How Bad Is AI for the Environment? 2026 Energy, Water & Carbon Data

A single AI prompt uses surprisingly little electricity. The infrastructure behind billions of prompts is another story. Google estimates that a median Gemini text prompt uses 0.24 watt-hours, about the energy of watching TV for nine seconds. The problem is volume. Data centers draw enormous amounts of electricity and water, and the chips inside them eventually become waste.

This article covers what the 2026 data says about energy, carbon, water, hardware, land, and the grid. It also explains what each statistic measures, because many of the most-quoted numbers describe something other than AI.

Most global figures in this article describe all data centers, not AI alone. AI drives a large share of new demand, so these figures show the pressure on infrastructure, but they do not measure AI’s standalone footprint. This article labels each figure with its scope.

Is AI bad for the environment?

AI adds real, fast-growing demand for electricity, water, and hardware. It can also cut energy use in other industries. The net result depends on how operators power data centers, where they build them, and whether heavier use absorbs the efficiency gains.

Right now, the costs are easier to measure than the savings.

What does each AI environmental number actually measure?

Check the boundary before you compare two figures. A statistic may describe:

• one prompt or one model
• one company’s data centers
• all data centers worldwide, including non-AI workloads
• direct cooling water or the full water footprint, including power generation
• operating emissions or lifecycle emissions
• a measured result or a modeled 2030 scenario

That same problem appears across AI reporting: a number can look precise while hiding the definition, timeframe, or denominator behind it. The broader AI facts and statistics in 2026 show why apparently contradictory AI numbers can still be measuring different things.

How much electricity does AI use?

Global data centers, not AI alone, used about 485 TWh of electricity in 2025, according to the International Energy Agency (IEA). The IEA projects roughly 950 TWh by 2030, about 3% of global electricity demand. The United Nations University (UNU) estimates a similar path: 448 TWh in 2025 and 945 TWh in 2030.

The growth runs unevenly. Data center demand rose 17% in 2025, while demand from AI-focused data centers rose 50%. The IEA expects those facilities to triple their consumption by 2030.

Training vs. inference

That distinction matters because inference is tied directly to usage volume. A model can become more efficient per request while its total footprint continues rising if people use it far more often. The energy cost of AI inference is therefore a demand question as much as an efficiency question.

How much energy does one AI prompt use?

Google’s figure of 0.24 Wh, 0.03 grams of CO₂e, and about five drops of water covers only the median Gemini Apps text prompt. The methodology counts idle capacity, host CPUs, and data center overhead. Google reports that this median fell 33-fold over one year, and it notes that no independent party has verified the data.

The IEA agrees that simple text queries now use less electricity than running a TV for the same time. It also notes that video generation, reasoning, and agentic tasks can use hundreds or thousands of times more energy per query.

The surprising part is not that AI uses electricity. It is that efficiency and demand move in opposite directions. Per-prompt efficiency can improve while total demand rises, because cheaper AI invites more use. UNU calls this the rebound effect.

One more caveat: most data center estimates assume the work runs in the cloud. As more on-device AI runs on phones and laptops, some compute leaves data center servers, which makes today’s projections harder to read.

Does AI increase carbon emissions?

Yes, mainly through the electricity data centers draw. The IEA projects that data center electricity emissions (not AI-only emissions) will double to about 350 million tonnes by 2035, still roughly 2% of global electricity-sector emissions. UNU puts the carbon footprint of 2030 data center electricity at 399 million tonnes. Both figures cover all data centers.

Carbon and energy are different measures. Two data centers with identical consumption can emit very different amounts depending on the local grid. UNU also warns that low-carbon power does not automatically mean low-water or low-land power. For example, switching from coal to bioenergy can cut the carbon footprint of electricity by about 70%, while raising the water footprint more than thirtyfold.

Where will the electricity come from?

AI data centers need steady, round-the-clock power, which is why tech companies sign nuclear deals:

• Microsoft and Constellation: a 20-year agreement to restart Three Mile Island Unit 1 as the Crane Clean Energy Center, with 835 MW of capacity. Constellation targets a 2027 restart, and grid-connection timelines remain a risk.
• Amazon and Talen: a 1,920 MW supply agreement from the Susquehanna nuclear plant in Pennsylvania, running through 2042.

These contracts have limits. Buying output from an operating plant secures clean power for one buyer but does not by itself add new generation. It can, however, help keep an existing plant running that might otherwise face economic pressure to close. The IEA also reports that US developers pursue onsite natural gas generation, with roughly 15 to 27 GW possible by 2030. Nuclear is one part of the power mix, not the whole answer.

AI’s footprint includes the grid

Generating power is only half the problem. Data centers also need transmission lines, substations, and grid connections, and these take years to build. The IEA identifies electricity-system bottlenecks as a constraint on data center growth, and developers who wait for grid access turn to onsite generation. A second question is who pays. Regulators now decide whether grid upgrades cost data center operators or other electricity customers, which is why California’s new reporting rules (below) address it directly.

How much water does AI use?

AI’s water footprint comes from two places: cooling servers directly and generating the electricity that powers them. UNU projects that the water footprint of 2030 data center electricity will reach 9.3 trillion liters, equal to the basic domestic water needs of 1.3 billion people in Sub-Saharan Africa. The figure covers all data centers and projects forward.

