How Does AI Affect the Environment?

AI affects the environment mainly through the electricity used to train and run AI models, the carbon emissions from generating that electricity, the water used to cool data centers, and the materials and waste involved in making chips and servers. The size of the impact depends on how much AI is used, how efficient the systems are and how clean the local power grid is. AI can also help cut emissions in some areas.

The main ways AI affects the environment

AI runs on physical infrastructure: specialized chips in servers, housed in data centers and connected to power grids. Almost all of AI's environmental footprint comes from that infrastructure.

• Electricity: training and running AI models takes large amounts of power.
• Carbon emissions: much of that electricity still comes from fossil fuels.
• Water: data centers use water for cooling, and many power plants use water to generate electricity.
• Hardware: making chips and servers requires energy, minerals and manufacturing, and old equipment becomes electronic waste.

This article focuses on electricity, emissions and hardware, with a short section on water, which has its own detailed article.

Electricity: how much energy AI uses

AI models run on specialized chips, most commonly GPUs, that perform enormous numbers of calculations in parallel. Servers built around these chips draw far more power than ordinary servers, and they run in data centers that also need power for cooling and other equipment.

Data center electricity use

The International Energy Agency (IEA) estimated in April 2025 that data centers used about 415 terawatt-hours of electricity in 2024, around 1.5% of the world's electricity consumption. That figure covers all data centers, including those running email, streaming and cloud storage, not just AI.

AI is the fastest-growing part. In its April 2026 update, the IEA reported that global data center electricity demand grew 17% in 2025, while consumption at AI-focused data centers grew 50%. It projects data center consumption roughly doubling from 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030, about 3% of global electricity demand.

The effect is more concentrated in some countries. Lawrence Berkeley National Laboratory estimated that U.S. data centers used about 176 terawatt-hours in 2023, or 4.4% of all U.S. electricity. Its scenarios project 325 to 580 terawatt-hours by 2028, roughly 6.7% to 12% of forecast U.S. electricity use. The lab attributed much of the recent growth to AI servers.

Training vs inference

AI uses electricity at two stages.

Training is building a model by processing huge datasets, often on thousands of chips for weeks or months. Researchers from Google and UC Berkeley estimated in 2021 that training GPT-3, a model released in 2020, used about 1,287 megawatt-hours of electricity and produced about 552 tonnes of carbon dioxide equivalent. Developers of today's larger models generally do not publish comparable figures.

Inference is using a trained model, every time someone asks a question or generates an image. Each request uses far less energy than training, but popular products handle enormous volumes of requests. Research from Google and Meta published in 2021 found that running models for users often uses more energy than training them. Meta researchers reported that Facebook's AI infrastructure power capacity was split roughly 10% experimentation, 20% training and 70% inference.

How much energy does one AI prompt use?

There is no single figure, because it depends on the model, the task, the hardware and what the estimate includes. Two widely cited company figures:

• Google estimated in August 2025 that its median Gemini Apps text prompt used 0.24 watt-hours of energy in May 2025. Google said energy per median prompt had fallen 33 times over the previous 12 months. The figure covers text prompts only and was not independently verified.
• OpenAI CEO Sam Altman wrote in June 2025 that an average ChatGPT query uses about 0.34 watt-hours. No methodology was published.

Heavier tasks use more. Generating images or video, or asking a model to reason step by step, generally takes more computing than a short text reply. Mistral AI reported in 2025 that a model 10 times larger has roughly 10 times the environmental impact for the same amount of generated text.

Carbon emissions: it depends on the grid

AI's carbon footprint depends mostly on how its electricity is generated. The same AI task has very different emissions in a region powered by coal than in one powered by hydro, wind, solar or nuclear.

Globally, fossil fuels still supply a large share. The IEA estimated in 2025 that coal provides about 30% of the electricity used by data centers worldwide, renewables about 27%, natural gas 26% and nuclear 15%. It expects natural gas and coal together to meet over 40% of the additional electricity demand from data centers through 2030.

The IEA estimated that data centers were responsible for about 180 million tonnes of carbon dioxide in 2024 from their electricity use. In its 2026 update, it projected data center emissions roughly doubling to about 350 million tonnes by 2035, which would still be about 2% of global electricity-sector emissions.

Company reports show the pressure AI is putting on climate goals:

• Google reported that its operational emissions fell 2% in 2025 even as its electricity demand rose 37%, but its supply chain emissions grew 25%. It attributed the rise partly to the scale of new AI infrastructure.
• Microsoft reported that its total emissions rose 25% year over year in its 2025 fiscal year, driven primarily by expanding its data centers and by pausing its use of certain renewable energy certificates.

