the water use of datacenters I find to be a weird angle to take. does anything bad happen to water when you run it through a computer? surely you can just let it cool down and then reuse it?
i can speak with a little authority on this, i work in this field
DCs expend water by evaporation, either in cooling towers or evaporative coolers (water becomes vapor, taking latent heat energy with it into the air and leaving cooler liquid water behind)
so we expend liquid water resources 1/
we also lose some liquid water as wastewater through regular ol' drains for a couple of reasons which we consider to be inefficiencies and attempt to minimize, because thats water we pay for but didnt contribute to cooling.
2/
the wastewater is generally reclaimed by the water utility and resold, especially in regions with lots of DCs/other large industrial water users
we pay the utility both for providing water and for handling our waste, and we have monthly caps on both determined by our deal with the utility 3/
the water we take in is industrial/reclaimed/non-potable - basically treated greywater - because it's cheaper, and theres no reason to buy purer water.
how much this matters, from a policy/water resourse management/environmental management perspective depends entirely on the region.
4/
so its not drinking water, but it probably couldve been made into drinking water, or not drawn from water sources at all, or used for other industrial uses
in places with constrained water resources its perfectly reasonable to question our expenditure of water. they are very thirsty buildings.
5/
the amount lost as liquid to waste drains varies, but its a small fraction of the water intake. the supermajority of cooling water evaporates into the air, and is in fact lost to the local water economy, until the water cycle replenishes it.
6/
finally, to correct a misconception, the cooling water never passes through the computers (except as vapor in air). all racks in typical DCs are air cooled. this improves reliability, and lets us use outside air to cool the racks when its cold enough. my job is cooling air, not cooling racks 7/
cooling with outside air ('free cooling') is desirable because its cheaper, and is one factor in selecting locations to build datacenters (as is humidity, which determines how effective evaporative cooling can be)
in winter, or at night, datacenters often use zero water for cooling
8/
so, a rational policy on DC water use might be to make regulations and price controls that force operators to build in locations with very plentiful water resources, and where they are ulikely to compete with agriculture
9/
datacenters make loads of money, employ relatively few people given their physical and economic footprint, and drink lots of water and power, which implies other externalities depending on gridmix. it is worthwhile for communities to discuss their inputs compared to their local benefits 10/10
I'm going to bring up cold plates because the new MAIA accelerators Microsoft built to run OpenAI workloads use direct water pipes to the accelerator and they custom designed a nice pipe system they can add to existing DCs. They have immersion cooling with the 3M liquid but not for this workload
Would building DCs underground make any difference? I imagine the thermal mass of the surrounding earth could act as a giant โfreeโ heatsink, and would eliminate a lot of weather-related variables.
There are some locations in the US where the geology is good in existing caves but the ground is good insulation so usually it's not cold down there. Some countries have snowmelt that's pre cooled but it's not common either
to cool air, which cools the servers
in an evap-cooled dc (more common) outside air is cooled by evaporation (consuming water), and then passed through the datahall.
in a chiller DC, a refrigeration plant chills water, which cools the datahall air, and rejects waste heat to a cooling tower.
In '08 city of Denver decided to turn on water fountains on the downtown mall for the DNC convention.
The fountains hadn't been used in ages. Something gave out and flooded utility tunnels, shorting out the pump for our building chilled water service. Fun times in the DC. Too much water = no water.
Sure, it would be crazy expensive - but I think some folks misunderstand the word waste, that it's a loss to DC operations not that it vanishes from the planet!
i should have written, water vapor caries away latent heat, cooling both the liquid water and the air.
in a cooling tower we care about the cool water, and the air is discarded
in an evap cooler we care about the cool air, and the water is recirculated
One trick I learned recently is to keep a spray bottle full of water in my car. If my steering wheel is too hot to handle (parked in the sun), I spritz on some water and wait for it to evaporate. Repeat until evaporation slows down. Doesn't matter if the water is hot, the evaporation is the key.
Program manager here, originally a network engineer, who got involved in construction projects.
