- Data-center investment reached USD580bn in 2025, putting AI on track to become one of the world's fastest-growing sources of electricity demand. Installed capacity is expected to double by 2030, with AI workloads already accounting for 15–20% of data-center electricity use and potentially approaching 40% by the end of the decade. Yet the sector's environmental footprint remains underestimated as most analyses focus only on operational electricity use. This analysis takes a broader systems view across 26 countries (+93% of global capacity), adding lifecycle emissions, water use and AI's growing resource demand.
- Identical workloads can generate up to 24 times more emissions depending on the emission intensity of the grid, making location as decisive as demand growth. Fossil-dependent grids in Indonesia, India and Malaysia exceed 600 gCO₂/kWh, compared with under 30 gCO₂/kWh in Norway and Sweden. The US and China, which host the largest data-center clusters, sit in between at 384 gCO₂/kWh and 526 gCO₂/kWh, respectively, giving Europe's cleaner power mix a structural advantage for low-carbon AI growth. These disparities are amplified by transmission and distribution losses of 10–15% in some markets, while less reliable grids raise electricity needs and dependence on backup generation.
- At 286 MtCO₂ in 2025, the true carbon footprint of data centers is 57% larger than IEA estimates suggest. Electricity consumption (Scope 2) accounted for 76% of this footprint, at 218 MtCO₂, with hardware manufacturing and construction (Scope 3) contributing a further 66 MtCO₂, or 23%, and direct Scope 1 emissions remaining negligible (<1%). Emissions are also heavily concentrated geographically, with the US and China alone accounting for roughly 70% of the global total. AI accounts for an estimated 43-60 MtCO₂ of today's emissions, and this is set to climb steeply as deployment widens and computing demand grows.
- Without grid decarbonization, global data-center emissions would more than double to 643 MtCO₂ by 2030, leading to an estimated USD154bn in annual climate damages (up from USD68bn today). AI workloads already account for an estimated USD13bn in climate damages annually and could exceed USD50bn by 2030. By contrast, an ambitious decarbonization pathway would hold emissions to around 329 MtCO₂ despite continued growth in computing demand, and keep climate damages at USD79bn. This makes the pace of power-sector decarbonization the primary determinant of whether AI growth can be decoupled from emissions in the near term. Even under ambitious decarbonization, however, the footprint does not vanish but moves up the supply chain: as Scope 2 falls from more than 70% of the footprint today to around half by 2030, embodied emissions from servers, semiconductors and infrastructure become the binding constraint, approaching 50% of the total. Achieving genuinely low-carbon AI will therefore require not only cleaner power, but also closer attention to emissions embedded in digital infrastructure supply chains.
- Deployed across the economy, AI could cut global CO₂ emissions by around 1.4 Gt a year by 2035, more than offsetting the emissions generated by its own infrastructure and creating net savings of roughly 750 MtCO₂. According to the IEA, these reductions would result from efficiency gains, optimization and improved resource management across sectors such as energy, industry, buildings and transport, and are equivalent to around 2.6% of current global emissions. However, this outcome is not guaranteed. With most AI applications still at an early stage of deployment, its ultimate climate impact will depend on whether these economy-wide benefits can scale faster than the infrastructure required to support them.
- Data centers consumed 814bn liters of water in 2025 and could require 1.3–1.8trn liters by 2030, comparable to Switzerland's annual consumption, making water the overlooked resource constraint of AI. Most of this footprint is indirect, with roughly three-quarters originating from electricity generation and the remainder from on-site cooling and semiconductor manufacturing. This ties water use closely to the energy transition, since fossil and nuclear plants require substantial cooling water while wind and solar use little or none in operation, lowering both the carbon and water footprints of a cleaner grid. Although power-sector decarbonization can help moderate future water demand, water-related risks are becoming increasingly concentrated in water-stressed regions such as South Korea, India, Mexico and parts of China, where rapid data-center growth is colliding with existing pressure on local water resources, raising the risk of access constraints and community or regulatory opposition to new capacity.
- Realizing "green AI" will depend less on making data centers marginally more efficient than on transforming the energy systems that power them. Unlocking AI's environmental potential will require a broader policy framework, combining clean-power expansion, greater transparency on resource use, stronger incentives to price environmental costs, and faster deployment of AI applications that reduce emissions across the wider economy.