Critics Furious Microsoft Is Training AI by Sucking Up Water During Drought | Hacker News Reader

Critics Furious Microsoft Is Training AI by Sucking Up Water During Drought

futurism.com

54 pointscdme46 comments

A) Water use and resource management is resource business. Water needs to be allocated in both a fair and humane way.

B) This headline and article is some classic modern-day repackaging "journalism".

It leans incredibly heavily on the original AP piece (at least it links liberally to them), adds a few semi constructive comments from 'actual critics' (thus just barely nudging them into the zone where they can use their headline with a straight face), adds some links to barely related things (the link for 'isn't fit for human consumption'? that link is mostly about nitrates and other agricultural run-offs), and the classic anonymous social media (Reddit) quotes.

And most egregiously, awhile cheerfully lifting right out the AP piece, they neglected to at least lift from probably the single most human scale and impactful snippet from the original AP piece - "In a paper due to be published later this year, Ren’s team estimates ChatGPT gulps up 500 milliliters of water (close to what’s in a 16-ounce water bottle) every time you ask it a series of between 5 to 50 prompts or questions. The range varies depending on where its servers are located and the season."

And price it accordingly to its value.

Half a liter of water? Wow.

We can probably derive the power usage from that, considering that the heat is generated by consuming electricity.

I assume they do evaporative cooling, Anyone with enough physics knowledge to calculate this?

At least in my kitchen, evaporating 500ml of water takes significant time and energy.

First result on Google give the energy use as about 0.002 KWh per query.

And evaporating 500ml is about 0.3kwh.

So that's about 150 queries.

> Anyone with enough physics knowledge to calculate this?

You could probably ask ChatGPT.

What's unclear from the article is how they're actually using the water.

Servers aren't 'thirsty' for water, there's not some dude with a firehose frantically running around the datacenters keeping things cool.

This seems to be intentional obfruscation of the real problem, which is datacenters using cooling towers with open loop cooling without any attempt at reclaimation simply because it's cheaper to do so than other options.

How much cheaper than using a closed loop system can it really be? We've run cars on closed loop liquid cooling for decades and it would seem outside of the slight additional cost of a heat exchanger, cost should be pretty even initially and lower in the long run.

It is a closed loop, they run muni water through the towers in open loop to cool the closed loop.

As far as how much cheaper? Well, heat pumps basically double your power costs, air handlers are highly contingent on reduced density and/or cool environmental air, so, alot cheaper to run the tap, no doubt.

It's more an issue of waste. That water, at that scale, should be treated and fed back into the system, but it's just discharged instead.

Is it really true that datacenters use freshwater once and then just discharge it to sewer? If so, the criticism of cloud companies (including but not limited to Microsoft) is valid. Especially that it's reasonably straightforward to build a fully-enclosed system, possibly requiring to incur a one-off capital expense to build a dry cooling tower ([1])

1. https://en.wikipedia.org/wiki/Cooling_tower

IIUC, evaporative cooling is capable of far, far, far greater cooling capacity per unit of flow. I can't speak to all datacenters everywhere, but the one I did work with had astonishing water usage, as the water was vaporized as it absorbed the thermal energy, the gas conveniently full of all that extra molecular wigglin' and shakin'. Getting the water back via condensation, can't really do that in quantities that match the heat, not without spending another big pile of energy. No free lunch I guess.

You have an analogue in weapons design, as this is one of the big reasons we haven't seen caseless ammunition. The case - in addition to providing a hot seal for the high pressure gas - also is an absolutely amazing heat sink, ejecting all that thermal mass right out the top of the weapon. Caseless ammo, the breech, barrel, bolt . . honestly, the wholy weapon - gets real hot real fast, exacerbated by the fact that now, with no case, everything has to have tight seals.

I don't think this addresses the comment you're replying to. Sure evaporative is more effective, I don't see an argument against that, but I think the point of GP is that fully enclosed (like a dry cooling tower) is possible and we can be critical of these companies for not doing the arguably more responsible/moral thing.

Yeah, the "oh well, it's evaporated already, so no getting it back without spending even more energy" tack of the post that you're responding to to is insane. That's not how water cooling works!

Is it more responsible to waste electricity and save water? That sounds like a damned if you do, damned if you don't trade-off between two bad things, not a clear good vs bad choice.

My understanding is that a closed cooling tower uses more energy because evaporation is so much better than radiation at removing heat.