Built my first ground up corporate data center in the late 90s. We were proud that we had a full OC-3 of connectivity. (Thatโs ~155 Mbps, for the non-telco inclined.)
There are very similar systems in place for cooling skyscrapers. Water sensitive areas like Saudi Arabia use closed loop water systems to prevent wasting water. (Itโs too expensive)
Water used to be dirt cheap so there wasnโt a drive to preserve it.
Source: I am a Water Chiller engineer.
Evaporative cooling is not the only way datacentres are cooled. Seems to be the majority at present, but dropping, with many major players (such as Microsoft) designing future datacentres to be outright cooling-water-free. Google, Microsoft and Meta have all pledged to be net-water positive by 2030.
To pick a specific example: last year Apple withdrew 5,8B litres of water and and discharged just under half of it, 2,6B. Meta discharges about 1/3rd of theirs. It all depends on the specific cooling system design; you can't generalize.
sure, but theres principles in common, and the fact remains that the large extant DC fleet as well as the many now being built are some thirsty bitches.
im suspicious of claims towards being water neutral anytime soon, let alone "water positive"
while there are water-neutral options, like aircooled chillers, they have their own caveats like mechanical complexity and maintenance demands- evap systems are much simpler. theres a reason the modal datacenter is water-intensive, which is that water is great for cooling.
Water replacement is absolutely a legitimate strategy. Paying to build more efficient irrigation systems, fix leaks, cover aqueducts, etc etc etc (there's a million things you can do), you legitimately put more available water on the market.
im skeptical of 'offsetting' strategies generally, but i acknowledge thats my personal politics.
if i were in a role of stewardship over a watershed, I would be wary of these strategies being carried out dishonestly or causing other problems. but i'm not, and people who are should weigh them fairly
Ultimately data centres are usually air-cooled (though liquid cooling is gaining inroads), with water just used to create cool air. Omitting water generally increases capital and energy costs somewhat. From memory I think it's like a 10% increase in power consumption, though I could be off.
Governments are not unaware of the environmental limits of cooling water. Like power plants, once-through cooled data centers have thermal limits on their discharge. Depending on the weather (drought, heat waves, etc), they can sometimes be forced to throttle due to discharge thermal limits.
afaik, data centers tend to basically evaporate their water supply through non-closed systems until it gets excessively dirty, then they dump the waste water (presumably responsibly)
DC water use isnโt really water cooling like youโd see in a PC, but itโs an open system so the water gets reclaimed (or put back into the water cycle). evaporative coolers and such
yeah water use scaremongering I am always extremely sceptical about because itโs gonna be a drop in the bucket (ha) compared to industrial farming or w/e
thereโs a lot of interesting work happening now in terms of quantifying co2/water/whatever usage for cloud workloads in general (thereโs a whole sustainability TAG in the CNCF I try to keep up with)
I just find it difficult to gin up a lot of outrage about the impact of specific workloads. streaming is more intensive than โAIโ from an energy/co2 perspective but you donโt see people going on a jag about that
Yeah, been saying the same about render farms for like pixar and Disney movies. Those take an outrageous amount of energy and require cooling Iโm certain. yet no one seems to complain about that, or other equivalents even.
Not saying anything wrong with enjoying disney and pixar movies, I love them. But, I think that is kinda the point. Thatโs โGoodโ energy and resource use cause we like those, this other stuff, thatโs โBadโ use cause we donโt like what it is being used for personally. Seems a bit hypocritical.
Though I'd also expect those kind of companies to take their local impact more seriously and invest in closed loop systems to use recycled water. That's the real issue - how much water they're taking and the local impact of that over time.
I honestly have no idea how they cool their render farms, i just know they use a massive amount of energy and heat, but seeing how no one ever seems to have a problem about their systems, maybe there is something different **shrug**
I'm curious about this. How does one set up a 1:1 comparison of these two tech uses? Like, the processor for AI, vs the data center for streaming? And then take an equivalent fraction of the workload for each to come up with a number that corresponds to one use? I don't know what I'm talking about!