Water statistics depend on what they count:

• Withdrawal vs. consumption. Withdrawn water may return to the source. Consumed water, such as evaporated cooling water, does not.
• Direct vs. indirect. Direct use happens on site. Indirect use happens at the power plant. That is why a headline about how much water AI uses needs more context than a single liters-per-prompt figure. Cooling technology, electricity generation, and geography can all change the result.
• Location. The same volume matters more in a drought region. UNU cites Querétaro, Mexico, and Uruguay as places where data center plans collided with water stress. The distinction becomes especially important when discussing the water consumption of AI data centers, because infrastructure-level estimates include much more than the water used directly to cool an individual request.

What is AI’s hardware footprint?

AI’s footprint starts before the first prompt. It begins with mining, then chip and server manufacturing, then transport and construction, and it ends with replacement and disposal.

Manufacturing emissions fall under Scope 3, the indirect emissions in a company’s supply chain, and per-prompt figures rarely include them. Google’s published method lists operating energy, idle capacity, overhead, and water, but not hardware manufacturing. Ask whether a number includes embodied emissions before you compare it.

On e-waste, a study in Nature Computational Science models scenarios in which generative AI produces 1.2 to 5.0 million tonnes of e-waste cumulatively over 2020–2030, depending on how deployment unfolds. The same study found that reuse and recycling could cut that waste by 16 to 86 percent. UNU separately estimates that AI infrastructure could generate up to 2.5 million tonnes of e-waste a year by 2030.

How does data center location change the impact?

The effects are not limited to electricity and water. Researchers are also examining local environmental effects, including whether AI data centers can contribute to local temperature increases

Community resistance is growing, and the IEA names social acceptability as a rising constraint. Some companies hedge against ground-based limits: Google’s orbital AI chip test puts four of its TPUs on a small satellite, though cooling in orbit remains unproven.

Benefits and burdens also split unevenly. Only 32 countries host AI-specialized data centers, and about 90% of that capacity sits in two of them, according to UNU.

Can AI reduce emissions elsewhere?

Yes, but the gains depend on adoption and honest accounting. The IEA estimates that well-documented AI use cases could save over 13 exajoules of energy by 2035, about 3% of global final energy use, if adoption overcomes the barriers. It also says AI-enabled optimization could lower energy costs in energy-intensive industries by 3 to 10 percentage points. The IEA points to grid management, industrial process optimization, and energy planning as the main areas of opportunity. At a smaller scale, AI energy management can extend the battery life of a microgrid.

Any claimed saving needs a baseline. Say an AI system trims a building’s electricity use by 10%. A fair assessment asks how much electricity and hardware the system itself needs, whether the building would have saved some of that anyway, and whether cheaper efficiency leads to more consumption elsewhere. Savings that AI enables are not automatically net savings.

How do standards and regulators measure AI’s footprint?

Measurement is catching up. In February 2026, the ITU published Recommendation L.1801, a lifecycle-assessment guide for AI systems. It makes climate impact mandatory and treats water, resource use, and biodiversity as relevant categories. It also requires identical functional units for comparisons, which matters because you cannot compare a per-prompt figure directly with a per-image, per-task, or per-year figure. It covers rebound effects as well.

Governments start with disclosure. On September 21, 2026, California’s governor signed seven data center bills. They require reporting on water and electricity use and address who pays for grid upgrades.

How can you reduce your AI footprint?

UNU recommends fit-for-purpose use: pick the lightest model and lowest-energy format that does the job.

For individuals

• Use a search engine or calculator for simple lookups.
• Skip image and video generation when text will do.
• Keep prompts focused when a task does not need long reasoning.

For organizations

• Ask vendors what their energy and water figures include, such as data center overhead and hardware lifecycle.
• Ask for efficiency metrics such as PUE (power usage effectiveness) and WUE (water usage effectiveness).
• Ask where workloads run and which grid powers them.
• Test smaller models on routine tasks before defaulting to the largest one.
• Track hardware and replacement cycles, not just operating energy.

What does the 2026 evidence support?

Three conclusions hold up. Data center electricity demand is on track to roughly double by 2030. The per-prompt cost of text is small, while video, reasoning, and agentic tasks cost far more. And the net climate effect of AI remains unsettled, because savings depend on adoption and on volume growth not erasing efficiency gains.

When you read the next AI statistic, ask one question first: what exactly does it count?

FAQs

Q. How much energy does one AI prompt use?

Google estimates 0.24 Wh for a median Gemini text prompt. Image, video, and reasoning tasks use substantially more, and other providers’ figures differ by method.

Q. Does training AI use more energy than inference?

Training takes a great deal of energy, but it happens once. After deployment at scale, inference can account for most of a model’s energy use. UNU estimates 80 to 90 percent.

Q. Does AI use more energy than a Google search?

Per query, a text prompt can cost more than a conventional search. The IEA calculates, however, that running all conventional searches as simple AI text queries would use under 4 TWh a year, less than 1% of data center consumption today.

Q. What is the biggest environmental impact of AI?

There is no single number. Electricity, carbon, water, land, materials, and e-waste all depend on the system boundary and the location.

Q. Does AI create e-waste?

Yes. One Nature Computational Science study models 1.2 to 5.0 million tonnes accumulating over 2020–2030, with large reductions possible through reuse and recycling.

Q. Is AI’s water use a problem everywhere?

No. Impact depends on cooling design, the power source, and local water stress.

Q. Can AI help the climate?

It can improve grid monitoring, industrial efficiency, and energy planning. Those benefits are plausible but not automatic.

Q. Where can I find reliable data?

Start with the IEA’s Key Questions on Energy and AI and the UNU report on the environmental cost of AI. Company environmental reports add detail, but check what each one measures.

Related:  AI Compute Heat Islands: Data Centers Are Heating the Ground