Every argument about AI ends up at the water. Somebody says a single email written by ChatGPT drinks a whole bottle of it. Somebody else says the planet is cooking so people can make pictures of cats. The other side says it’s all nonsense, and the whole internet uses more power than AI ever will.
I didn’t know who was right, and I use AI every day as a digital assistant that helps me write. So I went and looked at the numbers, from the International Energy Agency, the national labs, utility market monitors, the companies themselves and their critics. What I found doesn’t fit neatly on either side.
Think of a single AI prompt as a drip from a faucet. One drip is nothing. A billion drips a day is a lot of water, but spread over the whole country it’s still a small share of what people use. The trouble starts when all the faucets get plumbed into one building, in one town, on one stretch of river. Then the drip turns into a diversion, and the people downstream notice.
Most of the honest disagreement about AI data centers comes down to which of those pictures you’re looking at.
How much electricity do AI data centers use?
Less than most critics think, more than defenders like to admit, and climbing fast.
The International Energy Agency estimates that data centers worldwide used about 415 terawatt hours of electricity in 2024, about 1.5 percent of global electricity consumption. That covers everything data centers do: email, streaming, banking, search, cloud storage and AI. In its base case, the IEA projects data center use doubling to about 945 terawatt hours by 2030, just under 3 percent of the world’s electricity. Servers built for AI are the fastest-growing piece, rising about 30 percent a year.
The AI-specific numbers are smaller. A study published in September 2026 in Communications Sustainability looked at the AI data centers of the six biggest tech firms, Amazon, Microsoft, Google, Meta, Oracle and Apple. Together those six account for roughly 70 to 75 percent of hyperscale and cloud data center demand. It put their use at about 118 terawatt hours in 2024 and projected 239 to 295 terawatt hours by 2030, or about 1 percent of global demand. The authors note that leaving out smaller operators probably makes that an undercount.
The United States is a different story. Lawrence Berkeley National Laboratory found that data centers used 176 terawatt hours in 2023, about 4.4 percent of all US electricity, up from 58 terawatt hours in 2014. The lab projected 6.7 to 12 percent by 2028. That’s a wide range, and it reflects how unsure everybody is about how fast the AI buildout keeps going.
For scale, the average American home bought 10,791 kilowatt hours of electricity in 2022, according to the Energy Information Administration. A one-gigawatt campus running around the clock uses about 8.8 billion kilowatt hours a year. That’s the electricity of roughly 800,000 homes in one fenced lot.
Globally, data centers are a modest part of electricity growth. Carbon Brief’s analysis of the IEA figures found that data centers account for about 8 percent of the expected rise in world electricity demand to 2030, behind industry, air conditioning and electric vehicles. In the United States, roughly half of expected demand growth comes from data centers.
So the critics who say AI is eating the world’s power grid are overstating it. The defenders who say it’s a rounding error are describing the world and ignoring the US, where the growth is concentrated.
Ireland shows where this can go. Data centers used 23 percent of Ireland’s metered electricity in 2025, according to the country’s Central Statistics Office, up from 5 percent in 2015.
Nobody would call that a rounding error.
Does one ChatGPT prompt use a bottle of water?
No. That claim is the one I see most in AI arguments, and it doesn’t hold up for current chatbots.
It came from a 2024 Washington Post piece built on research by Shaolei Ren, a professor at the University of California, Riverside. The Post reported that a 100-word email written by ChatGPT used 519 milliliters of water, about a bottle. The estimate was for an older model, and most of that water wasn’t used in the data center at all. According to a Deseret News review of the research, much of it was water evaporating from reservoirs behind hydroelectric dams and cooling water at power plants that generate the electricity.
Ren himself has warned against treating one old number as a fixed fact. He told the Deseret News that a figure from two years ago can’t describe today’s systems. It’s never correct, he said, to name one amount of water that AI uses.
