I think its easy to forget how much energy is required to boil a liter of water, and these charts are a great reminder of the mundane things like that which escape our attention.
So what about the other 3%? That's not "massive"? Or it gets a pass because it's not AI? The whole debate over datacenter resource consumption feels like a rorschach test. Either you think the technology is transformative and therefore the resource consumption is justified, or you think it's slop and therefore any resource consumption can't be justified.
If it was 2% for VOD streaming, and 3% for every other service that is delivered by a networked computer, I would think "wow, streaming is huge". 2% for one thing that most people hardly use at all, versus 3% for everything else on the computer, is still a remarkable statistic. I'm not making any value judgment in order to say that.
I also don't think it's particularly conclusive about anything though. It doesn't on its face make me think anything like "it must be stopped", but it's remarkable.
Every company is using it? The deli and the laundromat too? Every consumer? Maybe transitively because they use Google or what have you, but most people I know still don't really know what AI is, never mind have a puritan revulsion over using it (some people are quite performative about this on here), but they just don't use it themselves, except transitively on occasion.
Was the 2% figure counting all those things? I don't think so. This is changing the grounds of the discussion. Ridiculously so. Do you take me for some zealot who hates tensors or matrices or something? I don't even hate LLMs, though I have huge issues with how data has been gathered. I feel like this discussion is useless.
Look at Meta’s capital budget and tell me people aren’t using AI. People use Facebook, Instagram, etc. I’d say the same for Google, but with Google Cloud it’s hard to know where the dollars are going and for whom.
Who exactly is the 2% figure? Can’t all be HN people on side projects using Claude code.
A lot of it is training, not inference, right? I really do think the vast majority of use is HN Claude Code, with some lawyers, etc., preparing initial draft emails that they rewrite, sprinkled on top.
As for capex implying high usage, look at xAI's capex and how it turns out most of it is now being rented out to Google, etc. High capex does not imply capex that has already found its customers and margins. Meta is an extraordinarily bad example to use, given the failure of the whole metaverse/VR thing that is now their name, but is completely vestigial.
The deli is using it to make advertising flyers. Every consumer is using it as a replacement for google search as far as I can tell. I do a fair bit of work with people in the 18-25 range and almost all of them love to rail against AI, but none of them seem to be capable of completing even simple tasks without using it. Even my 63 year old mother uses it at work to generate reports. I would love to see some hard numbers, but I would be truly shocked if "most people still don't really know what AI is" is in any way justifiable across all adults in OECD countries.
Speak for yourself. For "most people", "streaming" is basically TV for them, and the average american watches 2.6 hours per day[1]. It's probably safe to say that at least half of that is "streaming" of some sort, even more if you include other types of "streaming" like music streaming. Moreover nowadays even if they're not plopped in front of a TV, they're watching short form videos.
The "hardly use" comparison quickly runs into issues, though. If I use AI and it completes a job that would have taken an hour in 1 minute, should that be counted as a factor against AI, because I'm "hardly" using it (1 minute) compared to the alternative (1 hour)?
and while ambiguous, I meant that quantifying the number of people that use it, results in a not large number. I've had discussions in other threads about how to talk about transitive use, etc., so as with everything at all times, there are certainly ways to skin the cat that justify or invalidate the statement. One has to rely upon good judgment to decide which way is best, and reasonable people may disagree.
By all means, cull them all. I'm sure the remaining 3% support (in one way or another) terrible web apps that use 150 MB of javascript and data brokers.
The author claims it's surprising that AI uses less electricity than the rest of data centers. Why?? Data centers do a million different things other than AI
It's surprising because that's all anyone hears about right now.
Many are surprised to learn that data centers use much less water than golf courses or pools. We hear about it so much in the news it sounds like it's the overwhelming largest consumption problem around.
If they use so little water and power, why aren't they being built in the empty office buildings? Seems like commercial real estate was made for ... commercial activity.
This makes no sense? How is one even correlated with the other? Data centers obviously have much different requirements than office buildings for people. On a very small scale this works but not on a larger scale.
Nobody is saying that they use "so little water and power", just that the average person has no clue about the orders of magnitude at which anything outside of a private household uses these things, including things they previously didn't really have a problem with (like golf courses or almond farms)
Most people just hear <gigantic number of m³ or kWh> and their brain goes "damn, that must be a lot, i don't use nearly that much at home"
Water usage for a data center is nowhere near any agricultural use (which is subsidized in California so they aren’t incentivized to change much). Data centers don’t need water at all. It’s just the municipality offering cheaper way to cool by offering water cheaper than power for cooling.
Traditionally they have been. Several data centres in London are colocated in office buildings for example. I used to manage servers in one that most tenants wouldn't even have known about.
AI focused data centres also could, but require a higher amount of power and cooling per rack, so it's less practical. The point is not that they don't use more than a typical office, but that they don't use the world-ending amounts people assume they use.
