables

In plain terms

Three broad avenues are being mobilized to meet this demand.

Efficiency. Each generation of chips computes more per watt consumed (NVIDIA's Vera Rubin systems claim a leap in inference efficiency). Add to that quantization and specialized small models (Chapter 9), which do as much with far less. This is probably the most immediate lever.
Nuclear. This is the big trend of 2025–2026: tech giants are signing deals to power themselves with nuclear electricity, whether by restarting old plants or financing small modular reactors (SMRs). Nuclear energy offers stable, low-carbon electricity, but it runs into a major obstacle: it takes more than ten years to build a plant, whereas a data center rises from the ground in three.
Renewables. Solar and wind are advancing fast and cheaply. Strikingly, in 2025, renewable energy surpassed coal in the global electricity mix for the first time; in the United States, however, low-carbon sources (renewables and nuclear) still supply only around 43% of electricity, the rest remaining mostly fossil-fueled. But their intermittency requires pairing them with storage or other sources. Gas, quicker to deploy, often serves as a bridge solution, at the expense of climate targets.

In context

The grid-connection wall

Having electricity is not enough: a data center still has to be connected to the grid, and this has become a major bottleneck. In several countries, hundreds of gigawatts of projects (both generation and consumption) are waiting in a connection queue that can last years, for want of high-voltage lines and sufficient transmission capacity. For operators of "AI factories" who want to build fast, this delay is a headache. Hence workaround strategies: setting up where electricity is already available, signing direct deals with power plants (including nuclear ones, previous section), or generating one's own electricity on site (gas, solar, batteries), at the risk of worsening the carbon balance. The real limiting factor on AI's expansion is therefore not always the chip: increasingly, it is access to connected electricity, and the grids are struggling to keep pace.

10.4The climate debate

On one side, AI represents a new and massive load on grids already under strain, at the risk of prolonging reliance on fossil fuels and passing the bill on to citizens. On the other, its defenders stress its potential for the climate: optimizing electricity grids, accelerating the discovery of new materials (better batteries, carbon capture), improving climate modeling, or advancing research such as nuclear fusion (themes we will revisit in Chapter 14 on science).

The question that sums it all up, and that gives this chapter its title, remains open: who will pay for the gigawatts? The companies that profit from them, or society as a whole? And will the scientific and economic benefit justify the footprint? There is, in 2026, no clear-cut answer, but one certainty: energy has become the number-one limiting factor in AI's trajectory, as much as chips.

In context

AI bubble or rational bet?

The energy bill is only part of a broader question: are these colossal outlays sustainable, or are we witnessing a bubble? The figures are dizzying. The tech giants' AI investments ("capex") run into the hundreds of billions of dollars a year and are constantly being revised upward; some estimates put global AI spending at several trillion for 2026. Three signals fuel the fears. First, the gap between spending and revenue: leading players spend several dollars for every dollar of revenue, and an often-cited study found that the vast majority of companies were still seeing no measurable return from generative AI. Next, cross-financing (a chip maker invests in a lab, which uses that money to buy its chips; a cloud provider invests in a customer who rents its cloud), an arrangement that analysts deem "circular" and reminiscent of the internet bubble. Finally, debt and systemic risk: central banks, the IMF, and rating agencies have warned of a possible correction and its potential spread to the economy. Optimists counter that the spending answers real and growing demand, that the infrastructure (unlike the marketing of the 2000s) remains reusable, that it is backed by very solid giants, and that investors now "police" profitability (a sharp stock-market drop at the slightest hint of overinvestment). The truth comes down to a single question: will this spending create real use value to match the promises, or is the money merely going in circles? Notably, AI investment now accounts for a significant share of US growth: if the bubble were to deflate, the shock wave would reach well beyond the sector (Chapters 7 and 18).

Key takeaways (Chapter 10)

Data centers' electricity consumption (~415 TWh in 2024) could approach 1,000 TWh in 2026 (the equivalent of Japan) and triple by 2030.
Beyond electricity, AI weighs on water (cooling), carbon (backup gas turbines), and local residents' bills.
A bottleneck looms: a data center is built in 3 years, but a power plant (especially nuclear) takes far longer, hence a "scissors" gap between demand and generation.
Three solutions: efficiency (better chips, quantization, small models), nuclear (giants' deals, but long lead times), and renewables (booming, but intermittent).
The climate debate is open: a new load on grids versus the potential to accelerate the transition. Energy has become AI's foremost limiting factor.

So ends Part III. We have seen the fuel (compute, hardware, energy) that propels AI. Part IV explores the great convergences where AI meets other disruptive technologies: blockchain, quantum computing, and robotics.

10.1AI's energy bill#

10.2Water, electricity, carbon footprint#

10.3Solutions: efficiency, nuclear, renewables#

10.4The climate debate#

Key takeaways (Chapter 10)

10.1AI's energy bill

10.2Water, electricity, carbon footprint

10.3Solutions: efficiency, nuclear, renewables

10.4The climate debate