14 SEP 2026 · 09:25 · INFRASTRUCTURE
Why AI Uses Energy and Water: The Environmental Cost Explained
AI’s environmental cost comes from physical infrastructure: computation uses electricity, equipment produces heat, cooling can consume water, and hardware has a manufacturing footprint.
AI is not automatically bad for the environment, but using it is not immaterial. Training and running models happens on servers in data centres. Those servers draw electricity, cooling systems remove their heat, and specialised chips must be manufactured and replaced. The impact depends on the model, the task, the hardware, the data centre, its location and the electricity supplying it.
The short answer
This is why a single universal figure for “one AI query” is usually misleading. Text generation, image creation, video generation and model training have different workloads. Providers use different hardware and batching strategies, while power grids differ by hour and region. The responsible question is not whether AI has a footprint—it does—but which use produces which benefit at which resource cost.
Electricity is the largest visible input
Servers carry out large numbers of numerical operations when a model is trained and again whenever it is used. The International Energy Agency reported that data centres accounted for about 1.5% of global electricity consumption in 2024. That total includes much more than AI, but AI is an important source of current growth.
The same IEA analysis emphasises concentration. A data centre can be a modest share of global demand while still placing a large new load on a particular grid connection or local power market. Environmental consequences therefore depend on both total consumption and where new demand arrives.
Training and inference are different
Training adjusts a model’s weights over a large body of examples and can require sustained computation. Inference uses the trained weights to answer requests. One inference may be much smaller than a training run, but a popular service may perform an enormous number of inferences. Efficiency per request can improve while total demand rises because usage grows or applications become more computationally intensive.
Model size is only one factor. The length of the input and output, response latency targets, numerical precision, hardware utilisation, caching and whether requests are batched can all alter energy per useful result. Comparing systems requires a fixed workload and a clearly stated boundary.
Why data centres use water
Computers turn much of their electricity into heat. Some facilities use evaporative cooling, which can consume water on site. There is also indirect water use associated with generating electricity. Lawrence Berkeley National Laboratory distinguishes direct site water from water consumed in the electricity supply chain and reports water-use effectiveness in litres per kilowatt-hour of IT load.
Cooling choices involve trade-offs. A design that reduces direct water consumption may use more electricity, while an efficient evaporative system may place pressure on a water-stressed region. Climate, time of day, water source and local scarcity matter more than a global average alone.
Hardware has an upstream footprint
The operational footprint is only part of the lifecycle. Semiconductor fabrication, server manufacturing, construction and the supply of metals and chemicals require energy and materials. Replacing accelerators quickly can increase embodied impacts and electronic waste. A full assessment should therefore include both operating resources and the hardware lifecycle rather than counting electricity at the wall only.
What the strongest public data says
The 2024 U.S. Data Center Energy Usage Report from Lawrence Berkeley National Laboratory estimates that electricity used by GPU-accelerated AI servers rose from less than 2 terawatt-hours in 2017 to more than 40 terawatt-hours in 2023. Its scenarios vary widely because future shipments and operating patterns are uncertain. That range is itself a useful finding: projections are sensitive to adoption and efficiency, not fixed destiny.
The report also models direct water use and explains why power-usage and water-usage effectiveness must be interpreted together. It is evidence for rapid infrastructure growth, not a licence to assign the whole data-centre footprint to every AI application.
How to judge whether an AI use is worth it
Start with the outcome. If a smaller model, search system, rules engine or conventional software can do the job, compare it. Measure the complete task—including retries and discarded outputs—on representative inputs. Ask providers for energy, water and hardware reporting with defined boundaries, locations and time periods.
Then reduce avoidable work: shorten prompts where that does not harm quality, cap unnecessary output, cache stable results, batch requests when latency permits, route easy cases to smaller systems and retire experiments that never reach users. Environmental evaluation improves when teams count useful results, not merely model calls.
AI can also support energy optimisation, scientific work and other beneficial applications. Those benefits do not erase resource use, and resource use does not erase benefits. The sound conclusion is conditional: disclose the footprint, compare alternatives and spend computation where the value is defensible.
Sources and further reading
- International Energy Agency — Energy and AI
- Lawrence Berkeley National Laboratory — 2024 United States Data Center Energy Usage Report
- U.S. Department of Energy — Cooling water efficiency opportunities for data centers