Somebody told me that one paragraph I’d written with AI wasted tons of water. People say things like that about AI all the time now, and they say them with total confidence.
I’m the wrong person to say it to, and probably the right person to check it. I pay $200 a month for Claude Max and hit the limit just about every week. My whole working week is arranged around that allowance. If anybody’s draining a reservoir with a chatbot, it’s me.
So how much do I use? I read my own meter to find out.
There’s an electric meter on the side of your house. Nobody knows what their toaster costs, or the porch light, or the phone charger. You find out what the household uses by reading the meter at the end of the month, and then you compare the number with things you understand: last summer’s bill, the neighbor’s pool pump, the second refrigerator in the garage. A number by itself means nothing. A number next to your other habits tells you whether to worry.
How much electricity does a heavy Claude user consume?
Anthropic doesn’t publish energy figures per request, and I couldn’t find a public number for Claude. Nobody outside the company can read this particular meter exactly.
What exists are careful outside estimates. In January, a programmer named Simon Couch worked out the electricity behind his own use of Claude Code, Anthropic’s coding agent, and published the math. He built his per-token figures from Epoch AI’s energy estimates for GPT-4o and from Anthropic’s API prices, and he’s upfront that they’re guesses. His answer came to about 1.3 kilowatt hours on a heavy workday, for roughly $15 to $20 of usage at API prices.
My usage is much bigger than his.
When the research firm SemiAnalysis ran Claude’s top plan to its limit, Tom’s Guide reported it came to about $8,000 of API-equivalent usage in a month. I use 90 to 95 percent of my allowance. Scale Couch’s electricity by those dollars and my month lands somewhere around 450 to 650 kilowatt hours.
Dollars make a decent yardstick for this, because API prices are set per token. More spending means more tokens processed, and more tokens means more electricity. The prices overstate what the work costs Anthropic, but they track how much work got done.
A second, independent estimate keeps that number honest. The climate scientist Zeke Hausfather logged eight weeks of his own Claude Code use from the exact token counts the tool records, and converted them three ways. His central figure was about 3 kilowatt hours on an average day, and about 11 on his heaviest. Hausfather’s logs also show what heavy agent work is made of. He sent 1,138 prompts that triggered about 14,000 model calls and 3.2 billion tokens, and 96 percent of those tokens were the agent rereading context it had already seen.
I hit my weekly limit, so I work closer to his heaviest days than his average ones. Thirty of his heaviest days come to about 330 kilowatt hours. Couch’s method, scaled by dollars, gave me 450 to 650. Put the two together and my honest range is roughly 300 to 650 kilowatt hours a month.
Treat that as a range with wide error bars.
Couch’s per-token numbers could be off by a factor of two or three in either direction, and mine inherit every bit of that.
| Item | Amount | Basis |
|---|---|---|
| My Claude Max use, electricity | About 300 to 650 kWh a month | Couch and Hausfather estimates, applied to a nearly maxed top plan |
| Average US home, electricity | About 900 kWh a month | Energy Information Administration, 2022 |
| My use, water at the data center | About 10 to 185 gallons a month | Amazon’s 0.12 liters per kWh to Google’s ratio of about 1 liter |
| My use, water at the power plants | About 360 to 780 gallons a month | Lawrence Berkeley National Laboratory, 2023 |
| My use, as gasoline energy | 9 to 19 gallons a month | EPA’s 33.7 kWh per gallon |
| My use, carbon dioxide | About 240 to 530 pounds a month | EIA, 0.81 pounds per kWh, 2023 |
| Average American family, water at home | More than 300 gallons a day | EPA WaterSense |
| One AI paragraph | About 0.24 watt hours and five drops of water | Google’s median Gemini text prompt, 2025 |
What does heavy AI use look like on a home electric bill?
By the Energy Information Administration’s count, a typical American household used 10,791 kilowatt hours in 2022. Divide by twelve and you get about 900 a month.