It would be pretty difficult to do from the outside because we donโt know every aspect of the networking/hardware stack companies are using to deliver bits to end users. Netflix, for example, often serves popular content from a local ISP while long-tail content is delivered from further upstream.
No soy experto pero supongo que para el streaming obtienes un archivo de un almacรฉn o de una cรกmara y lo envรญas para su reproducciรณn, no debe requerir mucho procesamiento, es casi todo transmisiรณn.
La IA requerirรก mucho mรกs procesamiento en relaciรณn al volumen de la respuesta que transmite al final.
Thereโs a lot of transcoding involved (you have to create versions for different codecs/resolutions/bitrates), and for variable-bitrate streaming you have to actively monitor the network conditions every 15 seconds or so in order to change the file youโre sending. Itโs not a passive process.
Many areas are under drought conditions with barely enough water for current use (drinking, watering crops etc.) So diverting huge amounts of water from those systems to data centers is a real problem. And the water is tied up for a long time to do the cooling, it doesnโt just flow through quickly
>โ20% of those data centers โdraw water from moderately to highly stressed watersheds in the western U.S.,โโ
ok this kinda answers my question. thanks!
Your removing more water from one place (ground water) where it's usually located and it's not sure it get back to that exact place. Resulting in that place (the ground?) drying out?
Guess you could use water from somewhere that doesn't have issue with getting dry. Like the ocean.
yeah I mean you *could* unsustainably drain an aquifer. but surely you could just build your DC somewhere wetter where thatโs not a concern? A data centre in a drought-stricken desert would be stupid yeah
Unfortunately there are a lot of data centers in places like Texas where water is an issue. Decisions about where to put them are made more by real estate prices than the water situation. And water is relatively cheap. So theyโll choose the cheaper area even if it stresses the local water system
Yeah. I mean they could do it properly and in the right places but theyโre probably not going to unless some one makes them. Water should probably cost a lot more in those areas than it does, at least for big companies but water politics is very complicated
Other considerations that lead to a proliferation of data centers in Arizona is the relative lack of weather that causes sustained power outages such as hurricanes and tornados. Also no earthquakes.
Oh thatโs an interesting angle. So as long as they can actually get the water they know their centers are safer than other places. But meanwhile the local area gets even more water stress. Ugh
Itโs so maddening. I used to work for a cloud services provider so got to understand a bit about the business. And of course Arizona is notably lax with their water laws. Ask me about hard rock mining in our most important refugia sometime. Iโm very active in the fight against that horror.
The recharge time constant on the Ogallala Aquifer is about 3k years. So you arenโt getting that water back. Similar issue for other ground water sources.
This long read from the Guardian a few months back had some great stats on the environmental costs of data centres. amp.theguardian.com/world/2024/f...
My limited understanding is that it then falls as rain somewhere else, and this depletes the local water source. That's a big deal locally (I don't know if there's a global impact).
I almost always read complaints about water usage as responses to a โTell me you donโt understand the hydrological cycle without telling me you donโt understand the hydrological cycleโ prompt
Where evaporative cooling's used, it's lost to the hydrologic cycle. (Well, okay, it never left the hydrologic cycle, but you know what I mean.)
But yeah, it's not like water used on, say, golf courses, where the stuff that's not taken up by foliage is now runoff with chemicals dissolved in it.
Depends how you do it. Industrial plants that use rivers for cooling might not contaminate the water, but they'll heat it up before sending it downstream, which messes with the fish and stuff.
And evaporation is also a potential negative as someone mentioned.
With current industrial designs the water is in a separate loop used to cool a bath of nonconductive coolant that the computers are submerged entirely into.
Unfortunately, the cooling loop is not as modern and optimised, with many larger buildings still opting for evaporative coolers that trade off regular water use for increased cooling. Hence the water use.
To avoid shorts or failure to cool the coolant bath loop has to remain as is on the computing side with regular maintenance and dialysis, but most such coolant evaporates very slowly even over a prolonged period with contamination a bigger concern
pumping hot water out into places that werent expecting it tends to disrupt the flora and fauna somewhat. check out things like algal blooms and their relation to temperature.