1) It should be possible to build a mostly passive dry cooling tower. It will cost more to build, but the cloud companies are not particularly struggling.

2) Electricity is much more portable than water. Droughts are regional and there's little that can be done about it (other than avoiding waste), but electricity can be and often is transferred from out of state. And we do have a lot of solar+wind power nowadays, should be a good fit for cooling.

Microsoft mainly uses adiabatic cooling today in the US, the water is evaporated not dumped in the sewer.

https://datacenters.microsoft.com/wp-content/uploads/2023/05...

Thanks. That's a little better, because it has a higher chance of turning back to rain somewhere around (or at least making air a little more moist).

It won't rain where it is being used, Florida getting the rain from Arizona evaporation doesn't help out the Colorado basin.

I agree. Which is why full water recirculation for datacenter cooling should probably be a law.

It should work out a lot better, because 25C of cooling only requires you to lose 5% of the water per loop.

If the end result is a lot of freshwater lost, it's still very bad.

The end result is 1/20th as much lost. So if that's still "very bad" then the other plan is "catastrophic". It's a very big improvement.

The problem is not just with the volume lost water but that they also dump concentrated brine water into the sewer system and most of these companies never actually came though with their promises to upgrade municipal treatment facilities.

Millions of gallons of consumption and brine sewage is not a trivial problem for places that are rural.

That sounds to me like the utility is charging them wrong.

When I buy water it has sewer charges by default, which means if I concentrated that to seawater level they'd be getting 100x or 200x payments to process it, and even higher for brine.

It's possible to meter the sewer separately, but there are extra charges for water with too much biochemical oxygen demand or suspended solids. Throw an extreme salinity charge on there so they can pay for full treatment and then some.

Is there a reason to have datacenters directly fund facility upgrades?

It is worse. In the sewer it renters the groundwater system, ready for immediate use. evaporate and it goes somewhere else.

But the real take away is don't let the black journalism scare you. using water for an evaporation cooling system is about the least bad thing you can do with it.

No. The least bad thing to do is to stop using so many resources for so little usefulness.

usefulness is subjective

> "They're giving all the good water to ChatGPT," wrote one Redditor.

Where "Critics" apparently includes some reddit posts they found.

Welcome to modern "journalism": We found some disgruntled randos on Reddit and Twitter that should make for good quotes.

I can't find any statistics in this article. All I can find is a global water consumption figure (1.7 billion gallons worldwide), which is meaningless.

I have found this quote (https://www.firstpost.com/tech/news-analysis/thirsty-ai-micr...):

> Nonetheless, the quantity of water required, especially during the summer months, remains substantial. In July 2022, just prior to OpenAI’s completion of GPT-4 training, Microsoft pumped approximately 11.5 million gallons of water into its Iowa data centres, constituting about 6 per cent of the district’s total water consumption, which also serves the city’s residents.

but the document itself isn't linked.

6% of a city's consumption sure is a lot, but shutting down the datacenters completely won't make that big a dent in a drought.

Moving all data centers to the coast to make cooling have less of an environmental impact would benefit the water ecosystem, at the cost of slower internet connections for land-locked states, of course, but Microsoft didn't exactly cause this drought.

Looking online, all I can see is the same bits of information published over and over again by news sources quoting each other. I'm not entirely sure why this specific story and its exact wording are being used over and over again, but there seems to have been very little journalistic effort put into it since AP first published the story.

Let's do the math, instead of just being scary by using gallons as a unit inappropriately. The article says Microsoft wanted ~15 million gallons per month. The water works says it delivers 63 million gallons per day. Which mean Microsoft's datacenters demanded <1% of the water. Which I assume they contracted for.

Microsoft's global sustainability report says they "used" i.e. evaporated 6399 megaliters in 2022, globally. That is only 5161 acre-feet, for Americans. That's ~$2 million worth, globally.

According to the original article, the water usage during particular hot months (when the water was needed most), Microsoft used 6% of the available water. According to their own spec sheet (which may just be greenwashing, it's Microsoft after all) they're using a cooling technique that only requires additional water when the air is over a certain temperature (just above 29℃).

Total water usage isn't very interesting in this case. Their water usage increases right when drought conditions become the worst, during the weeks or months where temperatures rise and remain high. Worldwide, their water usage is negligible, but water shortages are local problems.