My Claude use comes to somewhere between a third and three-quarters of that.
So a month of the heaviest AI use one person can buy on a flat plan draws roughly the power of a small second household. It’s a real amount of electricity. It’s also the same order of size as a lot of things people do without a second thought, like running central air through a Florida summer or keeping an old freezer humming in the garage.
What heavy AI use costs in water
My whole month of maxed-out AI comes to between one and three days of one family’s water at home.
Water gets used in two places.
The data center uses some to cool its machines, and the power plants that make the electricity use a good deal more.
For the data center side, it helps to know whose buildings Claude runs in. Anthropic doesn’t own its data centers. It rents computing from Amazon Web Services, its main cloud partner, and from Google’s TPU chips, so their cooling numbers decide my share.
Amazon reports that its data centers withdrew 0.12 liters of water per kilowatt hour in 2025, with a power usage effectiveness of 1.14. Google is thirstier on paper. In August 2025 it put its median Gemini text prompt at 0.24 watt hours of electricity, with 0.26 milliliters of water to go with it. That works out to a bit over a liter per kilowatt hour.
Those two ratios make the low and high ends.
At my usage, the data center water comes to somewhere between 10 and 185 gallons a month, depending on which cloud does the work and where.
The power plants carry most of the load. Lawrence Berkeley National Laboratory counted 17.4 billion gallons of water used directly by US data centers in 2023, and about 211 billion gallons more at the plants that powered them, against 176 terawatt hours of electricity. Divide one by the other and you get about 1.2 gallons per kilowatt hour at the plant. For me, that’s another 360 to 780 gallons a month.
Add the two together and a month of my heaviest use comes to roughly 370 to 970 gallons of water, or 12 to 32 gallons a day.
The EPA says the average American family uses more than 300 gallons of water a day at home. My whole month of maxed-out AI comes to between one and three days of one family’s water at home.
How does a month of heavy AI use compare with a tank of gas?
Kilowatt hours are hard to picture, so I converted mine into something everybody buys.
When the EPA rates electric cars, it treats 33.7 kilowatt hours as the energy in one gallon of gasoline. By that yardstick, my 300 to 650 kilowatt hours hold the energy of 9 to 19 gallons of gas. Call it one fill-up of a midsize car, at most.
Put the same electricity into an electric car rated at 100 MPGe, about three miles per kilowatt hour, and it would carry you roughly 900 to 1,950 miles.
Carbon is the next question. The Energy Information Administration puts American power plants at about 0.81 pounds of carbon dioxide per kilowatt hour in 2023. At that rate my month comes to roughly 240 to 530 pounds. Gasoline at the pump releases 18.73 pounds of carbon dioxide per gallon, so a month of my heaviest AI use carries about the same carbon as burning 13 to 28 gallons of gas.
The grid behind any particular data center is cleaner or dirtier than that national average, depending on where it sits and what it buys. The average is the fairest stand-in I have.
What does a casual AI user’s footprint look like?
Most people are nowhere near me.
Somebody who asks a chatbot 20 questions a day, at Google’s quarter of a watt hour each, draws about five watt hours a day. Over a month that’s around 0.15 kilowatt hours. A typical electric dryer uses about 3 kilowatt hours per load, so the casual user’s whole month of chatting comes to about a twentieth of one load of laundry.
Subscribers sit in between. My guess is that most people with a paid plan use 20 or 30 percent of their allowance. A Max subscriber at that level lands somewhere around 60 to 200 kilowatt hours a month by the same math, a slice of one household’s bill.
If you’re in that first group, your whole month of chatting uses about as much power as three minutes of a clothes dryer.
How much energy do AI images and video use?
Everything above is text and agent work. Pictures and video are a different scale, and a lot of people who complain about AI are really complaining about the image generators.
