Ask an AI system a question and the answer appears on a screen. The work that produced it happened elsewhere: in a building full of chips, cables, fans and cooling equipment, connected to an electricity grid. The physical footprint is real. Measuring it honestly is harder than attaching a dramatic number to every prompt.
Follow the computation
Modern AI relies heavily on specialised accelerators designed to carry out many mathematical operations in parallel. Training a large model uses clusters of these chips for an extended period; using the finished model, known as inference, also takes computation every time it serves a request. Servers, memory, networking equipment and storage all add to the load. Cooling and power-conversion systems consume electricity beyond what reaches the processors.¹
A data centre is not necessarily an AI centre. It might host banking software, streaming video or ordinary cloud services alongside AI. That matters when interpreting industry-wide figures. The International Energy Agency estimated that data centres used about 415 terawatt-hours of electricity worldwide in 2024, around 1.5% of global consumption, and projected roughly 945 terawatt-hours in its base case for 2030. AI is a major driver of the expected rise, but those totals cover all data-centre activity.¹
The numbers also describe different things. A watt is power at a moment; a watt-hour is energy over time. A facility demanding a large amount of power can strain a local grid even while data centres remain a modest share of worldwide electricity. Where and when demand appears are therefore as important to planners as the global percentage.
Heat has to go somewhere
Nearly all the electrical energy used by computing equipment eventually becomes heat. Air cooling uses fans and mechanical refrigeration; some installations circulate liquid close to the chips. Cooling choices depend on equipment density, climate, water availability and the price and carbon intensity of electricity.²
Water enters the story in several ways. Evaporative cooling can consume water on site. Power stations may also withdraw or consume water to generate the electricity a centre uses. Manufacturing chips and constructing facilities carry further upstream demands. A claim about a data centre's water use is incomplete unless it says which boundary it measures: direct on-site consumption, electricity-related use, or the wider supply chain. The US Lawrence Berkeley National Laboratory reports electricity and water estimates with these distinctions rather than treating a single number as universal.²
It is tempting to ask how much water “one AI question” uses. There is no stable answer across all models and locations. A short response from a small model on efficient hardware differs from a long output from a larger one; cooling and electricity sources differ too. Even dividing a facility's annual water consumption by its annual number of requests hides idle capacity and shared workloads. Researchers and operators can estimate per-request impacts under specified conditions, but a viral figure without those conditions should be treated cautiously.³
Carbon depends on the power system
The same computing workload can produce different emissions depending on the electricity supplying it. An additional megawatt-hour served by fossil generation has a different consequence from one served by low-carbon power. Annual renewable-energy purchases may reduce a company's accounting footprint, but they do not necessarily mean a particular data centre ran on carbon-free electricity every hour.¹
There is also an embodied footprint. Chip fabrication, building materials, server replacement and transport all require energy and resources before a machine is switched on. These costs are easy to miss when discussion focuses only on operational electricity. Conversely, an estimate that attributes all data-centre growth to AI overstates AI's share. Good accounting should identify the workload, the measurement period and what is included.²
Corporate reporting offers useful evidence, but it is not a universal ruler. Google's environmental material, for example, reports its own data-centre efficiency and water programmes. Those disclosures describe one fleet, using that company's definitions. They cannot be multiplied across every provider, nor do they reveal the footprint of every model running in a shared facility.⁴
Manufacturing the hardware introduces another measurement boundary. More powerful accelerators require fabrication plants, materials and supply chains, and frequent replacement can shift impacts from electricity use to production. A model that appears efficient during operation could still carry a sizeable embodied cost if it runs on short-lived equipment. Conversely, using existing machines more fully may spread that cost over more work. Sector reports are strongest on facility electricity and water, while product-level comparisons remain less standardised.²
Water and power choices can also pull in different directions. Evaporative cooling may reduce electricity needed for mechanical refrigeration while consuming more local water. A design that looks efficient by one measure may be unsuitable in a dry region. That is why both the local resource context and the accounting boundary belong beside any headline estimate.²
Efficiency can grow alongside demand
A more efficient chip can perform a task with less energy. Better software can reduce unnecessary calculations; a smaller model may be sufficient for a simple job. Cooling design and the siting of facilities can also lower particular impacts. Yet if cheaper AI leads to far more usage, total demand can still rise. The IEA's growth projection illustrates why efficiency gains alone cannot settle the aggregate question.¹
Consider two separate decisions: how much energy a given task needs, and how many tasks society chooses to run. The first is a technical problem. The second depends on products, prices and policy. Confusing them produces both exaggerated panic and unwarranted reassurance.
The clearest environmental questions are therefore practical ones: how much electricity and water does a facility use, at what times, in which place, with which generation mix, and for which workloads? We have credible estimates of a rapidly growing data-centre sector. We have fewer standardised public measurements that isolate AI precisely. Better disclosure would make the physical costs easier to compare, and easier to manage.¹,²