Understanding the Environmental and Community Impact of AI and Data Centers

Ask an AI chatbot a simple question, and its answer uses about as much electricity as a microwave running for one second.

Now give that same AI a big job and let it work on its own. This is called an agent. It plans its own steps, searches the web, tries things, and keeps going until it thinks it is done. One of those jobs can use as much power as running your laptop for almost six hours. Same technology. Very different bill.

That difference introduces a new choice with considerable environmental impact, and it was identified after the research many commentators on AI keep quoting was published. Moreover, electricity is also only part of the story. The rest of the cost is in water systems, on your monthly power bill, and in local tax revenue, which is where the real fight over AI is happening right now.

This article is not an argument for or against using AI. It is a look at AI’s environmental cost, where those costs land, and who gets to decide. With AI poised to transform life, if not make and unmake fortunes, we want to equip you to make better-informed decisions about how to use AI, or whether to accept a plan to build a data center in your community.

AI Wins By a Mile?

In February 2024, researchers at three universities published a comparison in the journal Scientific Reports. They found that an AI writing one page of text emits 130 to 1,500 times less carbon dioxide than a person writing the same page.

Here is how they got there. An AI query produces about 2 grams of carbon dioxide, including its share of the energy used to train the model. A person writing the same page produces about 1,400 grams.

But consider how they calculated the figure for the person. They took the average American’s total yearly carbon footprint, which is roughly 15 tons, and divided it by the number of hours in a year. That gives about 1.7 kilograms an hour. Then they charged the writer for the 48 minutes it took to create the page.

Readers caught the problem right away. Comments posted on the study’s web page pointed out that people emit carbon whether or not they are writing. The writer does not disappear when an AI takes over her work. She still heats her house. She still drives to the store. Those 15 tons go into the air either way.

What actually changes is the electricity her laptop used for the job. Over a multi-day project, that comes to about 2 kilograms of CO2 emissions, not the 36 kg the study’s method assumed.

The Rematch, With the Answers Graded

In November 2025, the same journal published a direct response. Its author, Nolan Woo, said the first study skipped the most important question: whether the AI’s work was any good.

So he set up a test with graded answers and factored in the time and energy to arrive at correct answers into the total emissions. He used programming problems from a national high school computing contest, which come with automatic pass-or-fail scoring. After running four different AI models against the coding challenges, he fed the errors back into the AI and asked it to try again, up to 100 times.

The results differed based on the AI model used:

• The smallest model, when it got the answer right, put out 20% to 59% of what a human programmer’s computer would.
• A larger model, like ChatGPT or Claude, produced five to 19 times as much as the human.
• Failed attempts used 8 times as much energy as successful attempts.

What drove emissions was not the model’s size or cost. It was how many tries it took to get the job done.

There is one more detail that is easy to miss. Woo could barely run the comparison at all. Across 4 years of contests, the AI solved every problem in only one of them. Most of the time, it simply could not do the work. However, recent models have dramatically improved their coding capabilities over the past year, achieving above 95% accuracy, according to Stanford University’s AI Index Report.

Then Agents Changed the Math

Both studies measured a single question and a single answer. That is not how people use AI in 2026. Today’s tools plan, search, read, revise, and loop, sometimes for hours. Jobs where AI reasons through many steps can use hundreds or thousands of times more energy than simply generating text.

A team at KAIST, South Korea’s top engineering school, measured the “hidden energy” cost of AI. Using a model similar in size to commercial ones, they found that a single-agent job used an average of 348 watt-hours. That is 136 times what the same model used to answer a single, simple question.

Two things explain where the power goes. These agentic tasks take much longer, up to 154 times longer, and involve many AIs performing different parts of the job. The team found that the chips doing the work sat idle for more than half the test, still drawing power while waiting for a web search or a program to finish.

In other words, more than half of the energy bill goes to expensive hardware sitting idle. On the other hand, the International Energy Agency found the technology is becoming more efficient an April 2026 report, saying that, “Measured per individual task, the energy efficiency of AI is improving at a rate unprecedented in energy history.” There is a lot of waste in the current infrastructure that will be wrung out simply to cut the cost of running AI, which will make it more profitable — it’s still a money-loser on the whole today — and lower its environmental cost.

