I wish this were surprising to me, but AI news reporting has come up with so many outright bizarre "as much power as the Eiffel Tower in 17 years" and "all the factories in Nepal in a month" type analogies that I've become extremely cynical on this topic. Or training electricity expressed as some apparently staggering number of households... that turns out to be equal to less than 1% of annual global population growth, which is furthermore cumulative.
Working journalists by and large don't like generative AI, don't like big tech, and are overly trusting of sources critical of both. At the same time, they are highly inclined to be convinced by narratives, particularly those that cast them in the role of the hero. This doesn't require any malice, just a desire to tell the story that they feel must be told:
"As the dark side of AI was gradually exposed, and the stochastic parrots fell far short of their promise, the bubble burst and people gradually swore off AI. After calls from artists, writers, medical professionals and environmentalists that Something Must Be Done, [some thing] was done, and then everything went back to how it was and always will be, and the news cycle moved on."
Some of the more "big picture" articles in recent weeks seem to express bafflement that this just isn't playing out and people are still using ChatGPT. Go figure.
That was an interesting read, but for the 2nd assumption stating that total data center water usage is 67 million gallons a day, I was looking at the LBNL report Brian Potter cited for that 67 million gallons per day figure and you put in the further reading section, and 2 things come to mind. First, the water consumption data within the cited report is for 2023, not 2024.
Secondly, Potter's reduction of the 579 million gallons/day to the 19 million gallons/day attributed to indirect consumption (electricity generation) seems to assume that the first figure is the water withdrawal rate, not the water consumption rate, but the LBNL report explicitly defines water consumption as the following in the page before (56): "Water consumption refers to the amount of withdrawn water that is permanently removed from the immediate water cycle due to evaporation or other irreversible processes."
Looking at an alternate source for verification, from Meta's own sustainability report (https://sustainability.atmeta.com/wp-content/uploads/2024/08/Meta-2024-Sustainability-Report.pdf), for 2023 they report 14,975,435 MWh of electricity consumed by data centers, for 58,475 megaliters of water consumption from purchased electricity (99%+ of their electricity by source). That comes to 3.90 liters of water consumed per KWh, and Meta's own definition of water consumption is the difference between the water taken out from a source (withdrawal) and what is discharged back into the water. For reference, the LBNL report used an estimate of 4.52 liters of water consumed per KWh to get to their consumption figure.
The LBNL report states that total electricity usage for all datacenters was 176 TWh, so 176,000,000,000 KWh * 3.90 liters = 686,400,000,000 liters = 181,327,696,738.633 gallons / year. Dividing by 365 gets 496,788,210 gallons, or 496 million gallons of water per day. There are other water consumption figures I looked at ranging from around 2 gallons/7.5 litres per KWh to 0.52 gallons/2 liters per KWh, but even at the lowest figure, that's still more than 345 million gallons per day of water consumption, which exceeds the 19 million gallons per day indirect consumption figure by a significant margin.
There’s currently no agreement among states as to how to deal with a troubled Colorado River. I suppose Big Balls or some other qualified person could step in and help figure out the best way forward with the river usage. They’ve got until 2026.
I do have a question: why did Microsoft put up its data centers in Maricopa County given that it is drought-prone? What's driving the decision-making behind where to put these data centers?
Can look into it later, best guess is lots of other benefits specific to Arizona. So air cooling (instead of water cooling) costs more in energy, but if other costs are lower it balances out. Land is super affordable, lots of tax incentives from Arizona, low electricity costs, no really bad inclement weather (like hurricanes or floods or earthquakes) and maybe lots of abundant solar power options. Here's an article that seems to confirm: "These are boom times in central Arizona for data centers, particularly in Maricopa County, which is one of the country’s largest data center markets. Factors in its favor, according to Alan Howard, an industry analyst at Omdia, a research firm, include relatively cheap land, tax incentives, low power rates, and proximity to the computing needs of the 5 million people in the Phoenix metro area." https://waterdesk.org/2025/04/data-centers-a-small-but-growing-factor-in-arizonas-water-budget
Do you understand the difference between a private well drying up and municipal water service where costs are or are not rising? Maybe, but the piece was too long to read. However, including the NYT headline re: someone's well going dry and then talking about municipal water rates is a classic bait and switch
Nope sorry you should've read that section of the post, the wells didn't run dry because the data center was using the groundwater, they ran dry because of construction-related sediment buildup. This is a problem that could've happened with any large building. It's misleading to say that the data center "guzzled" the water, it wasn't using it at all (it wasn't even on yet!). I wouldn't have done anything like a bait and switch where I said something like "Haha water rates don't rise, wells just dry up but that's different!" and I'd appreciate if you actually bothered to at least skim the article before jumping to that conclusion. It's very hard to write concisely about AI water use because there are so many angles people are looking to jump on that if I don't include something people will treat the article as deceptive, as you've just done.
Your 4th reference says "AI data centers in America have a consumptive use of 67 Mgal/day." The linked text doesn't seem to make this distinction, though? Would it just be "data centers use 67" or is it actually AI-specific data centers using 67?
I wish this were surprising to me, but AI news reporting has come up with so many outright bizarre "as much power as the Eiffel Tower in 17 years" and "all the factories in Nepal in a month" type analogies that I've become extremely cynical on this topic. Or training electricity expressed as some apparently staggering number of households... that turns out to be equal to less than 1% of annual global population growth, which is furthermore cumulative.