Even if you got it all back instantly, you'd use it all up again bc you still need it.
But you don't get it back, it evaporates and comes down somewhere else, not necessarily where it gets collected or used again any time soon, e.g. in the sea.
And even in parts of the US, water is getting rarer and rarer. And that's only gonna get worse with increasing climate change. States are arguing over water rights and native american protest movements have mostly been over water rights.
I gave a talk at @re-publica.com in Berlin earlier today on the growing backlash to data centers.
As tech companies build more hyperscale data centers around the world, they demand even more energy and water โ and some communities are fed up. Check it out:
So I live in St Pete, FL and we have a power plant that pumps hot water out to the bay and it has interesting effects on wild life, specifically certain species of big fish you wouldn't expect that far up in the bay who love it (cobia) and manatees vibing during the cold months. Prolly some algae...
... and seagrass implications (seagrass is a Huge Deal in Tampa Bay).
With the right purposeful management maybe there's a way to make it useful! But I'd wager that'd be in a wildlife / ecosystem context, not, like, energy production.
It's so dumb. People use "litres or gallons per year" to try to make the numbers sound big. Try putting the figures in *acre feet* (an acre flooded to the depth of one foot), the measure by which US agriculture consumes water. "A litre" or "a gallon" is a meaninglessly small measure of water.
i can speak with a little authority on this, i work in this field
DCs expend water by evaporation, either in cooling towers or evaporative coolers (water becomes vapor, taking latent heat energy with it into the air and leaving cooler liquid water behind)
so we expend liquid water resources 1/
The very fact that we donโt need as many data centers as we have and weโre running out of water anyway so why add to the problem with useless things like AI or crypto?
It's about mostly about ravaging ground water levels in areas that are already under water stress or even drought.
The water itself isn't necessarily bad after use, but using drinking water from the grid or even drilling your own wells lowers its availability for other uses.
I think it takes a lot of effort to get the water in a place to use it. Municipal water is usually potable, and that takes effort, plus it seems like a terrible use of potable water.
Itโs an additional demand on water resources. While itโs in the process itโs not able to be used somewhere else or is pulled from natural sources. Itโs continually in use.
As I understand it: that water is being evaporated to disperse waste heat, so if itโs a closed system sure but I assume itโs being lost atmospherically making it a global problem to โcool it down and reuse itโ.
When you heat water up, it cools down fastest by evaporating some of it. Then that water goes elsewhere and becomes rain. If you live someplace where rain doesn't happen much, that means the water is gone and it's a Problem. If you live where rain happens, you now have even more of it.
I imagine goes down the drain and so itโs consuming drinking water. Which is a renewable resource but finite throughput I guess. Seems like itโs less on an issue than almond milk, though.
You could probably make a closed-loop cooling system but it would cost more.
The water issue is specific to datacenters in arid places, where evaporative cooling is cheaper than air conditioning. But those are also places that don't have a lot of water. This is one reason why the Colorado River doesn't reach the ocean.
I am also very confused by this, like is the issue that after it evaporates it ends up raining down over an ocean or something? Like we are swapping freshwater supplies for saltwater? Or itโs that weโre locally depleting water in places that canโt afford to lose it?
I would have thought the resources weโd be most worried about are whatever fuel is consumed for the electricity, and weโd be concerned by the amount of heating released as well? Itโs not like the water vapor is hitting escape velocity, like it comes back somehow (maybe just not to a useful place?)
Yeah I have been arguing the same thing. Like the water doesnโt just get destroyed or become unusable to keep cooling, as far as I know these pcโs donโt run on radioactive fuel rods! Also why dies this not seem to be a problem for like massive render farms?
Think mainly it's just that we're approaching the limit of fresh water on earth at once. Obviously not completely, but our capacity, at least. Not bad yet, but supposed to get bad in 10-20 years.
Twtr used to thrive on the โyouโre either with us or against usโ dynamic. There are definitely valid criticisms but online discourse rarely involves discussing solutions to problems. It would also require a lot more text and time. Itโs easier to use hyperbole to highlight the problem.