AI's thirst for water is alarming, but may solve itself Once an abstract subject of science fiction and academic research, the concept of artificial intelligence has become the topic of dinner table conversations over the past two years. This shift has brought widespread awareness of the environmental implications of this technology, most prominently centered on the massive sums of …
Did the people concerned skip school?
If water evaporates, it doesn't get removed from environment - it is released back to the environment.
You know, water doesn't go to space, it comes back as something called "rain".
Rain is when you see droplets of water coming down from the sky and when you get wet if you don't get an umbrella -ella -ella
Well, it's not much good if you've drained the local aquifers for your datacentre and the rain falls elsewhere, or in the sea, is it
Look up the process of desertification, covered in GCSE Geography
Excessive draining of an aquifer means plants die out and no longer contribute to local rainfall, and then nothing grows in that area again
Plants (unlike datacentres) also have the ability to trap and store water. If rain falls heavily without them then it doesn't replenish the aquifers, it just flows out rapidly to sea
Nature didn't let us do anything. We've been defying nature since we first picked up a stick and hit something with it. Making ourselves unnatural by using technology to eliminate our shortcomings is the core of human development, it's what we do and what we are.
> So the nature will sort it out.
True.
Once humans have built, burnt and bombed the planet into a state it can no longer support mammalian life, Time and Nature will work together to restore an ecosystem.
So, proof positive we can all do whatever we damn well like, there won't be any consequences. Well, none that the newly-sapient society of motile fungoids will blame us for.
quote: datacentre no longer can cool itself off and burns down.
If it takes water from the local community, the local community will burn it down. And the single bloke staffing it won't be able to do anything about it.
Positive note: If it is an AI datacentre, nothing of value will be lost.
...training a GPT-3-sized model could generate enough heat to support roughly 4.6 greenhouses and grow over a million tomatoes.
Refraining from training a GPT-3-sized model would enable you to support roughly 9.2 greenhouses and grow several million tomatoes.
The real question then is: What did we get for all the greenhouses and tomatoes? An artificial conversation on how many greenhouses were spared and tomatoes weren't spoiled?
One possibly useful way around this would be to use the waste heat from computing to feed into district heating schemes. In the UK there are as yet very few of these, but I would expect this paucity to change fairly rapidly once the free market gets involved with green matters. Domestic properties are a major user of heat, and if industrial waste heat could be converted into steam and this steam used as a means of transporting heat around a suburb, then a fair amount of energy that otherwise would be wasted could be re-used.
The means to turn low-grade waste heat into higher temperatures already exists with grid-scale heat pumps; AI bit-barns would be an ideal supplier for such industrial heat pump systems.
"In the UK there are as yet very few of these, but I would expect this paucity to change fairly rapidly once the free market gets involved with green matters"
The free market has little to do with it. All energy and building matters are driven purely by government policy, you and I get the opportunity to buy what government policy allows to be delivered. I've worked for a company that built and operated district heating systems, and like all heavy infrastructure it's exceptionally difficult to make a commercial return on anything network based, without subsidy or mandate. In urban areas, building a new heat main in made up ground often costs around £10k per linear meter. For exceptional circumstances in city centres where the underground environment is heavily used I've even seen short network extensions or replacements that cost £100k per linear meter.
If people would be good enough to live in very high density housing then cost per property comes right down, and as experience increases unit costs should come down, but it remains an inconvenient fact that district heating systems are big and costly, and the cost per property of say Denmark or Sweden is amongst the highest in Europe, despite far more multi-dwelling properties and much better insulation.
One company in the UK has built a mostly low density heat network near Exeter - and to make the economics work they needed an 80 year non-competitive contract.
https://www.devonlive.com/news/cost-of-living/cranbrook-residents-fury-over-ridiculously-8062965
Yeah, while we've seen some use-waste-heat-from-datacenter schemes, the economics either have to be forced (subsidies, regulation) in all but very special cases. You could in theory build cold-climate datacenters within single-building settlements like Whittier, Alaska, but you'd have to convince people to live in them. I personally think it'd be interesting to live in that sort of data-arcology for maybe a few years, but I don't know that I'd want to do it permanently.