Newer figures are far smaller. In August 2025, Google reported that a median text prompt to its Gemini assistant used 0.24 watt hours of electricity and 0.26 milliliters of water, about five drops. Google compared the electricity to running a microwave for about one second, and said the energy per prompt had fallen by a factor of 33 in a year.
OpenAI’s Sam Altman wrote in June 2025 that an average ChatGPT query uses about 0.34 watt hours and roughly one-fifteenth of a teaspoon of water. Independent estimates cited by the Deseret News, such as those from the EcoLogits project, land between 1 and 10 milliliters per prompt.
Those company figures deserve some skepticism. They’re self-reported, and Altman’s number came without a published method. Google’s covers text prompts only. Generating images and video takes much more energy. And none of the per-prompt numbers include the electricity used to train the models in the first place.
Even so, the bottle-of-water claim is off by about two orders of magnitude for a text prompt today.
A writer worried that asking a chatbot to fix a paragraph is draining a lake can stop worrying about that.
What do billions of AI prompts add up to?
OpenAI told Axios in July 2025 that ChatGPT users send more than 2.5 billion prompts a day. Multiply that by Altman’s 0.34 watt hours and you get about 850 megawatt hours a day, or roughly 0.3 terawatt hours a year.
That’s a rough number built on a company’s unverified figure, so treat it as a sketch. Even so, it’s about the yearly electricity of 29,000 American homes. Against the 176 terawatt hours US data centers used in 2023, it’s a sliver.
So where is all that new demand going? Mostly into training ever-larger models, generating images and video, running AI inside business software, and building capacity ahead of demand that companies expect to arrive. The chat window you type into is the visible part of a much bigger machine.
Context helps here too. The Energy Information Administration’s preliminary estimate put cryptocurrency mining at 0.6 to 2.3 percent of US electricity demand in 2023. That’s in the same range as a big chunk of the AI buildout, and it gets a small fraction of the outrage in writing groups.
None of that makes the growth harmless. It means the useful argument is about the buildout, and the person typing a prompt isn’t the one building it.
Why doesn’t more efficient AI use less energy?
Between May 2024 and May 2025, Google cut the energy of a median Gemini prompt by a factor of 33. Over 2024 and 2025, its total electricity use rose 27 percent and then another 37 percent.
Both numbers come from Google.
That’s the Jevons paradox, named for a 19th-century economist who noticed that more efficient steam engines made Britain burn more coal. When something gets cheaper to use, people use far more of it. Microsoft’s chief executive, Satya Nadella, said so out loud in January 2025: “Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.”
He meant it as good news for Microsoft. It’s also the strongest answer to anyone who says efficiency will solve the energy problem on its own. Efficiency lowers the cost of each prompt. It doesn’t put a ceiling on the number of prompts, the size of the models, or the number of companies racing to build them.
Training is where the growth is steepest. Training GPT-3 in 2020 took about 1,287 megawatt hours, according to a paper by Google and Berkeley researchers, roughly a year’s electricity for 120 American homes. That was a small model by today’s standards. Epoch AI and the Electric Power Research Institute estimated in August 2025 that the largest training runs already draw more than 100 megawatts. The biggest single runs in 2030, they estimated, could draw 4 to 16 gigawatts.
A gigawatt is a nuclear reactor’s worth of power.
Those projections could be wrong in either direction. Companies could hit a wall on returns. Chips could get efficient faster than anyone expects.
The direction of travel isn’t in dispute.
How much water do data centers use in total?
Here the drip and the river matter.
Lawrence Berkeley National Laboratory estimated that US data centers directly consumed 17.4 billion gallons of water in 2023, mostly for cooling. The power plants supplying their electricity consumed about 211 billion gallons more. The Information Technology and Innovation Foundation, a think tank that generally sides with the tech industry, put the two together at less than 1 percent of total US water consumption.
Agriculture uses far more.
National totals hide the local picture. Bloomberg analyzed where new data centers have gone and found that more than two-thirds of those built since 2022 are in water-stressed areas, including parts of Texas and Arizona. Consumer Reports, citing an analysis by the nonprofit Ceres, reported that data centers around Phoenix use about 385 million gallons a year for cooling and could reach 3.7 billion gallons, enough for 34,000 homes.