The Break-Even Point

Put it together, and you can find the point where AI stops being the cleaner choice.

Picture a research project that would take a person 20 hours to complete. They will read about 40 sources and write a 5,000-word report. Count only what that job adds: the laptop, the monitor, and their share of the pollution from building them. That comes to roughly 2 kilograms of carbon.

So how many AI agent requests does it take to pass the 2 kilogram mark? It depends on how clean the power is at the data center.

That bottom row matters. Big AI data centers tend to get built where power is dirtier than average. One 2026 study found the electricity serving them runs about 48% dirtier than the national average.

Most serious AI research jobs today use far more than 11 requests.

The Catch Nobody Catches

Every version of these analyses that favors AI rests on one assumption: that the machine replaces the person’s hours rather than adding to them.

The evidence says otherwise. We work harder and longer with AI.

A National Bureau of Economic Research study, which is still under review and therefore not fully validated, found that people heavily exposed to AI at work put in an extra 3.5 hours a week. Surveys by two staffing and consulting firms also found that only about a quarter of employees turn AI time savings into personal time. The rest gets filled with more work.

Say an AI workflow produces 3 times as much as a person would. If it gets used 3 times as often because it is now fast and cheap, you are at 9 times as often.

This is an old pattern known as Jevons’ Paradox. In 1865, an economist named William Stanley Jevons noticed that more efficient steam engines did not reduce the amount of coal Britain burned. Cheaper power just meant more uses for it, and coal use climbed for decades.

Water Impacts

Water comes up in almost every local fight over data centers, and the number people quote is usually the smaller one. Lawrence Berkeley National Laboratory found that U.S. data centers used about 17.4 billion gallons of water in 2023 to cool their buildings. While annual water use at data centers is less than a single day’s irrigation of the nation’s corn crops, according to Orennia, that could double or even quadruple by 2028.

The bigger number is the one almost nobody reports. Producing the electricity those data centers use took roughly 211 billion gallons of water. That is about 12 times the cooling number. Power plants are the largest user of fresh water in the country, and every data center owns a share of that through its power bill.

A data center can switch to air cooling and then honestly say it uses almost no water on-site. But its real water footprint has not changed at all. Only a data center powered by clean energy without evaporative cooling brings both numbers down.

What companies report is patchy. Google said its data centers used 7.7 billion gallons worldwide in 2024. Its Council Bluffs, Iowa, site alone used about a billion gallons. In Texas, 83% of the state’s 341 data centers had not filed required water reports.

Industry groups point out, correctly, that all data center water use together is still under 1% of the country’s total. Both things are true. The national average is not much comfort to a town in the Arizona desert.

The Charge On Your Power Bill

The cost that moved fastest from technical detail to kitchen table argument is the price of electricity.

Here is how it works. PJM is the company that runs the power grid for 65 million people across 13 states and Washington, D.C. Every year it holds an auction, paying power plants to promise they will have electricity available on the hottest and coldest days. Those payments get added to customer bills.

That auction price rose from about $29 per unit for 2024-2025 to about $329 per unit for 2026-2027. The independent monitor that watches PJM’s market found data centers caused 63% of one year’s jump. That works out to $9.3 billion collected from customers. The Natural Resources Defense Council projects that the average family in the region will pay about $70 more per month by 2028.

The evidence is showing up in utility bills across the country. Customers of Pepco in Washington, D.C., saw about $21 a month added. Home electricity rates in Ohio and Pennsylvania rose by 9% and 14% over the past year.

The industry says that is only half the picture. Virginia’s state watchdog agency, JLARC, found that data centers support about 74,000 jobs and $9.1 billion annually for the state economy. It also noted most of that comes from building them, not running them. Developers also argue that adding one large customer can spread the grid’s fixed costs across more people and lower everyone’s rate. That can be true. It depends entirely on how the contract is written.

States are testing whether the grid can be managed to segregate household power and its price from energy destined for data centers. Virginia set up a separate rate category for data centers. Ohio approved a rule requiring 12-year contracts and minimum payments for the biggest users. The idea in both cases is to have the new demand pay for the power plants it requires, rather than spreading that cost to everyone.