MIT Technology Review measured open models it could run itself. One 1024 by 1024 image from Stable Diffusion 3 Medium took about 2,282 joules, roughly 0.63 watt hours. That’s about two and a half text prompts’ worth, or a couple of seconds of a microwave.
Video is where the numbers jump. A five-second clip from the CogVideoX model took about 700 times the energy of that image, around 0.44 kilowatt hours. Seven of those clips use about as much electricity as a load in the dryer, and a full minute of generated video would be roughly a dozen clips.
I make the featured images for my articles in batches with an AI image generator. A batch of ten at the Stable Diffusion rate comes to about six watt hours, a few seconds of that dryer. Commercial generators don’t publish their numbers and may run bigger models, so treat that as a floor.
The water follows the electricity here, the same as with text.
Images barely register.
Heavy video generation is the one everyday AI use that can add up as fast as an agent, and it deserves more of the scrutiny than the paragraph does.
Does one AI paragraph waste tons of water?
No, and the claim isn’t even close.
Each paragraph is one prompt. By Google’s figures, a prompt draws less than a quarter of a watt hour and about five drops of water at the data center. Throw in the water at the power plant and it’s a little over a milliliter.
A metric ton of water is 1,000 liters. You’d need somewhere around 750,000 paragraphs to use that much, counting the power plants, and close to four million counting the data center alone. Whoever told me my paragraph cost tons of water was off by a factor in the hundreds of thousands.
One AI paragraph uses about five drops of water. A ton of water would take roughly 750,000 of them. – Richard LoweShare on X
The famous bottle-of-water claim has its own history, and I traced where it came from and why it fell apart in my piece on whether AI data centers are bad for the environment.
Where the water complaint has a point
Heavy agent use is a different animal from a chat. Couch’s median Claude Code session came to about 41 watt hours, roughly 138 times a typical chat question. An agent reads whole folders of files, runs tools, checks its own work and tries again, and every one of those steps draws power. People who work the way I do are the reason the totals aren’t trivial.
The bigger issue sits above any one user.
My few hundred kilowatt hours a month mean nothing to the grid. Millions of heavy users, served out of a handful of data center towns, mean a great deal to the people who live next to those buildings and pay those utility rates. The backlash against data centers in those towns is real, even with foreign bot networks stirring it up.
None of my numbers include training the models in the first place.
That cost is spread across every user, and nobody outside the companies knows what my share of it is.
Each prompt keeps getting cheaper
The energy behind each request is falling fast. Google said the energy behind a median Gemini text prompt fell by a factor of 33 in a single year. Chips get better, models get tuned to do the same job with less computation, and providers stop redoing work they’ve already done.
That last part matters most for people like me. An agent rereads the same project files again and again as it works. Providers store that context so it doesn’t get processed from scratch each time, and both Couch and Hausfather assume that rereading cached context takes about a tenth of the energy of reading it fresh. In Hausfather’s logs, 96 percent of the tokens were those cheap rereads.
Today’s numbers will probably look high a year from now.
Does getting work done faster with AI save resources overall?
Per piece of work, almost certainly. In total, almost never. Both answers are true at once, and the argument usually gets stuck because people only hear one of them.
Start with the piece of work. An article, a plugin or a small browser tool that I build with AI in an hour would take a person, or a small team, days the old way. Those days come with their own computers, lights, air conditioning, commutes and meetings. Spread my AI electricity across everything it helps produce, and each item comes out cheaper than the version made without it.
The savings get big when AI replaces something physical. One 30-mile round trip in a car burns about 1.2 gallons of gas. By the EPA yardstick, that’s roughly 40 kilowatt hours of energy. That’s two to four days of my heaviest AI use, gone with one meeting I didn’t have to drive to.
There’s research on this side too. A 2024 study in Scientific Reports, from a team at UC Irvine, put AI’s carbon for a page of writing 130 to 1,500 times lower than a person writing the same page.