Google’s own numbers show the trend at one company. Its 2026 environmental report said water use climbed 34 percent to 10.9 billion gallons, more than double its 2021 level, with data centers accounting for most of the increase.
So both sides are right about different things. Nationally, data center water use is small. In a desert city that’s already rationing its river, a big new campus is a real draw on a scarce resource, and the people who live there have every reason to ask hard questions before the permits are signed.
Cooling design matters too. Some newer campuses use closed-loop or air-based cooling that consumes little water but uses more electricity, and more electricity means more water at the power plant.
There’s no free option. Every design trades one cost for another, and the right trade depends on where the building sits.
One AI prompt is a drip. A billion prompts plumbed into one building on one river is a diversion, and the people downstream notice. – Richard LoweShare on X
How are AI data centers cooled, and why does it matter?
Every watt a computer uses turns into heat, and the heat has to go somewhere. How a data center gets rid of it decides how much water and extra electricity it needs.
The industry measures the electricity side with a number called power usage effectiveness, or PUE. A PUE of 1.0 would mean every watt goes to computing. Anything above that is overhead, mostly cooling. The Uptime Institute’s 2025 survey put the global average at 1.54, roughly where it has hovered since about 2020. The big cloud companies report 1.10 to 1.15, and Google reports 1.09 across its fleet. Older enterprise and rental facilities run 1.58 to 1.80.
The newest AI campuses are far more efficient per watt than the ordinary server rooms they’re replacing. That’s a real point for the defenders, and it rarely comes up in the arguments I see.
Water is the other side of the trade. The cheapest way to cool a building in a hot climate is to evaporate water, the same way sweat cools a body. It saves electricity and it uses water. Closed-loop systems keep the same water circulating and lose almost none of it, but they need more electricity to run chillers.
Microsoft says all its new data center designs since August 2024 use closed-loop cooling at the chip. Each one, the company says, avoids more than 125 million liters of water a year. Pilots were scheduled for Phoenix and Mount Pleasant, Wisconsin, starting in 2026. Microsoft admits the design brings “a nominal increase” in energy use. Its existing fleet used 0.30 liters of water per kilowatt hour in fiscal 2024, down from 0.49 in 2021.
So a company can pick its poison. Evaporate water in Phoenix and use less power, or close the loop and use more power from a grid that also needs water. The useful version of the water debate is about which trade a particular site should make, and who gets a say.
Are AI data centers raising electricity bills?
In some places, yes. Nationally, the evidence is mixed.
The strongest evidence comes from PJM, the grid operator serving 13 states and Washington, D.C., in the Mid-Atlantic and Midwest. PJM runs auctions to pay power plants to be available years ahead, and the costs flow into customer bills.
Its independent market monitor, Monitoring Analytics, attributed $6.3 billion, or 38 percent, of the July 2026 auction’s $16.4 billion in costs to data centers, and $29.4 billion across the four most recent auctions. Joseph Bowring, the monitor’s president, said ratepayers in PJM pay capacity charges for existing and potential data centers, and also pay for the higher energy and transmission costs data centers have caused.
Bloomberg found that wholesale electricity in some areas near heavy data center activity cost as much as 267 percent more than five years earlier. Wholesale prices aren’t retail bills. State regulators set retail rates, and utilities spread costs across years, but a share of those wholesale costs reaches households sooner or later.
The other side has evidence too. A review by the consulting firm Energy and Environmental Economics, known as E3, found no evidence that data centers have historically been subsidized by other customers. It noted that Texas and Virginia, two of the states with the most load growth, had some of the smallest rate increases, while California and New York had large increases with flat or declining demand. More customers sharing the fixed cost of the grid can lower everybody’s bill.
