The key issue to always remember with water is that averages and totals are almost meaningless in assessing problems. Some watersheds have lots of water, others very little; some watersheds have lots of water users, some very few. You can transfer water from one watershed to another, but it's an expensive process. So even if water use by data centers were a small part of the overall total, it may very well be critical in specific watersheds, and contrariwise there may be watersheds where data centers account for most water use and yet it does not matter because there are few competing uses in that watershed. (Likewise, the often cited numbers showing that agriculture is the largest water user is less meaningful than it appears, as a non-negligible portion of that use is in watersheds where it does not compete with other uses, and could not easily be transferred to other watersheds where it would be more valuable.)
ona (and the Southwest), where water, all of it, is precious and getting more scarce. There is enough ground water and lake water in Michigan and nearby states for 1000x more data centers.
Information transmission is ultra low energy (fiber optics).
Is the corollary to this that the only meaningful estimate of data center impact will require a team of scientists to construct and estimate a complicated dynamical systems model? Or do you see some value, as a first order approximation, in crude measures like *amount of water used by such-and-such purpose*?
Water in Arizona is so precious that Arizona has 400 golf courses. Arizona's golf courses alone use more water than the entire global data center business. Literally just dumping it on the ground.
Some do. Even the "Arizona Alliance For Golf" who obviously have an incentive to deemphasize this aspect of the business says they use 75 million gallons per day of fresh and 47 million gallons of reclaimed water to irrigate.
It's still a consumptive use since most ends up evaporating. Clean it a bit more and humans could easily use that water instead (and do in many places, like Israel, Orange County, CA, parts of Arizona, etc.; more if you consider that treated wastewater dumped into a river is often just pulled out again by the next city downstream).
I didn't read the original article. I think it might've been nice in this article if you gave some other numbers for context like how much water farms and households and such use. Because I've got no idea if 600 million gallons a day is actually a big deal or not. But that might be just me.
I give these figures in the original article, but for comparison golf courses in the US use around a billion gallons of water a day, and cotton production uses around 2.5 billion gallons of water a day.
Interesting! Thanks for clarifying. The absence of a singular method to incorporate evaporation seems a poor excuse to simply include it all... For example, a multiplication of the % of downstream use during the driest season, or some other measured annual maximum, would be a good step while remaining very conservative. With the current waiting line for grid connections, not acknowledging PPAs is more understandable
The hard upper limit for consumption use in the data center should be the latent heat of evaporation equaling the power input to the facility (I'm assuming ambientl heat is low density enough to be discounted), or ~0.63L/kWh (2260kJ/kg evaporated).
If using evaporative / recycling cooling, *total* should be around 1L/kWh to include blow down, but the remainder is non-consumptive.
To generate that power from coal or nuclear, it should be around 2X that much on the generation side from thermodynamic considerations.
600 million gallons a day is still pretty small potatoes. That's only around 700,000 acre feet a year. The US users around 370 million acre feet a year, so this is something like 0.2% of water usage. CA uses around 40 million acre feet a year. Any use of water outside of agriculture, landscaping, and golf courses is basically a rounding error.
The “engineer” in me is interested in the “downstream” consumption- used for cooling the hardware in the centers. If these are closed loop cooling systems and the heat transfer to the water cooling is released into the atmosphere (as in a cooling tower), then water consumption would only be for makeup due to leaks in the loop. A minimal amount. Electricity would be needed for circulation pumps and cooling fans though. I’m simplifying this a bit, but a closed loop is far less resource consumptive than an open loop.
What you are describing is dry air cooling as opposed to evaporative cooling (which can also use a closed loop circuit for the actual internal use, which confuses a lot of people).
The issue is that dry air cooling performance is limited by the ambient *dry* bulb temperature rather that the wet bulb temperature, which limits for evaporative cooling. They also need *far* more contact area than evaporative, and are a lot more expensive and use ~2X+ the electricity (mostly for fans)
The issue with dry bulb temperature is that in the areas with water stress this is usually to high to be that useful during peak heat weather.
"Amazon, Meta, Google, and Microsoft all report that 100% of their electricity comes from renewable sources."
