What is the most serious environmental risk of AI?

AI weighs on the environment through 5 direct pressures: the electricity it consumes, the CO2 emitted to produce that electricity and the hardware, the water for cooling and manufacturing, the metals in its chips and servers, and the land occupied by its data centres. The figures for each circulate separately, in different units. Electricity is the origin of most of the other pressures: we treat it here for what it does to grids, and CO2 for what it does to the climate, without adding the two. To compare them, we ranked them by 3 criteria: their magnitude, that is, their share of the world or national total; their scale, local or global; and the duration and reversibility of their effects. Indirect effects, such as AI in the service of oil and gas extraction, are covered in the analysis on the climate crisis.

Data centres emitted around 180 million tonnes of CO2 in 2024 for their electricity, according to the IEA, 0.5% of global combustion emissions; AI is only a part of this. GreenIT estimates AI emissions at 99 million tonnes of CO2e in 2025, manufacturing included, and 636 million in 2030 in its projection at constant electricity mix. This is small on a global scale, and it is growing fast: the IEA expects gas and coal to cover more than 40% of the increase in data centre demand by 2030.

Each pressure, set against its own global total, cannot be compared with the others: 0.5% of combustion emissions for data centre CO2, 1.5% of electricity demand, 11% of current gallium demand in 2030. Only a weighting brings them to a common scale: in GreenIT's, global warming accounts for 35% of the footprint of AI, metals 2% and water deprivation 1%. CO2 then stands out by its duration. CO2 mixes throughout the atmosphere, wherever it is emitted, and part of it remains there for between one year and several thousand years, according to the IPCC. The changes caused by greenhouse gases, of which CO2 is the main one, in the ocean, the ice and sea level are irreversible on a scale of centuries. Part of the CO2 emitted in 2026 by a data centre's electricity will still be acting on the climate at the end of the century. GreenIT also points out that about 65% of this weighted footprint is not climate-related: freshwater eutrophication accounts for 23%, depletion of fossil resources 17% and fine particles 15%. These come mainly from the same source, since the use phase of servers, that is their electricity, accounts for 92% of impacts in 2025. Air and water pollution from power plants is not part of our ranking, which sticks to the 5 most frequently cited pressures.

Worldwide, data centres consumed a little over 1.5% of electricity in 2025, and they account for only about a tenth of the increase in global demand by 2030, less than air conditioning or electric cars. But they are concentrated: according to the IEA, they account for up to 20 to 30% of electricity demand in the regions that host them. This is where the pressure is visible, on grids, on prices and on the construction of power plants. The analysis of the electricity shortage and the one on bills gives the cases.

The pressure on electricity is partly reversible: consumption can fall or move within a few years. What it triggers is less so. A gas plant built to meet data centre demand keeps emitting for as long as it operates, and its CO2 adds to the first risk.

The world's data centres consumed about 560 billion litres of water in 2023, according to the IEA, counting the water used for electricity and chip manufacturing. Globally, this is small next to agriculture. The risk lies in location: in the United States, 2 thirds of the centres built or planned since 2022 are in water-stressed areas, according to the World Resources Institute, which draws on a 2025 Bloomberg analysis. In GreenIT's weighted analysis, the risk of water deprivation accounts for only 1% of the global footprint of AI, because it is averaged across the planet. For a given basin, it can be the first. The analysis of data centre water use examines 6 cases.

Land is the least measured pressure. There is no national figure for the area occupied by data centres, in France or in the United States; Arcep notes in 2026 that the literature remains sparse. The largest US AI campuses cover 0.4 to 11 km², according to Epoch AI's satellite surveys. The risk is local and lasting for land that has been built over; it is the origin of several challenges to projects in France.

Data centres will consume little metal on a global scale: in 2030, according to the IEA, up to 2% of current copper demand, more than 3% of that of rare earths, and 11% of that of gallium. Gallium shows the risk: China refines 98% of it. Electronic waste from generative AI would reach 1.2 to 5 million tonnes cumulatively from 2020 to 2030, according to Wang and co-authors, less than a tenth of what the world produced in the single year 2022, 62 million tonnes, of which 22.3% was collected and recycled in a documented way. The metal in waste that leaves these channels is hard to recover, which is what makes this pressure lasting despite its small volume.

Hardware

Electronic waste and AI

Waste electrical and electronic equipment, in million tonnes · 2020 to 2030

Source: ITU and UNITAR, 2024; Wang et al., 2024 · projetcelsius.com/en/ai · CC BY 4.0

PNGCSV

For us at Projet Celsius, the most serious environmental risk of AI is the CO2 from the electricity that powers it, and first of all that of the gas and coal plants that meet its demand. It is the only direct effect that carries the most weight in the weighted footprint, acts across the whole planet and lasts for centuries. Water and land matter just as much to the inhabitants of a basin or a municipality, and they are judged site by site. For metals, the risk lies in supply and recycling more than in volumes. Beyond these direct pressures, the indirect effects of AI on the economy may weigh more: they are the subject of the analysis on the climate crisis.