Why Towns Are Saying No

All of this has produced the biggest fight over land use in recent American memory. Data Center Watch, a group that tracks local opposition, counted at least 75 projects worth about $130 billion blocked or delayed in just the first 3 months of 2026. That roughly matches all of 2025. The number of active opposition groups more than doubled to 833, spread across 49 states. Over 300 related bills were filed in statehouses in the first 6 weeks of the year.

Opposition to data centers crosses party lines, which is rare for environmental issues. Republican officials tend to focus on tax breaks and strain on the power grid. Democrats tend to focus on water and pollution. But they end up voting the same way.

Two towns show what happens next, and each story ended differently.

Tucson, Arizona, is a winner. Amazon wanted a 290-acre campus that could have grown to 10 buildings. The project moved through Pima County under a confidentiality agreement signed in 2023 that hid Amazon’s involvement from the public and even from some elected officials. County supervisors approved a land sale in June 2025 without knowing the full picture. When residents found out how big it was, more than 1,000 of them packed a hearing. That August, the city council voted unanimously to end the project, refusing to annex the land or supply city water. The developer called it a missed chance at millions in tax revenue and thousands of jobs.

Then the developer bought the land anyway, under a different company name, and moved the project just outside the city limits, where the council’s vote did not apply. The organizing worked, but the city’s authority to stop the project ended at the edge of town.

Saline Township, Michigan, is the harder story. In September 2025, both the planning commission and the township board voted against rezoning 575 acres of farmland for a $16 billion data center tied to OpenAI and Oracle. Two days later, the developer sued the township. It argued that because the township had no land zoned for industry at all, it was illegally shutting out a use it could not ban.

The township, with about 2,900 residents, could not afford a long court fight. It settled within weeks. Construction started in November. Residents negotiated about $14 million in community benefits, including funding for the fire department and farmland protection. One township supervisor voted against the settlement and was the only one to do so.

Both fights are now happening at a much bigger scale. In Virginia, a coalition of data center companies, chambers of commerce, and construction unions spent months defending a multibillion-dollar tax break in a television ad campaign. Nationally, a political group called Leading the Future launched with $140 million to back pro-AI candidates, including $50 million each from a venture capital firm and OpenAI’s president. Employees at AI companies started their own group, the Guardrails Alliance, raising small donations to push back.

New York recently became the first state to pause new large data center construction. At least 19 Michigan towns passed their own building freezes after Saline.

What You Can Do

Your own choices matter less here than the infrastructure decisions do. But your use of AI will ultimately shape how widespread the data center infrastructure becomes.

• Match the tool to the job. A quick chatbot question costs a few grams of carbon, about what a car puts out in 2 seconds. Handing that same question to an agent for 100 steps can cost 19 kilograms, or about 47 miles of driving.
• Stop the loops. Failed and repeated tries drove emissions more than anything else. If a tool is spinning without getting anywhere, shutting it down is both the sensible move and the low-carbon one.
• Smaller is not always better. A small model that fails 20 times can put out more carbon than a big model that gets it right once.
• Read your power bill. Auction costs usually appear as a separate line item, often called a capacity charge. If you live in one of those 13 states, that line is where data center demand reaches you.
• Go to the utility hearing, not just the zoning meeting. State utility commissions decide who pays for new power plants, and those meetings take public comment. That is where the question of who bears the cost is actually settled.
• Ask what confidentiality agreements cover. Several of these fights turned on residents learning that their own officials had signed away the right to say who the customer was. Any resident can ask that at a public meeting.
• Ask who is counting, and how. When you see a claim that AI is a thousand times cleaner than a person, check whether someone put that person’s commute and house on the bill. That one choice changes the answer enormously.

Your Next Choice Matters

For quick, clear jobs an AI handles on the first try, it really is the lower-impact choice. For open-ended research where an agent must loop, search, and correct itself, the person is currently cleaner, and the gap grows with every extra step.

The bigger point is that the energy used per question is not the number that matters most. Where the data center gets built, which utility serves it, whose water it draws, and who signs the power contract will shape AI’s environmental cost far more than any efficiency improvement. Those are public decisions. They get made in rooms that take public comment. And right now, they are getting made very fast.