Know who paid for it, though. The Data Center Coalition, an industry group, funded the E3 report. And even E3 found that about half of PJM’s 2025/2026 capacity price increase came from load growth, with the rest from plant retirements and market design. Industry money doesn’t make the findings wrong. It means they deserve the same scrutiny you’d give a study paid for by the other side.
Residential electricity prices rose 7.1 percent in 2025, Consumer Reports reported, more than double the inflation rate. Data centers are one cause among several, including fuel prices, aging transmission lines, storm damage and power plant closures. Where they cluster, as in PJM, they’re a big one. In states where utilities make data centers pay for their own grid upgrades, everyone else feels little of it. E3 counted 30 new large-load tariffs written in 2025 and 2026 alone to do that.
Do AI data centers increase carbon emissions and air pollution?
Yes, and the companies’ own reports say so.
The IEA estimates data centers produce about 180 million tonnes of carbon dioxide a year through the electricity they use, about 0.5 percent of global combustion emissions. It expects that to reach 1 percent by 2030 in its base case, and 1.4 percent in a faster-growth case. The IEA calls data centers one of the few sectors, along with road transport and aviation, where emissions are set to rise. That’s still a small slice of the world’s emissions, and it’s growing while most other slices are expected to shrink.
Google’s 2026 environmental report showed its greenhouse gas emissions up 18 percent in a year, its largest annual increase, and its electricity use about 3.5 times its 2019 level. The company signed a record 12 gigawatts of clean energy agreements and still couldn’t keep emissions flat. “This rapid expansion in energy demand is a reality we must manage actively,” Google wrote.
Air pollution is a local issue again. The starkest example is xAI’s data center in Memphis. It ran gas turbines on site while it waited for grid power, and I go through what happened there below.
Some of the money is flowing toward cleaner power. Microsoft signed a 20-year agreement with Constellation Energy to restart Unit 1 at Three Mile Island, now called the Crane Clean Energy Center. Federal regulators are expected to rule on its operating license in 2027, according to WITF, and the plant would employ about 600 people. Google, Amazon and others have signed deals for new nuclear and renewable power.
The optimistic case comes from the IEA too. It estimates that wider use of existing AI tools in energy, industry and transport could cut emissions by about 1,400 million tonnes a year by 2035, three to four times the total emissions data centers are projected to produce. The IEA was blunt about the catch: “there is currently no momentum that could ensure the widespread adoption of these AI applications.” That’s a possibility, and nobody has earned credit for it yet.
Why are tech emissions rising when companies buy so much clean energy?
Because the clean energy and the data center don’t always run at the same time, or on the same grid.
For years, most companies used annual matching. If a data center used a million megawatt hours in a year, the company bought a million megawatt hours of renewable energy certificates somewhere and called the year clean. The data center might still have run on gas at 2 a.m. while a wind farm in another state got the credit.
The difference is large. A Guardian analysis in September 2024 recalculated the data center emissions of Apple, Google, Microsoft and Meta for 2020 through 2022 using the actual mix of the local grids, known as location-based accounting. It found emissions about 7.6 times higher than the companies’ official figures. Meta reported 273 metric tons of carbon dioxide from its data centers in 2022. The location-based number was about 3.8 million.
The companies’ method isn’t a trick they invented. Standard greenhouse gas accounting allows market-based reporting, and certificates do pay for renewable projects that might not otherwise get built.
A certificate doesn’t change what comes out of the local power plant on a windless night.
Google has started measuring the harder way. It tracks how much of its electricity is carbon-free hour by hour, on the same grid. In 2024 that figure was 66 percent globally, about 92 percent in Latin America, around 70 percent in North America, 12 percent in Asia-Pacific and 4 percent in Singapore. That’s a more truthful number than a 100 percent annual claim, and it shows exactly where the gaps are.
Critics who say tech companies’ green claims are inflated have a point. Defenders who say the companies are funding a lot of new clean power have a point too.
What happened with xAI’s data center in Memphis?
Memphis is the clearest case of the drip and the river.
