Measure the whole lifecycle
AI’s footprint begins before a prompt and continues after a server is retired. Semiconductor fabrication uses energy, ultrapure water, chemicals, and minerals. Data centers use electricity, cooling, backup generation, networking, land, and equipment. End-of-life hardware becomes an e-waste and recovery problem.
Public debate often quotes a per-query water or energy number as if it were universal. Actual use varies by model, hardware, data-center utilization, cooling, climate, grid, prompt length, output modality, and allocation method. Training and inference are distinct; inference can dominate over a product’s lifetime if use is large. Providers disclose too little for precise global AI-only totals.
What is measured today
The International Energy Agency estimated that all data centers used about 415 terawatt-hours of electricity in 2024, around 1.5% of global consumption. AI is a subset of that total. Nearly half of U.S. data-center capacity was concentrated in five regional clusters, so local grid effects can be much larger than the global share (IEA, April 2025).
The IEA’s updated central projection puts data-center use at roughly 950 TWh in 2030, about 3% of global demand. That is a forecast based on capacity, efficiency, adoption, and market assumptions, not a meter reading from the future. More energy-intensive reasoning, video, and agentic use could push demand upward; efficiency, financing constraints, or slower adoption could reduce it (IEA, 2026).
In the United States, Lawrence Berkeley National Laboratory estimated data-center electricity use rose to 176 TWh in 2023 and could reach 325–580 TWh in 2028 (U.S. Department of Energy, December 2024). The range illustrates uncertainty. It covers data centers generally, not only generative AI.
Emissions depend on where and when power is produced
An efficient server on a fossil-heavy grid may create more operational emissions than a less efficient server supplied by low-carbon power. Annual renewable-energy contracts do not always show the electricity serving a facility at the hour of demand. Marginal demand may keep fossil generation online even when a company purchases renewable certificates.
IEA estimated electricity-related data-center emissions at about 180 million tonnes of carbon dioxide in 2024, around 0.5% of global fuel-combustion emissions. Its base case rises toward 1% by 2030 (IEA climate analysis). Small global shares can still conflict with local climate plans or create rapid sectoral growth.
Credible reporting should include location-based and market-based emissions, hourly matching where feasible, backup generators, construction, and embodied hardware emissions. Avoided-emissions claims from AI optimization should be reported separately and validated against a counterfactual.
Water is local
Data centers may consume water directly through evaporative cooling and indirectly through electricity generation. Semiconductor fabrication also requires high-quality water. Withdrawal is not the same as consumption: withdrawn water may be returned, while consumed water is not immediately available to the local watershed.
A gallon used in a water-stressed basin has different impact from one used where supply is abundant. Annual global totals can hide seasonal competition with households, ecosystems, agriculture, and industry. Air cooling can reduce direct water use but increase electricity use; reclaimed water can reduce potable demand but requires infrastructure and safeguards.
The OECD notes that environmental assessment should extend beyond energy and carbon to freshwater and mineral extraction, and that data-center water reporting remains incomplete (OECD AI compute). Communities need facility-level withdrawal, consumption, source, seasonal stress, discharge, and emergency plans before permits are granted.
Chips have an upstream footprint
Advanced accelerators rely on geographically concentrated fabrication, packaging, memory, and materials. Manufacturing plants have large construction and utility requirements. Short model-upgrade cycles can retire still-functional equipment, although older accelerators may be reused for less demanding inference.
Mining and processing can produce habitat damage, pollution, occupational hazards, and geopolitical dependency. Exact attribution to “AI” is difficult because the same supply chain serves phones, vehicles, networking, and other computing. Lifecycle analysis should allocate impact transparently rather than assigning all semiconductor growth to AI.
UNEP identifies energy, greenhouse-gas emissions, water, mineral extraction, and e-waste as direct lifecycle concerns and calls for common measurement methods (UNEP, September 2024). UNEP has an environmental mandate, which shapes its focus; its measurement recommendation remains broadly useful.
AI can produce environmental benefits
Machine learning can improve weather and climate modeling, detect methane, optimize grids and buildings, forecast renewable output, design materials, and support conservation. Benefits must be measured in deployment, not inferred from benchmark accuracy.
Rebound effects matter. If efficiency lowers cost, total use can increase enough to erase savings. Optimization may also shift harm: reducing electricity cost while increasing water use or siting infrastructure in a vulnerable community. Net assessment should include induced demand and distribution.
The IEA projects renewables meeting roughly half the growth in data-center electricity demand to 2035 in its base case, but natural gas and coal also supply substantial near-term growth. “Powered by renewables” is therefore not a description of the global sector.
Better infrastructure choices
Operators can improve utilization, use efficient models and hardware, schedule flexible workloads when cleaner power is available, extend equipment life, recover heat where useful, reduce idle capacity, and select cooling appropriate to the local watershed. Smaller task-specific models can outperform wasteful use of a frontier model.
Utilities and regulators can require realistic interconnection studies, cost allocation, demand flexibility, clean supply, public water review, and decommissioning plans. Data-center customers should not silently shift grid upgrades or scarcity costs to households.
Disclosure should use common functional units: energy and water per training run, per standardized workload, and per service volume, alongside absolute company and facility totals. Efficiency ratios without absolute growth can mislead.
What AGI could change
AGI energy demand is a scenario. Systems that automate research and economic activity could sharply increase compute use. They might also improve chip design, materials, grids, and clean-energy deployment. Jevons-style rebound could turn efficiency gains into more total consumption.
Physical constraints remain: fabs, transformers, generation, grids, water systems, and permits take time. Claims of immediate intelligence abundance should account for these bottlenecks. Environmental governance may become a practical limit on deployment rather than an external issue.
A balanced standard
Ask four questions: What is the absolute lifecycle footprint? Where and when does it occur? Who receives the benefit and bears the cost? What alternative delivers the same function?
The evidence supports neither “AI is environmentally negligible” nor a universal per-prompt catastrophe. Data centers are a minority of global electricity use but a fast-growing, geographically concentrated load with material water and supply-chain impacts. AI can enable environmental improvements, but only measured net outcomes count. Mandatory facility disclosure, lifecycle standards, local participation, clean-grid investment, and efficient model choice are the foundations of an honest environmental policy.