The Environmental Impact of AI Data Centers, Impact by Impact

The environmental impact of AI data centers gets argued about in very broad terms. In practice it breaks into a handful of separate effects, each with its own cause and its own fixes. Some are global, like carbon. Others are felt next door.

This guide takes them one at a time, with a public source wherever a number exists. It is written by a company that plans to power AI capacity with natural gas burned on site, so it is plain about what that means: on-site gas emits carbon dioxide and nitrogen oxides, and it needs air permits. The table near the end lists what to ask any developer, us included.

Where does the environmental impact of AI data centers come from?

Most of it comes from electricity. A 2026 working paper on US AI data centers finds that their emissions are driven mainly by how much electricity they use, and that the outcome depends on the grid, water and land systems around each site, not on the building alone.

It helps to sort the effects by where they land:

• Electricity and the grid. AI racks draw far more power than conventional servers, and large sites add load that the grid has to supply.
• Carbon. Set by the power source. The same building has a very different footprint on a coal-heavy grid, a gas plant or a hydro-rich grid.
• Local air. Diesel backup generators, and on-site gas engines or turbines where a site makes its own power, release nitrogen oxides and other pollutants near homes.
• Water, land and noise. Cooling can use a lot of water, campuses take up land, and fans, chillers and generators make noise.
• Hardware. Servers and accelerators are replaced every few years, and their manufacturing and disposal carry their own footprint.

Not every site has every problem. A site with closed-loop liquid cooling uses little water. A site on a clean grid may still run diesel generators. A fair answer starts with the specific site.

How do AI data centers affect the power grid?

They add a lot of new demand, quickly. Lawrence Berkeley National Laboratory estimates US data centers used 192 TWh in 2024, or 4.7% of US electricity. Its reference case reaches 649 TWh by 2030, about 11.8% of US electricity. For the full set of figures, see our guide on how much power a data center uses.

The strain on the grid shows up in three places:

• Connection waits. Utilities have to study and build for each large load. In busy regions the wait in the data center interconnection queue runs to years.
• More generation. New load has to be met by someone's power plants. The US Energy Information Administration expects most of the near-term increase to come from gas. In its higher data center demand case, gas-fired generation rises 7.3% from 2025 to 2027.
• Local peaks. A single large campus can change the load on a substation or a whole utility zone, which drives new lines, transformers and, sometimes, higher costs shared by other customers.

Efficiency helps but does not erase the load. LBNL estimates an average power usage effectiveness of 1.145 in 2024 for facilities serving AI equipment, so little power goes to cooling and losses. The servers still need the rest.

How big are data center emissions?

Data center emissions depend almost entirely on how the electricity is made. The building's own fuel use is small, apart from generators. So the useful question is not how much a data center emits, but how much each kilowatt-hour it uses emits.

EIA's figures for US power plants in 2023 show the spread:

Read that table honestly. Gas emits well under half as much CO2 per kWh as coal. It also emits more than the US average, because the average includes nuclear, hydro, wind and solar, which emit almost no CO2 when running. In 2025, fossil fuels made about 58% of US electricity, nuclear about 18% and renewables about 24%, according to EIA.

Two more things shape the real number:

• Methane upstream. Natural gas is mostly methane, and leaks from wells, pipelines and equipment add to its climate impact. The EPA says methane is more than 28 times as potent as carbon dioxide at trapping heat, and names oil and gas systems among the largest human-caused sources in the US.
• Equipment efficiency. A large combined-cycle plant turns more of its fuel into electricity than small engines or simple-cycle turbines do. Less efficient equipment burns more gas, and emits more CO2, per kWh.

Do AI data centers cause air pollution?

Yes, wherever they burn fuel on site. Almost every large data center has backup generators, most of them diesel, and a growing number make their own power from natural gas. Both release nitrogen oxides (NOx), plus carbon monoxide and particulates in varying amounts.

NOx matters locally. The EPA says nitrogen dioxide comes mainly from burning fuel, can worsen asthma, and reacts in the air to form ozone and fine particles. Children, older adults and people with asthma are most at risk.

How much a site emits depends on how often its engines run:

• Backup generators mostly sit idle and run for testing and outages. For permit estimates, EPA uses 500 hours a year as a default worst case for an emergency generator. Certain emergency engines may also run up to 50 hours a year outside emergencies to support the grid, under EPA's rules.
• On-site gas generation that powers the site day to day runs nearly all the time. It is a small power plant, and it is regulated like one, with emission standards for engines and turbines and, usually, controls such as catalysts to cut NOx.

In the United States, state and local air agencies issue most air permits for data centers, under Clean Air Act rules set by the EPA. Permits set limits, equipment and sometimes operating hours, and large sources may need air modeling to show they will not push nearby air quality past federal standards. Canadian provinces run their own approvals. Our guide to gas-powered data centers covers the equipment side.

What about water, land, noise and hardware?

Each of these is real, and each depends heavily on design and location.

• Water. Evaporative cooling consumes water on site, and power plants that feed the site use water too. Closed-loop and liquid cooling use far less at the site. We cover the details in our guide to data center water usage.
• Land. Large campuses cover many acres, plus substations, transmission lines and, for on-site power, a generation yard and fuel connection. Reusing industrial or already disturbed land avoids clearing new ground.
• Noise. Cooling fans, chillers, transformers and generators make a steady hum, and generator tests are louder. Distance, enclosures, quieter fans and barriers reduce it. Local noise rules vary widely.
• Hardware and e-waste. AI servers are typically replaced every few years as new chips arrive. Building them takes energy and materials, and retired equipment needs reuse, resale or proper recycling.

There is little standard public reporting on these effects, site by site. That makes direct questions to the developer more useful than headlines.

Is on-site gas better or worse than grid power for the environment?

It depends on the grid it is compared with, and honest answers go both ways. Against the US grid average, gas generation emits more CO2 per kWh. Against a coal-heavy grid, it emits less. Against a grid rich in hydro, nuclear or wind, it emits much more.

What on-site gas changes:

• It owns its emissions. The CO2 and NOx come from a stack on the site, not from a plant somewhere else, and they show up in the site's own air permit.
• It adds no load to a strained grid. A site powered behind the meter does not draw on the shared grid, so it does not compete with homes and businesses for scarce grid capacity. Where the grid would meet that same load with new gas plants, the emissions are similar in kind; they simply move onto the site.
• Fuel source matters. Landfill gas is roughly half methane, and the EPA ranks landfills as the third-largest source of human-related methane in the US, about 14.4% in 2022. Capturing and burning it prevents that methane from escaping. Burning it still releases CO2 and NOx.

On-site gas is not a clean option. It is a fast, firm option with a visible footprint that has to be permitted, controlled and reported. Whether it is the right trade-off for a given site is a fair question to ask.