“ forced (subsidies, regulation)”
For our situation I have to disagree. We simply duct the hot isle into the office building we are in. Return air is from the outside. The net gain from not running the cooling system far outweigh the cost of the duct fans. This does not even include the savings from the building.
That said, it does take a lot of water to keep the humidity high enough to prevent the servers from burning up. (Moisture is how heat is transferred)
For replacement of 20 tons of cooling from outside at temperature of 30 degrees F requires 35 gal of clean water per 24hrs. Atomizing that much water required us to design and build or own humidifiers.
This post has been deleted by its author
"One possibly useful way around this would be to use the waste heat from computing to feed into district heating schemes. "
If the water used can be recovered and potable, perhaps a more comprehensive solution including using the heat for other applications will work.
Fresh water has been and is a limitation to populations in a particular area. Humans have rigged the game through massive works to bring water from where it is to where people are, but that's getting more difficult and the remote sources drying up means there won't be those options, but there will still be large populations in places without enough local water resources. Los Angeles is such a place. Also look up the Salton Sea which was created when a project to bring water from the Colorado River to southern California burst it's banks. Phoenix and Las Vegas rely on fresh water brought in from many miles away to support their populations which keep on growing as developers build more houses.
The trading of humans for a new group of digital overloads that compete for the same resources will need to be formally considered. So far, it's just been "build, build, build" without any planning. We hope that it would be the data centers that get turned off if there's a choice between fresh water and electricity for people or the machines, but big companies have more sway even if they can't vote.
Capitalism is about allocating capital. If the datacentres flock to your local water source and gobble it all up, it means the water is too cheap. Of course, making the water hella spendy for everyone isn't popular either. But maybe selling water for evaporation should have a different tariff than selling water for domestic use, where it comes back through the sewer system.
I'm quite sure you can employ all sorts of tricks, like painting the bit barn superwhite, adding PV on the roof, and so on, to keep energy and water use down. So find incentives for bit barn owners to go and use them, too. (Whatever happened to that "park a computing thingy in your basement, get free space heating" initiative?) Or you can try and find synergies, despite that being a horribly overused idiot management buzzword by now.
One feature of this latest AI hype is that it needs most of its number crunching to go through all that data and build the model, which isn't latency sensitive. Meaning you can park the bit barn wherever you want to, and it doesn't have to be just for the cost of 'leccy and water. So you could, oh, park an evaporatively cooled datacentre in your newly planted regreening project in the desert in hopes of helping it rain more there. You'd need to bring the water, probably sea water and deal with the salt somehow. (Get it out first. Then sell it? Build molten salt batteries with it? Something useful, please.) But at least it'd be "dual use", number crunching AND regreening. Now for some computational ecologists to do a bit of modelling to see if that idea could possibly fly, and if so in what circumstances. (What, that's people meddling in the environment? Yes, yes it is. Consider that IIRC the Sahara got enlarged a good bit because humans started to keep goats there. Goats eat everything, so no more plants to keep the sand in check.)
It's been said that good rulers never do anything for just a single reason. It seems such rulers are a tad scarce.
That's not what it says. It says you could use (influence) the mechanism of capital allocation to get some grip on the problem, too. That's real capitalism, and when done properly, without giving room to unfettered greed.
Thus capitalism is not a politically correct term for greed. Recall that political correctness is a leftist vehicle, so rather, it's the other way around. Doesn't change that capital allocation is the lifeblood of the economy. Fuck it up and you end up with a shambles, regardless of the ideology you use to legitimise your rule.
"When done properly..."
Imma stop you right there. It has NEVER been "properly."
https://www.theguardian.com/commentisfree/2021/oct/06/offshoring-wealth-capitalism-pandora-papers
Trashing the planet and hiding the money isn’t a perversion of capitalism. It is capitalism
.
All of these ideas of linking together evaporative cooling, desalination, greening the desert, finding uses for waste heat....they're all very well, but they cost money.