Working journalists by and large don't like generative AI, don't like big tech, and are overly trusting of sources critical of both. At the same time, they are highly inclined to be convinced by narratives, particularly those that cast them in the role of the hero. This doesn't require any malice, just a desire to tell the story that they feel must be told:
"As the dark side of AI was gradually exposed, and the stochastic parrots fell far short of their promise, the bubble burst and people gradually swore off AI. After calls from artists, writers, medical professionals and environmentalists that Something Must Be Done, [some thing] was done, and then everything went back to how it was and always will be, and the news cycle moved on."
Some of the more "big picture" articles in recent weeks seem to express bafflement that this just isn't playing out and people are still using ChatGPT. Go figure.
That was an interesting read, but for the 2nd assumption stating that total data center water usage is 67 million gallons a day, I was looking at the LBNL report Brian Potter cited for that 67 million gallons per day figure and you put in the further reading section, and 2 things come to mind. First, the water consumption data within the cited report is for 2023, not 2024.
Secondly, Potter's reduction of the 579 million gallons/day to the 19 million gallons/day attributed to indirect consumption (electricity generation) seems to assume that the first figure is the water withdrawal rate, not the water consumption rate, but the LBNL report explicitly defines water consumption as the following in the page before (56): "Water consumption refers to the amount of withdrawn water that is permanently removed from the immediate water cycle due to evaporation or other irreversible processes."
Looking at an alternate source for verification, from Meta's own sustainability report (https://sustainability.atmeta.com/wp-content/uploads/2024/08/Meta-2024-Sustainability-Report.pdf), for 2023 they report 14,975,435 MWh of electricity consumed by data centers, for 58,475 megaliters of water consumption from purchased electricity (99%+ of their electricity by source). That comes to 3.90 liters of water consumed per KWh, and Meta's own definition of water consumption is the difference between the water taken out from a source (withdrawal) and what is discharged back into the water. For reference, the LBNL report used an estimate of 4.52 liters of water consumed per KWh to get to their consumption figure.
The LBNL report states that total electricity usage for all datacenters was 176 TWh, so 176,000,000,000 KWh * 3.90 liters = 686,400,000,000 liters = 181,327,696,738.633 gallons / year. Dividing by 365 gets 496,788,210 gallons, or 496 million gallons of water per day. There are other water consumption figures I looked at ranging from around 2 gallons/7.5 litres per KWh to 0.52 gallons/2 liters per KWh, but even at the lowest figure, that's still more than 345 million gallons per day of water consumption, which exceeds the 19 million gallons per day indirect consumption figure by a significant margin.
Excellent, thank you. Will circle back on this over the weekend. I'm extremely lucky to get such quality breakdowns as comments.
I've updated the post with a much higher number to be safe until I circle back
There’s currently no agreement among states as to how to deal with a troubled Colorado River. I suppose Big Balls or some other qualified person could step in and help figure out the best way forward with the river usage. They’ve got until 2026.
Good point, water management’s definitely far from perfect
Access will become untenable without management
I do have a question: why did Microsoft put up its data centers in Maricopa County given that it is drought-prone? What's driving the decision-making behind where to put these data centers?
Can look into it later, best guess is lots of other benefits specific to Arizona. So air cooling (instead of water cooling) costs more in energy, but if other costs are lower it balances out. Land is super affordable, lots of tax incentives from Arizona, low electricity costs, no really bad inclement weather (like hurricanes or floods or earthquakes) and maybe lots of abundant solar power options. Here's an article that seems to confirm: "These are boom times in central Arizona for data centers, particularly in Maricopa County, which is one of the country’s largest data center markets. Factors in its favor, according to Alan Howard, an industry analyst at Omdia, a research firm, include relatively cheap land, tax incentives, low power rates, and proximity to the computing needs of the 5 million people in the Phoenix metro area." https://waterdesk.org/2025/04/data-centers-a-small-but-growing-factor-in-arizonas-water-budget
Makes sense!
Thanks for writing this and for all the links!
Do you understand the difference between a private well drying up and municipal water service where costs are or are not rising? Maybe, but the piece was too long to read. However, including the NYT headline re: someone's well going dry and then talking about municipal water rates is a classic bait and switch
Nope sorry you should've read that section of the post, the wells didn't run dry because the data center was using the groundwater, they ran dry because of construction-related sediment buildup. This is a problem that could've happened with any large building. It's misleading to say that the data center "guzzled" the water, it wasn't using it at all (it wasn't even on yet!). I wouldn't have done anything like a bait and switch where I said something like "Haha water rates don't rise, wells just dry up but that's different!" and I'd appreciate if you actually bothered to at least skim the article before jumping to that conclusion. It's very hard to write concisely about AI water use because there are so many angles people are looking to jump on that if I don't include something people will treat the article as deceptive, as you've just done.
Good luck with the Colorado River….
Your 4th reference says "AI data centers in America have a consumptive use of 67 Mgal/day." The linked text doesn't seem to make this distinction, though? Would it just be "data centers use 67" or is it actually AI-specific data centers using 67?
Good catch, should flip to all data centers. Lots of claims and math in this one so there are probably other mistakes I'm missing. Thanks!
>A lot of journalism is misleading
FIFY
You want water issues? Look into water rights and water usage in Arizona. Using water to irrigate corn and cotton. True insanity.
Thanks for catching