This is complete rubbish and the most cynical sort of greenwashing by voracious corporations. Electrons flow according to physics, not contracts data centers might have with PE owned solar, PV or hydro. Most if these monstrosities are built in close proximity to fossil fueled generation. Whatever sorts of contracts the Magnificent 7 might have for renewables from distant sources, doesn't change the fact their obscene consumption of electricity and water displaces other consumers. Some reality checks:
Meta's planned Louisiana AI data center is going to be supplied by Entergy with 2+GW of gas fired generation that is folded into the consumer rate base.
Largest US gas-fired power plant planned for data centers in Pennsylvania (Utility dive)
The 4.5-GW, $10 billion project at the former Homer City coal-fired power plant could start operating by 2027, according to the project’s developers.
In NY State, governor Hochul believes int SMR nuclear fairies will rescue the state with its all-electric future and massive loads from the Micron and Global Foundaries chip fabs and data centers. One can only imagine what residential rates become in the mid term future.
I don't see any comparison against other water usage, such as waste or runoff from inefficient agricultural methods. Without that, I don't think we have an accurate picture.
"Amazon, Meta, Google, and Microsoft all report that 100% of their electricity comes from renewable sources."
Hm, I'm not sure that's exactly what they say. They say they "matched" their energy consumption with purchases of 100% renewable energy. Not the same thing.
The thing is, PPAs don’t guarantee any kind of local production nor time of the day for production and therefore you could argue that the power plants closest to the consumer should be taken into account. That would be the most accurate « physical » representation of the system. If we start to count PPAs and CCS it will all be compensated and the real environmental cost will be hidden behind contracts and such that makes you feel like these datacenters don’t have a huge environmental impact.
Maybe PPAs are different in the US though.
Water consumption, carbon footprint, electricity consumption may suffer the same fate as money, you can always « optimize » when it does not look good…
The key issue to always remember with water is that averages and totals are almost meaningless in assessing problems. Some watersheds have lots of water, others very little; some watersheds have lots of water users, some very few. You can transfer water from one watershed to another, but it's an expensive process. So even if water use by data centers were a small part of the overall total, it may very well be critical in specific watersheds, and contrariwise there may be watersheds where data centers account for most water use and yet it does not matter because there are few competing uses in that watershed. (Likewise, the often cited numbers showing that agriculture is the largest water user is less meaningful than it appears, as a non-negligible portion of that use is in watersheds where it does not compete with other uses, and could not easily be transferred to other watersheds where it would be more valuable.)
Great point!
We don't want these data centers here in Ariz
ona (and the Southwest), where water, all of it, is precious and getting more scarce. There is enough ground water and lake water in Michigan and nearby states for 1000x more data centers.
Information transmission is ultra low energy (fiber optics).
Move photons! not water or electrons
The nuclear power plants in AZ also used reclaimed sewage water for cooling.
This! I don't care about water usage in very wet areas. But growing cotton and corn in the Arizona desert is insane.
Which of those two gets the most attention, though?
Is the corollary to this that the only meaningful estimate of data center impact will require a team of scientists to construct and estimate a complicated dynamical systems model? Or do you see some value, as a first order approximation, in crude measures like *amount of water used by such-and-such purpose*?
Water in Arizona is so precious that Arizona has 400 golf courses. Arizona's golf courses alone use more water than the entire global data center business. Literally just dumping it on the ground.
Golf courses used “reclaimed water” in the Phoenix area. They are using the water coming from waste treatment facilities.
Some do. Even the "Arizona Alliance For Golf" who obviously have an incentive to deemphasize this aspect of the business says they use 75 million gallons per day of fresh and 47 million gallons of reclaimed water to irrigate.
It's still a consumptive use since most ends up evaporating. Clean it a bit more and humans could easily use that water instead (and do in many places, like Israel, Orange County, CA, parts of Arizona, etc.; more if you consider that treated wastewater dumped into a river is often just pulled out again by the next city downstream).
Very interesting article. Thanks for writing it. The one nit to pick is that the Grand Coulee is on the Columbia.
I didn't read the original article. I think it might've been nice in this article if you gave some other numbers for context like how much water farms and households and such use. Because I've got no idea if 600 million gallons a day is actually a big deal or not. But that might be just me.
I give these figures in the original article, but for comparison golf courses in the US use around a billion gallons of water a day, and cotton production uses around 2.5 billion gallons of water a day.
