Why can't data centers recycle their water?

Factual Questions
41

Do people need to be near the data center often? IOW, are on-site staff a constant requirement, or could it be a now-and-then thing? Do data centers need to be built close to “where the workers are”, or can they be placed somewhere with cold climates that are hard for humans to reach with frequency – for instance: remote Alaska, Northern Canada, Greenland. Perhaps Russia and China (in a very different political future) could place data centers in Siberia.

Maybe with future tech – Antarctic data centers. Build enough of them down there, and maybe the ice continent sort of becomes terraformed into a climate like modern Scandinavia (that wouldn’t have any follow-on effects anywhere else on Earth, no siree).

42

It’s already being done where practical.

The issue then becomes access to significant amounts of electrical power.

One of the issues with building out in some place like Siberia or Antarctica is you aren’t just building the data center, you are also building the plant to power it. Or investing in long distance power transmission and maintenance. That’s less of an issue for places in Canada or Iceland or Scandinavia.

But another testbed has been sealed undersea data centers. A hundred meters down provides a pretty good amount of passive cooling, as you might imagine. You still need to place them near a city or some source of power, though. And maintenance becomes much more difficult.

43

Can data centers use reclaimed wastewater?

44

There’s also the need for access to the internet, without which data centers would be useless.

45

In principle, yes, but in practice, no. The efficiency of a heat engine depends on the difference in temperature between the heat source and the cold reservoir. Actual power plants use temperatures of thousands of degrees. Hot CPUs are maybe 100 degrees C, not much hotter than the environment.

46

Some datacenters do use fully closed loop systems, but it costs more:

https://www.oracle.com/news/announcement/blog/closed-loop-cooling-in-oracle-ai-data-centers-2026-02-09/

47

I’m trying to keep this FQ, but may step over the line.

This is actually going on right now in my city of Colorado Springs - and there’s a lot of concerns that current claims are not going to hold up down the line.

https://www.cpr.org/2026/04/09/colorado-springs-proposed-data-center-backlash-public-meeting/

Green said the facility, as planned, would use a closed-loop water system that would initially require 200,000 gallons of water, around the amount of water for four residential pools. It would recirculate that water without needing more.

“200,000 gallons, day one, one time only. We don’t have to replace anything else. We don’t need any more water. We don’t want any more water,” he said.

The trade-off, he said, is the center will need to use more power to cool the recirculating water.

There was a great deal of concern about the whole “as planned thing”, and a huge secondary concern about the tradeoff.

Colorado Springs just had a near-record setting low snowfall over the winter, and has been in and out of drought conditions for the better part of a decade now.

So, sure, if completed as planned, needing minimal new water input is a big help. But… that last line I quoted has all the locals going “say WHAT?!”

Because that’s the increased power demand. Our utilities have been increasing above the national average cited in the article - 42% increase in the national average since 2019.

And starting this year, they forced everyone to a “peak” vs “off-peak” rating system which they promised was cost neutral, to deal with ongoing challenges. And on June 1st, despite doing that, they announced they were DOUBLING the peak rate all summer because of challenges to managing the summer grid demand.

So, many/most of us are looking at CSU (Colorado Springs Utilities) claims of being totally ready for the data center’s demand for power being ‘totes okay!’ with utter disbelief. And the cynical among us are betting that the recycled water use plans are going to go out the window to “address power grid concerns”.

I’ll close with another section from near the end of the article:

Green said the proposed data center would have “no tangible accountability” to the public, but he told residents that he would give them his personal number to call if they had any issues.

Under the proposal, Raeden is responsible for retrofitting the building, then the company would rent it out to an AI company. Green said because of non-disclosure agreements, he couldn’t share which company or companies would end up occupying the space.

So few-to-no details because of NDAs, no accountability, and complicit city government and utilities.

48

never mind, off topic.

49

We already have that with current computers. You can measure performance in calculations per second, but you can also measure performance in calculations per watt (with lower watts resulting in lower temperature). And modern computers produce multiple orders of magnitude more calculations per watt than old computers. A data center with the capacity of a modern data center but using Cray 1 computers would probably take a Kardashev Type I civilization to power it.

50

Aside from buffering, I assume the used water still has to go through the water treatment system before getting added to the water main supply, right?

Or are there places that would allow used, untreated water directly back in the supply (maybe with some oversight)?

51

You’d at least need some sort of testing on the outflow to verify that it’s not contaminated.

But downstream customers probably wouldn’t appreciate turning on their cold-water spigot and getting hot water out.

52

Which is really pretty stupid, because once the data center is built, it requires very few people to keep it running. So the local governments are making long-term concessions for a very temporary burst of jobs.

53

Heat and power problems have been part of high performance computing from the beginning. In a previous life I looked after a university supercomputer facility. Even back then the power and heat problems were a constant worry. (Especially if the university realised just how much power they were paying for to keep the supercomputers alive.)