Power, water, air, jobs and promises all collided there within two years.
In June 2024, Elon Musk’s xAI announced it would build what it called the world’s largest supercomputer in a former factory in South Memphis. The Greater Memphis Chamber called it the largest capital investment by a new-to-market company in the city’s history. The project was approved fast, and local officials later said they’d been told little.
The grid couldn’t supply all the power at first. In November 2024 the Tennessee Valley Authority approved 150 megawatts for the site, and xAI agreed to fund a $24 million substation, provide discounted Tesla battery storage to help stabilize the local grid, and build a wastewater recycling plant. Those were real commitments with real value to the utility.
To run in the meantime, xAI brought in portable gas turbines. According to the Southern Environmental Law Center, aerial images in April 2025 showed 35 of them. The mayor said only 15 were operating. In July 2025 the Shelby County Health Department permitted 15 turbines after a hearing where hundreds of residents spoke against it. The surrounding area already failed federal smog standards. After the NAACP served notice that it intended to sue, xAI removed the unpermitted turbines from the Memphis site, according to the SELC.
Then the expansion moved across the state line. An xAI affiliate bought a former power plant site in Southaven, Mississippi, and in March 2026 state regulators approved a permit for 41 turbines there, three weeks after a public hearing where no one spoke in favor. In April 2026 the NAACP, represented by the SELC and Earthjustice, sued xAI under the Clean Air Act over turbines it said were running at the Southaven site without permits.
Water was the other promise. The Memphis region drinks from the Memphis Sand Aquifer, and xAI pledged to cool its data centers with recycled municipal wastewater to protect it. The plant broke ground in October 2025. In April 2026 construction stopped. According to E&E News, xAI said it needed “to focus on finishing Colossus 2 and ensuring it is extremely stable, then will build the water recycling plant.”
The company had no legal obligation to finish it.
Look at what Memphis shows. The investment and the grid upgrades were real. So were the turbines, the smog, the unanswered questions from local officials, and a water promise that slipped as soon as something else took priority. None of that is visible from a chat window, and none of it shows up in a per-prompt number.
It shows up in one neighborhood.
Do AI data centers help the local economy?
They pay a lot of taxes and employ very few people.
Loudoun County, Virginia, has the largest concentration of data centers in the world. The county collected about $1.1 billion in taxes on the computer equipment inside them last year, roughly 38 percent of its general fund, according to Moneywise. It has cut its real property tax rate every year for a decade, from $1.145 per $100 of assessed value in 2016 to $0.805 in 2026.
That’s a real benefit to every homeowner in the county.
Jobs are a different matter. A data center needs thousands of construction workers for a couple of years and then a small permanent staff. When Vantage Data Centers planned a 1.1 million-square-foot campus near Reno, 2024 business records projected more than 4,000 temporary construction jobs and 73 permanent jobs over the following decade. Researchers at the University of Michigan concluded that data centers don’t bring high-paying tech jobs to their communities and compared them to infrastructure like bridges and highways.
Nationally, the spending is enormous. The five biggest US cloud and AI companies committed $660 billion to $690 billion in capital spending for 2026, according to the Futurum Group, nearly double 2025. Harvard economist Jason Furman calculated that investment in information processing equipment and software accounted for 92 percent of US GDP growth in the first half of 2025. He also pointed out that without the AI boom, lower interest rates and electricity prices might have made up about half of that growth elsewhere.
That kind of concentration cuts both ways. It props up the economy while the money flows. If the buildout slows, a lot of growth goes with it. I follow that money in more detail in the economy article.
Why are communities fighting AI data centers?
Opposition is growing fast, and it doesn’t break down along party lines. Data Center Watch tracks those local fights, and it counted 75 projects worth about $130 billion blocked or delayed in the first quarter of 2026, about as much as all of 2025. Moratorium proposals came up in 14 states from both sides of the aisle.
