I was taken aback recently when a Gen-ish Z person told me AI was 'destroying all the water'. I've done data center work, and while I know it is used for cooling, I don't think I've ever personally destroyed any large bodies of water.
There is a perception out there about GenAI and water that goes surprisingly deep. I was told we are will be living in a drought-stricken hellscape, and AI is to blame.
I'd like to know the equivalent energy consumption of a single TikTok video, but that is probably arguing the wrong thing. My bigger question is ... where do they think that water goes? Steam? The assumption is that it is gone forever, and I can't get over how people could just take that at face value.
I was also surprised when someone asked me about AI's water consumption because I had never heard of it being an issue. But a cursory search shows that datacenters use quite a bit more water than I realized, on the order of 1 liter of water per kWh of electricity. I see a lot of talk about how the hyperscalers are doing better than this and are trying to get to net-positive, but everything I saw was about quantifying and optimizing this number rather than debunking it as some sort of myth.
I find "1 liter per kWh" to be a bit hard to visualize, but when they talk about building a gigawatt datacenter, that's 278L/s. A typical showerhead is 0.16L/s. The Californian almond industry apparently uses roughly 200kL/s averaged over the entire year -- 278L/s is enough for about 4 square miles of almond orchards.
So it seems like a real thing but maybe not that drastic, especially since I think the hyperscaler numbers are better than this.
Water shortages are highly local problems. Sure, running a data center in Arizona might have some genuine water concerns. But even then, it can be mitigated by changes like using wastewater. The Palo Verde plant does that for its heat exchangers.
says Google's datacenter water consumption in 2023 was 5.2 billion gallons, or ~14 million gallons a day. Microsoft was ~4.7, Facebook was 2.6, AWS didn't seem to disclose, Apple was 2.3. These numbers seem pulled from what the companies published.
The total for these companies was ~30 million gallons a day. Apply your best guesses as to what fraction of datacenter usage they are, what fraction of datacenter usage is AI, and what 2025 usage looks like compared to 2023. My guess is it's unlikely to come out to more than 120 million.
I didn't vet this that carefully so take the numbers with a grain of salt, but the rough comparison does seem to hold that Arizona golf courses are larger users of water.
Agricultural numbers are much higher, the California almond industry uses ~4000 million gallons of water a day.
Data centers don't just heat up the water and return it - they evaporate the water into the atmosphere (yes, I know, the H2O still exists, but it's in a far less usable form when it's gaseous atmospheric H2O)
Yeah, the funny thing about the water claims is that there's a pretty simple technological fix: closed loop exists, and seems to be simply the better choice going forward as new DCs are built (excepting of course things like musk's 'rolling coal' bullshit).
The claim that they would recondense the water is surprising to me. :)
I'm genuinely curious how recondensing would work. It seems like a Rube Goldberg machine to me. To recondense the water vapor, you would need to move a large amount of heat out of the water vapor into a heat sink so that it could cool down to the boiling point and undergo a phase change. What would that sink be, and how would you move the heat into it fast enough?
Air is the obvious answer, but if you are dumping heat into the air, why would you do it by evaporating and then condensing water, rather than transferring it directly (e.g. via a heat exchanger and fan)?
Destroying the water originates as a NIMBY talking point (to stop data centers) that was co-opted by anti-capitalist groups (higher membership rates among GenZ) as a boogeyman to blame for why AI is bad.
Data centers do consume a lot of water, but even more power. AI is causing us to forget our climate change carbon goals.
The global increase in drought is probably more a result of climate change temperatures than direct consumption. (Yes AI is doing it but not in the way most people thought)
Stopping AI development in the name of Climate Change so that we lose to China (who pollutes 3x as much as we do) is idiotic as best, self-destructing at worst.
Since pulling out of the Paris Climate Accords, defunding climate research and doctoring past data, it's no longer possible to take the moral high ground here.
It's probably best, for your mental health, to ask these questions in earnest, and stop dismissing people as illogical. The communities living near data centers have real water problems, making it believable. If you're wondering why NIMBY happens, just watch the first 30 seconds.
And there isn't solid evidence that this was connected to the data center construction:
> Ben Sheidler, a spokesman for the Joint Development Authority, which manages the industrial park that Meta’s facilities occupy, said the cause of the water issues was unknown. The Joint Development Authority did not do a well water study before construction to determine any potential effects, but the timing of the problems could be a coincidence, he said.
> “I wouldn’t want to speculate that even the construction had something to do with it,” he said. “One thousand feet away is a pretty significant distance.”