I'm pretty sure they're the number one customer for water in WDM during the hot summer months. [1]

> Yes, but: They (Microsoft West Des Moines Datacenter) consumed as much as 11.5 million gallons of water a month for cooling, or about 6% of WDM's total usage during peak summer usage during the last two years, according to information from West Des Moines Water Works

[1] https://www.axios.com/local/des-moines/2023/09/18/ai-iowa-ep...

Is there actual demand for AI? I don't think there is. It's just being shoved down everyone's throat unwittingly.

In the question of datacenter water usage, its application (AI or not) is actually not very relevant to the problem at hand. I challenge the fact why (if true) the datacenters even need "new" water, instead of having a closed system with full recirculation of water.

> instead of having a closed system with full recirculation of water.

The water would need to be cooled before reuse, which could look like a car radiator and fan. It's about cost, and new water is likely cheaper when dumping heat from the fixed location of the data center.

> The water would need to be cooled before reuse, which could look like a car radiator and fan.

Dry cooling tower is the word, I believe: https://en.wikipedia.org/wiki/Cooling_tower

it's only cheaper because the cost is externalized.

Everyone thinks their AI product is the one that consumers will finally want. It's like the Cambrian explosion of coins except those didn't involve scraping all of our personal data.

It’s the next Web3. Comes and goes in waves.

Cue the AI fanbois saying “well actually, AI has a ton of positive uses because I recognize them, while Web3 has none, because I won’t recognize any uses as valid”

The upsides are only one part of the equation, the downsides of AI are destruction of all trust in digital media and claims, chaos on the web and social forums, job loss on a massive scale and and possible human exinction. AI will affect you with downsides whether you want it or not. By contrast the downsides of Web3 are bugs in smart contracts that can make some people lose some of the money that they voluntarily sent over to that contract.

It's a repeat of Home Assistants heh

Amazing how much all the companies forgot how the bumrush of Alexa, Google Assistant, Siri and Cortana went

I think there is huge demand and faith from corporate executives. They can secretly propose their wild ideas without any employee resignations, ethics concerns (spending millions on a contract, you won't get consumer-restricted "safe" AI), etc.

Don’t give yourselves to these unnatural men - machine men with machine minds and machine hearts! You are not machines! You are not cattle! You are men! You have the love of humanity in your hearts! You don’t hate! Only the unloved hate - the unloved and the unnatural! Soldiers! Don’t fight for slavery! Fight for liberty!

Maybe everyone else on Hacker News are all super geniuses who don’t benefit from it, but I’m a developer about 9 years into my career and have found ChatGPT4 to be absolutely astonishing for helping me learn new skills and improve my productivity. It helped me figure out the design flaws that were causing me considerable difficulties when writing some generics, and we’ve got a new subsystem in an important product up and running that ChatGPT helped me and the other engineers design.

Furthermore I use GitHub Copilot literally every day. The combo of chatGPT for “big picture” discussions and copilot for writing code feels like I iterate faster and learn all sorts of things I wouldn’t dredge up in the documentation, or just wouldn’t feel like I had time for (like using Theories instead of Facts in a unit test and running multiple sets of data through the test, I was vaguely aware of it but just hadn’t done it).

I’ve made amazing art for my entire D&D party. I’ve got at least a few hundred hours racked up on Colab Premium renting A100s and generating pictures and iterating until I get something really really nice in an hour or two. Someone offered me $250 to make him a picture using the pretty standard Stable Diffusion toolset after I showed him some of the stuff I’ve made.

I don’t see how people aren’t absolutely amazed at what these tools can accomplish, and they’re just going to get better. To be clear, I didn’t buy a single crypto coin and thought Web3 was total snake oil (still do), despite my friend literally making millions of dollars off of Bitcoin. But I think this is a radically different situation and the benefits (at least in the short term to someone like me, who is already technical) seem overwhelming and obvious.

ChatGPT had over 100 million users a few months after launch, the demand is there.

I and a lot of other devs I know use chatgpt to assist then in various capabilities, be it writing code, debugging errors, setting up environments or suggesting libraries.

Other non devs enjoy humorous queries with chatgpt and other less constrained models, image generation is also popular

Afaict the city grew 1300% since 1995, and sells of most of its own water to neighbouring regions in long stading contracts running through to 2045.

What are the actual effects of using 6% of the water? Was 6% of the city's population left without drinking water? How did they survive?

Reply on news.ycombinator.com