You can make semiconductor systems dissipate a lot less power by running them slower. It doesn’t take that much of a slowdown of the clock to drop the dissipation significantly. Your smart phone isn’t running its processor silly hot. However more modern chip processes have larger intrinsic gate leakage, so the baseline dissipation has been slowly increasing and the gains diminish.

Use slower but more efficient processors, your capital costs go up for the same performance. In a market where there is already a scarcity of every component, this isn’t going to happen. Everyone is running headlong into the great AI gold rush. Whether anyone comes out the other side and pays off the investment is another matter.

In the extreme we have the tech bros talking up AI data centres in space. Radiate the excess heat out into the cold dark. From a technical point of view it can be done. Again, whether there are enough customers for this as yet unbuilt capacity is a very open question.

My bet is that many of these data centres will never be built, and there is going to be a lot of money lost. If you are not already making money with your AI offerings, there is a good chance you never will.

54

Yes, but at a much greater cost than even the most-regulated terrestrial data centers. Even if launches were free, it’d still be a much greater cost.

55

Back into what system? The water input to a datacenter is a drinking water supply line. The only output is sewer. 5 million gallons of water a day is a lot of extra capacity to put through a sewer and water treatment system. Depending on how nearby the treatment plan is, the elevated temperature could mess up the chemical processes in the wastewater treatment process.

The right solution would capture hot waste water and use it for some municipal benefit like heating, but that won’t happen in America because we’re dumb about that kind of planning, and a lot of the datacenters are being built in low-regulation states where the climate is too hot to make much of industrial thermal waste water, and nobody really gives a shit about negative externalities anyway.

It’s best to think of a data center like any other kind of factory that also boils off 5 million gallons of water a day. Soda bottler, paper mill, semiconductor fab, take your pick. It just consumes a lot of power and water to produce valuable data services and heat, nowhere near the nasty stuff that other factories would produce.

The environmental objections to data centers are really just a few other disguised complaints - first, people don’t like AI, second, companies are throwing up big boxy structures on what used to be bucolic landscapes, and 3, the part people should actually be getting mad about, is they’re often getting outrageous tax abatements that get them off the hook for paying for the externalities that they actually do generate.

56

In my town, the waste heat from our crematorium is used to heat the municipal swimming pool, which is a couple of miles away.

57

From the water main, through the brand new clean factory pipes, through a filter, into a water tower, back into the water main.

58

Some numbers might help focus the thoughts.

  • Specific heat of water is 4182 J/(kg·K)
  • Latent heat of evaporation is 2257 kJ/kg

The first problem one will have with cooling a data centre is that the heat is relatively low grade. There are efficiencies if you can keep the chip cooler - they don’t thermally throttle, and you get an an overall gain in compute for the power - so there is incentive not to let them get hot. Package temperature is not going to be more than 50C. So you won’t heat the coolant higher than this.

If you ingest water at say 20C, and heat it to 50C, that is a 30K gap, for 125kJ per kg of water. If you evaporate one kg of water, you need 2257 kJ. So twenty times the energy/cooling of just heating the water a 30K delta. Or another way, you circulate the water twenty times through a cooling tower before it is all gone. (20C input may be too low, higher input temperature and higher flow rates may be more efficient.)

The water running through the compute racks won’t be the same water as in the cooling tower. There is going to be a heat exchanger in between. The water/coolant in the main cooling loops is going to be a lot more carefully curated and managed than just plain municipal tap water. There is no reason that municipal water could not safely flow through the heat exchanger. The problem is how much you would need to flow, and how hot you are prepared to put up with it becoming. Also, you need a buffer. Data centres run 7x24. Where I live, even the normal tap water is too hot during the summer, without some AI data-centre heating it up even more.

59

When a data center says “closed loop”, they are of course sending the same water back in again. So then, like your car’s engine, you need a way to cool it before sending it back. For a car, you have a few square feet of radiator and blast ambient air through it at whatever speed the car is going, aided by a fan. So missing in discussions about “closed loops” is what means to dispose of the heat before the water returns to cool the computers again. As mentioned, water evaporation is the most efficient. ambient water, from the ocean or river, is second best but with environmental implications. Forced air requires massive “force” (no doubt noisy fans) and a very large radiator surface to accomplish the same, and the fans require even more electricity.

Many bigger cities back in the early 1900’s had central steam heat, more efficient for large blocks of buildings. Steam rising from manholes was an iconic New York thing, that I recall famously to add mood in movies such as Taxi Driver. However, newer buildings got away from using central steam and many cities progressively shut it down, which doomed grand old buildings where installing their own steam heating would be too expensive and adapting to forced air impossible.

60

Yes this was said, but are there any stopgaps/penalties in place just in case a “Whoopsie! It looks like our original plan didn’t work and we are going to have to use more water than projected-Our bad!” situation happens?