People are also more likely to click into web content that helps them learn more — such as an in-depth review, an original post, a unique perspective or a thoughtful first-person analysis
So... not the blog spam that was previously prioritized by Google Search? It's almost as if SEO had some downsides they are only just now discovering.
1) Clicking on search results doesn't bring $ to Google and takes users off their site. Surely they're thinking of ways to address this. Ads?
2) Having to click off to another site to learn more is really a deficiency in the AI summary. I'd expect Google would rather you to go into AI mode where they control the experience and have more opportunities to monetize. Ads?
We are in the "early uber" and "early airbnb" days ... enjoy it while it's great!
No SEO is required for quality content, as by definition qualitative content contains the words and terms that make it qualitative. The problem is the low quality spam that's deliberately SEO'd to masquerade as quality content.
Even though Varnish may not be in fashion any more, there were many companies happily using it for free and still demanding security updates.
I like their transparency about who actually supports them, and what the whole community gets for it. I wish other projects would do that, if for no other reason than to make it obvious that FOSS isn't just something that happens.
Given that Factorio has logic gates and people have built various programs (including Doom, iirc) how long will it take before someone runs an LLM inside the game?
I'm reminded of Hitchhikers Guide to the Galaxy, where the whole planet is one big computer. I would expect that any model directly implemented in Factorio would take up most of the game-world.
I once had a (older) Russian colleague tell me about how he learned to program. He would write it out longhand, because access to compilers was limited to certain time slots in University. He know it had to work first time, otherwise he had to wait for the next chance to fix the bug.
I'm sure that was true for everyone back in the punchcard days. It would enforce a kind of rigor that I can blissfully ignore.
edit: I see the exact same story in the linked thread, so clearly a lot of Russians are very proud of that skill
Quite simply, when you had to walk across campus or at least to a different room to submit your card deck, wait (perhaps hours) for the results (printed on fan-fold paper, that again you had to go to a different building or room to pick up) only to find your program didn't compile due to a syntax error or didn't run due to a careless bug, you learned to "desk check" your code, run the program in your head, and be as sure as you could be that there were no errors.
Even when we got connected terminals, it could still take hours for your compile job to work its way through the queue of pending jobs, because it was in a development region that only got resources when the production queue was clear. You didn't use the compiler as a syntax checker in those days.
That all started to change when we got PCs or workstations, or at least good interactive multiuser development environents, and a "code-compile" loop or repl became part of the standard toolkit.
I have a old guy that I work with - PhD in Math (because CS didn't exist then) - who does lots of algorithm development with me. I often get Word docs of pseudo code from him. I'll do a search-and-replace on things like "LET" and "ELSE IF" and a very high percentage of the time if I run it in Python it works on the first try. Kind of amazing to me.
The hard part for me is then translating his ideas into vectorized numpy for speed, but at least I get the right answer to check against.
I sometimes wonder whether following some of these practices may promote more mediocre programmers, if they so wish, to become better ones.
- Think through and write on paper in pseudo code first;
- Run written code in head, or on paper if they don't have the capacity, a couple of times before pressing that BUILD menu item;
- Restrain from using libraries if a custom, better solution is possible;
But then I think, it probably doesn't make a lot of sense if you work as a frontend pushing out JS, or as a data eng writing Python code which calls some Scala code which runs in JVM which is highly complicated, because the whole industry is "AGILE" and the chain is way too long. You need to get into the right job, to nurture such a mindset. The old timers got lucky. They started with 6502 bare metal.
That's why I'm pushing myself to get out of Data Engineering, to do something lower level, until I get deep into the basement. I probably won't succeed here but it's fun to try.
Not sure if actually writing it out on paper is necessary. But along these lines I will often start my code by just writing comments explaining what the code in this file does. Then as I get into writing code, I break up the comments and insert the functions/classes below the comments that describe what they do. So sort of literate programming, but I don't usually go to the lengths that would really qualify that description..
I disagree about not using libraries. Libraries are almost always going to be better tested, better reviewed, and handle more edge cases than anything you come up with in-house. Of course if no suitable library exists you have no choice.
I agree with the library being better tested part, so that's why I think it's better to find a job that actually doesn't allow the use of libraries (or too many of them), than to try to go to the bottom in a job that has 5 layers of abstraction.
It's good to hear the literate programming thing. I sometimes do that on paper when I need to clear my mind. Basically code in English but with C type {} as scope.
I don't think this practice completely disappeared until laptops became commonplace. As late as 1996, I remember hand-writing several pages of assembly code during some down time at a conference; I had an idea I wanted to try out, but there were no computers in the conference center.
There's a little 'update' blob to say now (Oct 23) 'Expanding to Pro and Max plans'
It is confusing though. Why not a separate post?