How long is a sea water main going to need to be? What's the capex for that? What about the capex for de-sal? How will you address the eco-challenges that the desert is home to rare <insert species of choice> that will be harmed by greening? Where does the cheap energy come from in the desert? You've got around 6-7 peak sun hours per day, how much for batteries and PV over-provision to support night and non-peak hours DC ops, plus the de-sal? What's the power backup arrangements - expensive but rarely used grid connection, or diesel standby that they have to freight the fuel in for? Where will the maintenance and ops staff live? Why will the evaporative cooling result in precipitation over the desert? As a rule, deserts have such low atmospheric humidity that there's no feasible prospect that an evap cooling system would result in beneficial precipitation...and then there's the challenge that even if you can force it to happen, desert soils aren't fertile as there's little nitrogen, and they're sandy, chalky and high in phosphates. if you can add organic material as well as water you can overcome that (eg Nile Valley), but good luck finding the organic matter to green a desert.
As with almost any technical ideas, each step is demonstrably resolvable with existing technology. The problem is that the chained "solutions" don't solve the whole problem unless somebody is prepared to burn money. A LOT of money.
All good points and I'm glad you're attacking the content rather than making fun of it based on a misreading. Honestly, I haven't the faintest, haven't done the math. Just throwing example ideas out there. I will note that occasionally people do stand up, willing to spend a lot of money on a project. Like that "the line" city idea, which appears to make no economic sense. Simply plunking more datacentres down wherever it seems economically convenient is perhaps the other extreme. The article correctly points out that easily becomes problematic.
Point being, how about someone with a bigger plan gathering up a number of goals and to try and work out solutions for multiple of them together in a way that brings us more than each individual solution to each individual problem would bring? I think we increasingly may have to, having more or less exhausted the viability of single-purpose solutions in the face of mounting environmental and other concerns.
After "how about we try this?", of course, we get "but then who to do it?" And from that "how do we find capable people of sufficient vision?" Back in the day you could have an energetic king try it himself or find a suitable deputy to be grand architect of the overarching plan. How about today in our modern democracies?
"How will you address the eco-challenges that the desert is home to rare <insert species of choice> that will be harmed by greening? Where does the cheap energy come from in the desert? You've got around 6-7 peak sun hours per day, how much for batteries and PV over-provision to support night and non-peak hours DC ops, plus the de-sal? What's the power backup arrangements - expensive but rarely used grid connection"
All of these questions need to be asked and addressed. I live in a desert community and it 'rains' energy most days of the year. I've not used central heat since I bought the house in 2015. As time goes by, I'm adding to my passive solar heating system so each winter I can remove another article of clothing on a sunny day while inside. I'll know when I've hit the useful amount of panels. I have plans and many of the parts for a thermal battery to bridge cloudy days and I'm slowly moving appliances to solar operation with battery back-up (made with recovered cells). I am hooked up to the grid or the city would condemn the house as not suitable for occupation. I must have water service as well. The cost to build electrical generation on-site gets more expensive as I reach zero input so using a bit and just eating the standing charge is fine. I could have water delivered, but I keep my usage down to the included baseline so adding hygienic storage, the plumbing and etc would take a lot of time to recoup if it were allowed. The vast majority of the automated gadgets shown on "eco-houses" have no ROI. Lucky for me, collecting rain water is not banned where I am so I plan on doing that so I have water for the garden. If I can buy the land next to mine, I'll have the space for more water storage. Starting with a basic IBC container will let me experiment first.
How a house is sited on a parcel can be a giant factor in how well it can perform, but most people don't take that into consideration. Commercial buildings can strongly benefit from that as well. I've noticed that old factories will have many roof pitches facing south (in the northern hemisphere) with windows on the vertial North faces to bring in light with it being direct sunlight. If the pitch is at the correct angle, it would do well to have solar panels on. What's more common and cheap to build are giant buildings with a flat roof. With expensive power and limited supplies, buildings were designed with efficiency in mind. The MBA's have run their first approximations and don't look past the build cost. Companies don't look past the sale price of a building to what it might cost to operate.
...is this innovation? It doesn't look like it. It looks like we're just throwing masses and masses more raw compute at doing the same dumb thing faster and faster. Innovation would be getting the same results with fewer resources, or better results with the same.
