What Does AI Actually Cost the Environment?
The impact is real. The simple answers usually aren't.
AI can feel almost weightless.
You type a question. Words appear. There is no smoke, no engine, no visible pile of raw materials sitting beside your laptop.
But AI is not weightless.
Behind the screen are data centers, computer chips, cooling systems, electricity grids, water systems, manufacturing plants, mines, transportation networks, and eventually electronic waste.
All of that has an environmental cost.
The difficult part is that there is no single number that tells us what that cost is.
There is no single environmental cost of using AI
One reason this topic gets confusing so quickly is that AI and data centers are not the same thing.
Data centers support AI, but they also support cloud storage, streaming, websites, financial systems, business software, and much of the rest of the modern internet.
So when someone says, “AI uses this much electricity,” it matters whether the number actually measures AI or whether it measures data centers more broadly.
The International Energy Agency projects that global data-center electricity demand could rise from about 485 terawatt-hours in 2025 to roughly 950 terawatt-hours by 2030 — around 3% of global electricity demand.
AI-focused data centers are expected to grow even faster, with their electricity use roughly tripling over that period.
That is significant growth.
In the United States, Berkeley Lab researchers estimate that data centers could account for about 11.8% of national electricity use by 2030 in their reference case.
But even that number is not presented as a certainty. Their modeled range runs from 9.5% to 15.3%.
That range is not a flaw in the research.
It is a reminder that the future depends on things that are still changing: how quickly AI use grows, how efficient hardware becomes, what kinds of AI systems people use, where new data centers are built, and how those facilities are powered and cooled.
One prompt isn't a very useful unit
A lot of environmental claims about AI get reduced to something like:
One AI prompt uses X amount of electricity.
That sounds wonderfully simple.
It usually isn't.
Google measured the environmental impact of its own Gemini Apps service in production and reported that, in May 2025, a median text prompt used about 0.24 watt-hours of electricity.
That is a useful measurement.
It is not a universal conversion factor for “an AI prompt.”
It measures one company's service, for one type of interaction, using its infrastructure, at one point in time.
Different AI tasks can require dramatically different amounts of computing.
The International Energy Agency reports that energy use for basic AI tasks has been falling rapidly as systems become more efficient. But more demanding uses — including reasoning systems, AI agents, and video generation — can require hundreds or even thousands of times more energy per task than a simple text interaction.
So asking, “How much energy does one AI prompt use?” is a little like asking, “How much fuel does one trip use?”
A drive to the grocery store and a cross-country freight route are both trips.
That does not make them equivalent.
Water is even more local
Water-use claims are another place where a simple viral number can become misleading very quickly.
Data centers can use water directly for cooling.
They can also have an indirect water footprint because producing electricity can require water.
And the hardware itself carries additional water and resource demands through manufacturing and the broader supply chain.
But those impacts vary enormously depending on where and how the computing happens.
A 2025 Berkeley Lab study found that water use across data-center workloads can vary by more than 10,000 times.
That variation comes from differences in server efficiency, electricity sources, cooling systems, utilization, infrastructure efficiency, climate, hardware age, and other factors.
This is why a claim that “one AI prompt uses a bottle of water” should not be treated as a universal rule.
Location matters.
A gallon of water used in a water-rich region does not create the same local pressure as a gallon used in a drought-prone area.
The same computing workload can have very different environmental consequences depending on the local grid, cooling design, climate, and watershed.
Global averages can help us understand scale.
They cannot tell us everything about local impact.
The footprint starts before the server turns on
Electricity and water are only part of the story.
NIST's AI Risk Management Framework Playbook specifically identifies environmental effects from:
energy consumption,
water consumption,
greenhouse-gas emissions,
mining and extraction of raw materials,
production of computing equipment and networks,
hardware transportation,
and electronic-waste recycling or disposal.
The OECD makes a similar distinction between the direct environmental impact of AI infrastructure itself and the indirect effects created by what AI causes people and organizations to do.
That matters because a system can become more efficient without necessarily reducing total resource use.
If an AI task becomes dramatically cheaper and more efficient, people may simply use much more of it.
That is known as a rebound effect: efficiency improves, but increased use absorbs some or all of the savings.
So efficiency matters.
But efficiency alone does not answer the environmental question.
Scale matters too.
AI can also reduce environmental impact
There is another side to the equation.
AI is already being explored and used to improve electricity grids, building efficiency, industrial systems, transportation, forecasting, and other resource-intensive operations.
The International Energy Agency has identified AI applications that could produce substantial energy savings if they are deployed successfully and widely.
But potential savings are not the same thing as measured net benefit.
An AI system that identifies a more efficient way to operate a building only creates environmental value if the recommendation is actually implemented and the resulting savings outweigh the resources required to build, run, and maintain the system.
The same standard should apply to environmental claims that we apply to other AI claims:
What changed in the real world?
Not what the technology theoretically could do.
Not what the vendor says it enables.
What actually changed?
So what does responsible AI use look like?
For an individual user, the most useful conclusion probably is not that every ordinary question should produce environmental guilt.
One text prompt is rarely the right level at which to understand the problem.
For organizations building, buying, or deploying AI at scale, the questions are much larger:
How much computing is actually required?
How efficient is the infrastructure?
Where does the electricity come from?
How is the facility cooled?
What is happening with local water resources?
How often is hardware being replaced?
What happens to that equipment afterward?
Are efficiency gains reducing total resource use, or simply making much more use possible?
What effect does the supporting infrastructure have on the communities where it is built?
And if AI is being justified partly because it will save energy or resources somewhere else, are those savings actually being measured?
Those questions are harder than turning every prompt into a single scary number.
They are also much more useful.
The bottom line
AI has a real environmental footprint, and that footprint is growing as AI infrastructure expands.
But there is no honest universal number for the environmental cost of “using AI.”
The impact depends on the system, the task, the hardware, the data center, the electricity source, the cooling method, the location, the lifecycle of the equipment, and the scale at which the technology is used.
Simple numbers travel well.
Environmental systems are not simple.
Understanding AI's impact means measuring the right things, being clear about where the numbers came from, and resisting the temptation to turn a complicated physical system into one frightening statistic.
Sources and further reading
International Energy Agency (2026)
Key Questions on Energy and AI
Lawrence Berkeley National Laboratory (2026)
United States Data Center Energy Usage Report: 2025 Update
Lei, N., Lu, H., Shehabi, A., & Masanet, E. (2025)
The Water Use of Data Center Workloads: A Review and Assessment of Key Determinants
National Institute of Standards and Technology
AI Risk Management Framework Playbook — MEASURE 2.12: Environmental Impact and Sustainability
Organisation for Economic Co-operation and Development (2022)
Measuring the Environmental Impacts of Artificial Intelligence Compute and Applications: The AI Footprint
Google (2025)
Measuring the Environmental Impact of Delivering AI at Google Scale
Last reviewed: August 25, 2026