The complaints are concrete. Residents near Loudoun’s campuses complain about the hum of cooling equipment and backup generators, and about new high-voltage transmission lines. One county supervisor put it bluntly to The New York Times. “We’ve become addicted to the data centers for their tax revenues, but at what cost?” In Louisiana, police records obtained by the Gulf States Newsroom showed vehicle crashes around Meta’s Hyperion construction site jumped more than 600 percent in the first nine months of 2025.
Then there’s the hardware. A 2024 study in Nature Computational Science projected that generative AI could produce 1.2 to 5 million tons of electronic waste between 2020 and 2030, as servers get replaced every few years. The same study found that reuse and recycling strategies could cut that by 16 to 86 percent.
None of these complaints are imaginary. They’re the cost of a big industrial building, and the people who live next to it pay that cost while the benefits go to the county budget and to customers all over the world.
What would the world look like without AI data centers?
Picture the same world with the AI buildout erased. What changes?
Data centers wouldn’t disappear. Email, streaming, banking, search and cloud storage ran in them long before anybody had heard of ChatGPT. What would disappear is most of the recent growth. Lawrence Berkeley National Laboratory found that data center power demand more than doubled between 2017 and 2023, largely because of AI servers.
The grid would look like the one we had for years. US electricity use was nearly flat for close to two decades, from the mid-2000s to the early 2020s, according to the EIA, as efficiency gains offset growth. Now it’s climbing, and data centers are a big reason. Without AI, the capacity crunch in the Mid-Atlantic would be smaller, Google’s and Microsoft’s emissions would probably be lower, and Memphis wouldn’t have had those turbines.
The other column has entries too. Without the buildout, Loudoun County loses more than a third of its general fund revenue and has to raise property taxes or cut services. The Three Mile Island reactor stays shut, because Microsoft’s 20-year contract is what’s paying to restart it. The construction jobs vanish. By Jason Furman’s math, a large share of US economic growth in early 2025 vanishes with them. The IEA’s possible 1,400 million tonnes of emission cuts from AI applications drops to zero, though nobody had earned that credit yet anyway.
And the work AI does would still need doing. A 2024 study in Scientific Reports by researchers including Bill Tomlinson of the University of California, Irvine, estimated that AI emits 130 to 1,500 times less carbon per page of text than a human writer, counting a share of the person’s footprint during the time spent writing.
The authors said their numbers leave out job displacement, legality and rebound effects. The obvious objection is that the human writer goes on living and emitting either way. Still, the study makes a point the per-prompt arguments miss: the comparison that matters is AI against whatever would have done the job without it.
Last, the demand doesn’t vanish if one place says no. The IEA expects the United States and China to account for nearly 80 percent of data center growth to 2030. Block a campus in one county and the computing gets built in another county, or another country, on whatever grid that place has. Google’s hourly carbon-free figures run from 92 percent in Latin America to 4 percent in Singapore. Where the building goes changes its footprint more than whether it gets built.
Which AI environmental claims are true?
The bottle of water per email is false for today’s text chatbots. One person’s chatbot habit makes almost no difference to the planet. A median text prompt uses about as much electricity as a microwave running for one second. If you’re worried about your personal footprint, your car, your air conditioner and your diet matter far more than your chatbot.
AI’s share of global electricity is small today. It’s also growing fast, and most of that growth lands in the United States. Data centers already use about 4.4 percent of US electricity and could use as much as 12 percent within a few years.
Data centers are raising power bills in parts of the Mid-Atlantic grid, according to that grid’s own market monitor. Across the whole country, the evidence is mixed, and the answer depends on how each state makes the companies pay for the grid upgrades they need.
Local water stress, air pollution from on-site gas turbines, noise and new power lines are real where they happen. Those are the strongest arguments the critics have, and every one of them comes down to one building in one place. A county that approves a campus without asking where the water and power will come from has made a mistake, and the residents will be living with it for decades.
As for AI saving the climate, that’s a hope.
The IEA’s own projections say so.
Is the environment argument against AI fair?