The article makes mention of ChatGPT-4 having 1.6 to 1.7 trillion parameters, or 2-3x what ChatGPT-3 had.
At some point the model size will quit going up, because there won't be any more information to feed it. Once you've fed it the internet (or at least that portion of it you're legally allowed to scrape) and the Library of Congress, and so forth, there isn't anything else. I don't know where the current models are in terms of the percentage of all possible information they are fed vs could be fed, but that seems like a pretty hard limit.
Sure the internet will continue to grow, but if that growth is mostly Reddit trolling and the babblings of AI placed there by SEOs hoping to get a high search ranking, it isn't gobs of new information that will make the models much larger. The yearly growth in actual new human knowledge and culture will be below the yearly growth in AI computing power.
So by necessity at some point the problem will turn to how to make them smarter with the information they have, and that will require some new technology that supplants LLMs. There are some useful things they can do, but only a fool would believe that path will lead to AGI.
" Once you've fed it the internet (or at least that portion of it you're legally allowed to scrape)..."
[Insert Homeric laughter here] Legal issues regarding access to and use of information are being challenged by AI as we speak. If information can be accessed through the Internet, it's fair game. If I read a copyrighted newspaper article I'm training the neural network in my organic brain on it and I'm better informed thereby which enables me to do my work better and increase my income. The idea that an AI operator would have to pay for using the same information to train a neural network in an artificial construct will not be legally defensible for much longer. Legal battles are being fought over it already as we speak, and have been for some time.
But that's not the biggest issue here. Which is: there is always more information to feed the models! Not only is there a lot of information still in cold storage (from data tape vaults full of NASA data from countless space probe missions to historic records that still have to be digitized) but new information becomes available on a daily basis. AI training is an ongoing and never-ending process.
Then there's the next big thing for AI: the processing of all the information that the models have already been trained on into new bits of information. Correlation, extrapolation, what not. Which could consume even more power that current AI deployment.
So "running out of information for AI to work with" is not going to happen.
Legal and ethical aren't the same. It's technically legal for me to look at some art in a window, copy the brushwork and make a painting in a similar style but that wouldn't be ethical without their permission if the artist didn't intend for their work to be copied or inspire other works. All this shows is that the average creative needs better copyright laws to protect their honest work against the big corporations. Corporations have always hated copyright and have done all they can to take it from us.
"So "running out of information for AI to work with" is not going to happen."
So much "information" is repetitive and low grade to begin with. The Foundation series does a good job of pointing out how history can become more and more fragmented through the choices being made about what to preserve. Hari Seldon had a devil of a time finding contiguous histories to develop Psychoshistory. While a fictional character and story, the premise is a good one.
Since AI's don't think, problems occur with memes getting reinforced even though they weren't useful to start with. A human might discard a meme that doesn't fit, but how does an AI make that sort of qualitative decision? You get outputs that tell you spotting melanoma is easy, all you need to do is look for a ruler. Obviously, that's wrong, but every photo example of melanoma includes a ruler to calibrate the size of the lesion. A human would ignore the ruler, the AI doesn't. This is a simple example where the problem is easy to spot. For very complex analysis, it might be nearly impossible.
I would think that AI training is best based on relevant materials. If you are trying to analyze binary star systems, having lots of historical data on ladies skirt lengths won't help and could hinder.
The problem is that AI won't know that, as it is not possible to know what has come out of an AI orifice and what hasn't. It will continue to "consume" data that has been generated by AI and the output will become more and more bland, and it will asymptotically converge with a result that may or may not be sensible.
Not your downvoter btw.
Cities use *a lot* of potable water *but* do return almost all of it in a "slightly used" form with additional elements.
I would think it poetic if this return product were diverted to be piped into these AI factories but then it wouldn't be my taking the piss or even, perhaps, my giving a shit.
Unfortunately The Vulture doesn't offer Space Karen's emoji as a forum icon.
Modern wastewater treatment can easily return municipal sewage to safely drinkable standards (if the plant is so designed and operated), so it's feasible what you suggest However, DCs are always built where land and power are cheap and planning approval is easy, and those locations are rarely adjacent to a city with a large enough wastewater treatment works to make a difference.