How companies count emissions matters. Many report "market-based" figures that subtract clean energy they purchase, which can differ substantially from the emissions of the local grid their data centers actually draw from.

Water: a local issue

Data centers often use water to cool their equipment, commonly by evaporating it in cooling towers, and power plants use water to generate electricity. Google, for example, reported consuming 10.9 billion gallons of water across its data centers and offices in 2025, up 34% from 2024.

The impact depends heavily on location: the same volume matters more in a drought-prone region than in a water-rich one. For the full explanation, including why per-prompt water estimates vary so widely, see Why Does AI Use Water?.

Hardware: chips, minerals and electronic waste

Manufacturing emissions. A 2021 study by Harvard and Facebook researchers found that most emissions associated with modern mobile and data center equipment come from hardware manufacturing and infrastructure rather than from operating it. The balance depends on how clean the electricity used for operation is. For AI specifically, researchers estimated that training the BLOOM language model emitted about 24.7 tonnes of carbon dioxide equivalent counting only the power drawn during computation, and about 50.5 tonnes once all processes, from equipment manufacturing to the energy used to run it, were included.

Critical minerals. Chips rely on specialized materials. Gallium, a metal used in advanced chips, is almost entirely refined in China, which accounts for around 99% of global refined gallium supply. The IEA flags this as a supply risk and projects that data centers could need more than 10% of today's gallium supply by 2030. The UN Environment Programme notes that the chips powering AI depend on rare earth elements, which are often mined in environmentally destructive ways.

Electronic waste. AI servers are replaced as newer, faster chips arrive. A 2024 study in Nature Computational Science estimated that generative AI could generate a total of 1.2 to 5 million tonnes of electronic waste over 2020 to 2030, depending on how the technology develops. The authors found that circular economy strategies, such as reusing and recycling equipment, could cut that by 16% to 86%. For context, the world produced a record 62 million tonnes of e-waste from all electronics in 2022, according to the UN's Global E-waste Monitor, and less than a quarter was documented as properly collected and recycled.

Is AI bad for the environment?

AI has a real and growing environmental footprint, driven mostly by the electricity and hardware behind data centers. Whether its overall effect is harmful depends on how that electricity is generated, how efficiently AI systems are built and used, and what AI is used for.

Several things are true at once:

The footprint is growing quickly. Electricity use at AI-focused data centers grew 50% in 2025, and major technology companies report rising electricity use and emissions linked to AI infrastructure.

It is a small share of global electricity, but can be a large share locally. Where data centers cluster, their demand can put significant pressure on local grids and water supplies.

Efficiency is improving fast. The IEA reported in 2026 that energy use per AI task has dropped by at least an order of magnitude annually in recent years. Total demand keeps rising anyway, because usage is growing even faster.

AI can also reduce emissions. The IEA says AI can improve forecasting for wind and solar power and could free up to 175 gigawatts of additional transmission capacity on existing power lines. Europe's weather forecasting center, ECMWF, reported that its AI system, which began running alongside its traditional physics-based model in February 2025, uses about 1,000 times less energy to make a forecast. The IEA found that widespread use of existing AI applications could cut emissions by far more than data centers emit, but far less than is needed to address climate change, and that such adoption is not guaranteed.

Why the numbers are uncertain

• Limited disclosure. Lawrence Berkeley National Laboratory has said that a lack of available data significantly limits its analysis of data center energy use. Many AI developers do not publish energy or emissions figures for training their largest models.
• Different methods. Some estimates count only electricity at the chip; others include cooling, idle equipment, manufacturing or power plant water. Emissions can be counted using the local grid or after subtracting purchased clean energy.
• Unverified company figures. Google notes that its per-prompt figures have not been verified by an independent third party.
• An uncertain future. The IEA says there is substantial uncertainty about data center consumption today and even more about the future, depending on how quickly AI is adopted and how fast efficiency improves.

When you see a striking number about AI and the environment, check what it covers, what year it refers to and who produced it.

Why this matters

AI's environmental impact affects where data centers are built, how electricity grids are planned, what local communities are asked to accept and how companies meet their climate commitments.

For individuals and organizations using AI, useful questions include:

• Does this task need AI, or a large model? Smaller models and simpler tools use less computing.
• What does the provider disclose? Look for published energy, emissions and water figures, and whether they are independently verified.
• Where does the computing happen? The local power grid and water situation shape the real impact.

For how AI systems are trained and used, and why that takes so much computing, see How Does AI Work?.