Interesting! Thanks for clarifying. The absence of a singular method to incorporate evaporation seems a poor excuse to simply include it all... For example, a multiplication of the % of downstream use during the driest season, or some other measured annual maximum, would be a good step while remaining very conservative. With the current waiting line for grid connections, not acknowledging PPAs is more understandable
The hard upper limit for consumption use in the data center should be the latent heat of evaporation equaling the power input to the facility (I'm assuming ambientl heat is low density enough to be discounted), or ~0.63L/kWh (2260kJ/kg evaporated).
If using evaporative / recycling cooling, *total* should be around 1L/kWh to include blow down, but the remainder is non-consumptive.
To generate that power from coal or nuclear, it should be around 2X that much on the generation side from thermodynamic considerations.
600 million gallons a day is still pretty small potatoes. That's only around 700,000 acre feet a year. The US users around 370 million acre feet a year, so this is something like 0.2% of water usage. CA uses around 40 million acre feet a year. Any use of water outside of agriculture, landscaping, and golf courses is basically a rounding error.
The “engineer” in me is interested in the “downstream” consumption- used for cooling the hardware in the centers. If these are closed loop cooling systems and the heat transfer to the water cooling is released into the atmosphere (as in a cooling tower), then water consumption would only be for makeup due to leaks in the loop. A minimal amount. Electricity would be needed for circulation pumps and cooling fans though. I’m simplifying this a bit, but a closed loop is far less resource consumptive than an open loop.
The cooling towers mostly use evaporation, since it is far denser to build than heat transfer to air. The losses are not "leaks", they are by design.
They could also return it to the river, but I doubt that is allowed.
What you are describing is dry air cooling as opposed to evaporative cooling (which can also use a closed loop circuit for the actual internal use, which confuses a lot of people).
The issue is that dry air cooling performance is limited by the ambient *dry* bulb temperature rather that the wet bulb temperature, which limits for evaporative cooling. They also need *far* more contact area than evaporative, and are a lot more expensive and use ~2X+ the electricity (mostly for fans)
The issue with dry bulb temperature is that in the areas with water stress this is usually to high to be that useful during peak heat weather.
Why is conversion of water from liquid to gaseous form considered “consumption”?
What is considered to be “production” of water?
Don’t forget to promote your book!!
"Amazon, Meta, Google, and Microsoft all report that 100% of their electricity comes from renewable sources."
This is complete rubbish and the most cynical sort of greenwashing by voracious corporations. Electrons flow according to physics, not contracts data centers might have with PE owned solar, PV or hydro. Most if these monstrosities are built in close proximity to fossil fueled generation. Whatever sorts of contracts the Magnificent 7 might have for renewables from distant sources, doesn't change the fact their obscene consumption of electricity and water displaces other consumers. Some reality checks:
Meta's planned Louisiana AI data center is going to be supplied by Entergy with 2+GW of gas fired generation that is folded into the consumer rate base.
Largest US gas-fired power plant planned for data centers in Pennsylvania (Utility dive)
The 4.5-GW, $10 billion project at the former Homer City coal-fired power plant could start operating by 2027, according to the project’s developers.
In NY State, governor Hochul believes int SMR nuclear fairies will rescue the state with its all-electric future and massive loads from the Micron and Global Foundaries chip fabs and data centers. One can only imagine what residential rates become in the mid term future.
And then there the centers along the Columbia getting power from the dams.
It seems to me the huge data centers are going to be dinosaurs soon, due to advancements in fractal computing.
I don't see any comparison against other water usage, such as waste or runoff from inefficient agricultural methods. Without that, I don't think we have an accurate picture.
"Amazon, Meta, Google, and Microsoft all report that 100% of their electricity comes from renewable sources."
Hm, I'm not sure that's exactly what they say. They say they "matched" their energy consumption with purchases of 100% renewable energy. Not the same thing.
I love it when someone is honest enough to admit they made a mistake. Thanks for your integrity! Love your posts!
The thing is, PPAs don’t guarantee any kind of local production nor time of the day for production and therefore you could argue that the power plants closest to the consumer should be taken into account. That would be the most accurate « physical » representation of the system. If we start to count PPAs and CCS it will all be compensated and the real environmental cost will be hidden behind contracts and such that makes you feel like these datacenters don’t have a huge environmental impact.
Maybe PPAs are different in the US though.
Water consumption, carbon footprint, electricity consumption may suffer the same fate as money, you can always « optimize » when it does not look good…