There's also the fact that often treated effluent is returned to a watercourse from which water is abstracted for drinking downstream, so that using a treated effluent for evap cooling does take away somebody else's raw water resource. As an example, I wouldn't like to guess how many drinking water treatment plants use the Mississippi and it's tributaries, but the cities of St Paul's, Minneapolis, St Louis, Memphis, and New Orleans all rely on the river for drinking water.
It's usually only near the coast that you could presume the treated effluent is mere "waste", and in that case you might as well use a sea water cooling system.
"Modern wastewater treatment can easily return municipal sewage to safely drinkable standards (if the plant is so designed and operated), so it's feasible what you suggest However, DCs are always built where land and power are cheap and planning approval is easy, and those locations are rarely adjacent to a city with a large enough wastewater treatment works to make a difference."
There was a recent financial study commissioned by the city where I live and reclamating water so it's safe to use to water the local golf course and fill the lake at the park was more expensive than using potable well water and not treating the waste water to a higher standard.
Cities often have baseline charges that are calculated by the size of the pipe going to a property. A DC with a large input pipe is going to make the city wet themselves since it's going to be a nice income stream. I highly doubt they're looking as much at the draw on wherever that water is coming from and how the usage might impact adding other homes and businesses in the area before that source is maxed out.
A few municipalities in very dry areas have a "grey water" system. They aren't piping sink and shower water from single family homes, but large apartment buildings and industrial uses that result in grey water is recovered for uses like irrigation of golf courses to avoid using treated water for such purposes. That would work for evaporative cooling, but it is available in so few places it can't be any sort of standard solution.
If it was close enough to the treatment plan they could build a pipe to deliver "lightly treated water". i.e. filter the suspended solids and put in enough chlorine or other chemicals to kill any bacteria so it is nowhere near potable but adequate for the needs of evaporative cooling.
Why are we* building data centers in hot, arid climates? And then arguing about the economics of pipelines? Data can move throgh a kind of pipeline as well. About the thickness of my hair (which is pretty thin these days ).
That some VC investors insist on financing stuff in their back yards doesn't speak well for their faith in the Interwebs. Or their sudden reticence to buy yet another corporate jet.
*Yeah. I get that the "we" putting up the money gets to pick where it goes. But when IPO time rolls around, I'm going to think twice about buying into the heat sensitive tech built in the middle of Death Valley.
"Semi-desert is generally very cheap land, far from people likely to complain about a big ugly building."
That's where They are putting the solar farms. On a trip I made yesterday, I noticed a whole bunch more being built in an area that's not useful at the moment for agriculture. I expect that due to drought conditions, the farmers have had to sell off parcels they can't pay for any longer. I have been seeing these solar installations being built on taller pylons. I'd love to find out if that's to allow animals to more easily graze under and around them. Every time I see a flock of sheep being deployed I never have the time to stop and chat with the shepherd. The sheep will be brought in for a day and taken someplace else the next. That would be a synergistic way to maintain the solar farms. Since they are large fully fenced parcels, there's no worry that the sheep will wander off.
There is water in the desert, but it's a long way down. A compromise might be to say to data centres that they should drill the necessary holes and use that source, provided that a proportion of it is used to irrigate the desert. The problem here is that will have a knock-on effect with the current (sorry) outflow from that water table.
"The trouble with aquifers under deserts is that they are incredibly slow to refill.
Because there's very little rainfall."
Where I am, the rain water flows down into the desert valley from the surrounding mountains and soaks in. Of course, the aquifers aren't infinite so while they can refill at a better rate than some might expect, that doesn't mean that they'll remain useful if abused. Just like a lead-acid battery, bad things happen if you try to use too much of the capacity.
evaporative cooling? In DC? I'm surprised, swamp coolers don't work worth a damn when it's humid out, Which it often is in DC,. In Arizona etc. where it's like 10% humidity they work great (but you are in a desert so it'd use too much water for a data center I would think.). I mean obviously it must work since they're using it, I'm just surprised it does.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