Now the catch. That study charged a slice of the writer’s whole life to the writing time, as if the person stopped existing while the machine worked. I don’t stop driving, eating or cooling my house when Claude writes a draft. Saved time saves resources only when it replaces something that burns them. My own desk is a small example: a desktop computer draws a few kilowatt hours on a long day, a fraction of what my AI use draws.
Then there’s the rebound effect. Economists have known about it since coal. When work gets cheaper, people do more of it. The authors of that study said themselves that they left rebound effects out.
I’m a textbook case. I didn’t use AI to do my old workload with less. I fill the allowance every single week.
So my footprint per piece of work went down, and my total went up. Anyone who tells you AI pays for itself in saved resources is skipping the second half, and anyone who says it’s pure waste is skipping the first.
What about chips, land and the rest?
Electricity and water are only the running costs. The machines themselves take mining, factories and energy to build before a single prompt gets answered. Researchers writing in Nature Computational Science in 2024 estimated that generative AI could leave behind 1.2 to 5 million tons of discarded electronics over the decade ending in 2030, because servers get swapped out every few years.
Land, buildings, transmission lines and grid capacity are real costs too. They mostly land on the counties that host data centers. That’s why the local fights get so heated.
All of it is real, and nobody publishes a per-user share of any of it, so I won’t invent one.
Can AI’s water and power problems be solved?
Most of them already have working answers. The open question is how fast companies adopt them, and whether anyone makes them.
Water-free cooling. Microsoft says every new data center design it has drawn since August 2024 cools the chips with a closed loop that reuses the same water. By the company’s count, each site saves more than 125 million liters a year. The trade is a bit more electricity for chillers. As the power gets cleaner, that trade matters less.
Cleaner power. Since most of the water sits at the power plant, clean electricity fixes much of the water problem along with the carbon. Microsoft signed a 20-year deal to restart a reactor at Three Mile Island. Google tracks its carbon-free power hour by hour on the same grid, and it reached 66 percent worldwide in 2024. Wind, solar and nuclear plants use little or no cooling water.
Rules that make data centers carry their own load. Ireland now requires a new data center to bring on-site generation or batteries big enough for its full demand and to buy at least 80 percent of its yearly power from new renewable projects. Texas passed a law in 2025 making very large customers help pay for their own grid hookups, so the neighbors don’t foot that bill.
Efficiency. The newest big cloud facilities waste far less power on cooling than the older server rooms they replace, and the energy per prompt keeps dropping, as Google’s numbers show.
Hardware reuse. The same Nature study found that reusing and recycling servers could cut that electronic waste by 16 to 86 percent.
The rebound effect is the one problem none of this makes disappear. People will keep using more AI as it gets cheaper. Clean power and closed-loop cooling are what make that growth matter less, because more use of clean electricity and recycled water costs the planet far less than more use of coal and evaporated rivers.
My take
Telling someone their one paragraph wasted tons of water is wrong by a factor in the hundreds of thousands. My own use is about as heavy as one person’s gets on a flat plan, roughly half a home’s electricity and a few days of a family’s water each month. Per piece of work, it’s efficient. In total, it’s a lot, and the fixes already exist. Push data centers onto clean power and closed-loop cooling, and check a scary number before you repeat it.
Read your own meter
If AI’s footprint worries you, run your own numbers the way you’d read the meter on your house. Look at how much you use, multiply by the best public figures, and set the result next to your dryer, your car and your sprinklers. Most people will find their AI use is a rounding error on the bill.
Heavy users like me will find something closer to a second household. That’s still a long way from a drained lake.
Why a flat plan can carry users like me without losing money is its own question, and I answered it in the gym membership math behind AI subscriptions. For help building AI into your own work, my AI services page explains how I do it, and the rest of my writing on the subject lives in the AI and writing hub.
The meter on the side of the house doesn’t care how loud the argument gets. It just counts.
Once you’ve decided AI is worth what it costs, in money or in power, the next question is what goes right when you put it to serious work and keep your own judgment in charge.