Some of it is. A town deciding whether to approve a two-gigawatt campus beside a stressed aquifer should argue about water and power, loudly, and nobody should call those people cranks for asking. Asking hard questions before the permits are signed is how communities protect themselves. And nobody should sneer at a family in South Memphis worried about turbine exhaust in a neighborhood that already fails federal smog standards. Those are the people the environmental argument was made for.
But I see a different version of the argument in writers’ groups every week. An author uses an AI image on a book cover, and somebody tells her she’s draining the planet’s drinking water. Somebody else posts that AI is killing the planet and everyone will run out of water. That’s AI Derangement Syndrome borrowing a real issue to win a fight about something else. The author didn’t pick the site for the data center. She didn’t sign the utility contract or decide whether the turbines got permits.
The water she supposedly drank was a few drops, if that.
Somebody who shames a writer over a few drops of water does nothing for the river. They get to feel righteous while treating a stranger like dirt, and they make the people with real grievances easier to ignore.
The decisions that matter get made by utility commissions, county boards, state legislatures and the companies building the campuses. An argument about data centers can change something there. I’ve written more about the fears that turn out to be justified in when the worried are right about AI, and about the anger behind the writers’ side of this in why writers are so angry about AI.
There’s also a fair point the defenders miss. Telling a farmer in a drought county that data centers use less than 1 percent of the nation’s water is technically true and useless to him.
He doesn’t live in the nation. He lives next to the building.
What rules would reduce the environmental impact of AI data centers?
I think AI data centers are inevitable. They’re going to get built. You can bitch and moan about it all you want, and it won’t stop a single campus.
What we can do is make sure rules, policies and procedures get written to reduce the impact on the environment. Push them toward more energy alternatives, and make sure they aren’t dependent on the grid the rest of us rely on.
That isn’t a fantasy. Ireland did a version of it. After a four-year freeze on new grid connections for data centers around Dublin, the country’s energy regulator set new rules in December 2025. A data center seeking a connection now has to install on-site generation or batteries able to meet its full demand, send power back to the grid when it’s needed, and source at least 80 percent of its yearly electricity from new renewable projects.
Texas took a different route. A state law signed in June 2025, Senate Bill 6, requires customers over 75 megawatts to pay a share of the cost of connecting to the grid, lets the grid operator cut them off or make them switch to their own generators in an emergency, and makes them disclose backup power that can cover half their demand. Add the 30 large-load tariffs E3 counted in 2025 and 2026, Microsoft’s closed-loop cooling, and Google’s hour-by-hour clean power accounting, and the toolkit already exists.
One warning from Memphis. Off the grid doesn’t automatically mean clean. xAI’s turbines were off the grid too, and they were burning gas in a neighborhood that already failed smog standards. The rule that works is off the grid with clean power, or on the grid paying its own way and giving power back in a crisis. Write that into the permits before the concrete gets poured, and the drip stays a drip.
What should a writer who uses AI do about its environmental impact?
Use it for work that’s worth doing, and don’t waste it. Generating fifty versions of an image you’ll throw away costs far more than asking for help with a paragraph.
Text is cheap. Video is expensive.
If the issue matters to you, take it where decisions happen. Show up when your county considers a data center. Read your utility’s filings on large-load tariffs. Ask the companies building near you how they’ll cool the buildings and where the power will come from. Those questions can change an outcome.
Yelling at a stranger about her book cover changes nothing.
And check the numbers before you repeat them. The bottle-of-water claim spread because it was vivid, and it was wrong. The true story is more complicated and more useful.
The drip and the river
Your prompt is a drip. It always was. The question worth asking is where all those drips get collected, who pays for the pipes, and whose river they come out of. Ask that, and the argument stops being about whether AI is evil and starts being about whether a particular building belongs in a particular place.
The AI and writing hub has more on using AI as an author with your eyes open. If you’d like help figuring out where AI belongs in your own writing work, my AI services cover that. Turn off the faucet when you’re done. Then go to the meeting about the river.