The operators should still be made to pay for the externalities, and regulated to ensure they don't make life miserable for people around them.
But that doesn't mean we need to stop building them.
The data centers in L.A. are in old bank buildings. I used to locate my servers at One Wilshire.
Water is too expensive and wasteful to just run through pipes and send out to the ocean just for cooling. Power though… the reason bank buildings are loved for data centers is that they are on two power lines so if one goes out the other is available.
Our World in Data is great, but this article would've been improved if it offered a chart that offered a comparison to it's leading "1.5% of global electricity number".
For example: 7% to 10% of all global electricity is consumed by air conditioning. It's not hard to think of relatively small interventions that could save 1.5% of global energy use annually.
Increasing global electricity production by 1.5% (or 500 terawatt-hours) would also be a solution. Production from solar power increased by 600TWh from 2024-2025 and its rate of adoption is increasing as costs go down, so this doesn't seem like a problem worth worrying about.
Going by those figures, would mean that AI datacenters built so far, consume enough energy to undo almost a year of solar power buildout. AND growing like crazy. That's running costs. Not energy going into a building's construction or fabrication of the gear going in there.
I'm dismayed by the current political discourse on this topic. Seems like the two camps are 1) let mega corps build data centers with impunity or 2) use NIMBY tactics to oppose them in all cases.
I want to see a pro-growth approach that makes the mega corps assume the costs of their negative externalities. I want to see an increased build out of renewable and nuclear energy sources. I want to see environmental protections in place to limit impacts on aquifers and other water sources.
Instagram's algorithm is based on what you interact the most with (even stopping to watch the reel, to read comments or to write one counts as interacting). You can start marking all that content as "Not interesting" and in the worst case just blocking those profiles to fix your feed and algorithm (or even go to your settings and tweak your interests directly).
But uninstalling the app is never a bad choice regardless
One of the camps has all the money, hold all the power, and does indeed behave with impunity. The other camp has no power, has no way to affect anything, holds no political relevance, and NIMBY tactics is one of very few ways they can even hope to have at least some tiny chance of a minor change in their favor.
I think it does the discourse a disservice to take these two as equal.
There’s the pro-growth Abundance faction, but they’re captured by special interests, so they’re predictably advocating for supply-side YIMBYism and relaxed regulation, without addressing the political economy that led us to the current situation.
Seriously, it can't be that hard to build a datacenter and not piss people off. The economics demanded scale so fast that the capitalists don't care enough to thoughtfully plan this stuff out.
This is America, anything you do pisses people off[0]. The problem is that the rest of us listen to them because it is easier to understand one person's complaint than the distributed pain of electricity being a penny per kwh more expensive.
[0]I've seen people whinging about building solar plants in the desert, windmills off shore, data centers in silicon valley...
I am pro-AI datacenters. I just think they should all be built in Atherton and the Hamptons, with the stipulation they're only allowed to be built adjacent to the most expensive residential property in the area, in the way of its most impressive view.
Democracy isn't to blame, capitalism is. The mega corps have every incentive to dodge the costs of their negative externalities and they have them money to lobby for that. The "reasonable" citizen supporting growth with caveats will be drowned out by the mega corp and their mega wallets. The loud NIMBYs are the only group that have any chance of stopping the negative externalities which they achieve by blocking the growth. Your enemy isn't the NIMBY, your enemy is the mega corp, the NIMBY is your ally, they are the backstop.
It's less a fault of "democracy" and more a fault of partisan extremism where all dimensionality and nuance is removed, it's intentionally made into an "us" vs "them" issue with only two extreme sides, where either you're fully with "us" and everything we believe, or you're evil/horrible/etc. If you dare to fall in the middle somewhere, that's not allowed by the extremists on one side; you're lumped in with and seen as no different than the opposite extreme.
I think the problem is that "make them build responsibly" has the same outcome as "ban altogether" while being more effort on both sides.
It's too much easier and cheaper to drive one town over and bribe a local politician to let you pollute the river with impunity.
Maybe if we had some sort of Environmental Protection Agency or Federal Trade Commission that could enforce more broadly it would be a different story, and every hyperscaler DC could come with a guarantee of boosting clean energy and water reclamation. That would be a cool thing to have.
I presume running an electric shower includes heating up the water? Otherwise I don't understand how it could be so energy intensive. The energy is the mgh of the water lifted, plus the heat you put into it. Heat tends to be by far bigger in that calculation unless I've missed something.
We all just look at energy as a price. I think more people are concerned if our prices will go up, not how much energy it uses. The bigger question is water.
How much water do they use? Is it potable to begin with? Is it usable after? Will they drive prices up. These are one time builds with minimal jobs created after for a tech that could end my career. If I lose my job, I don’t also want to have no water and higher bills.
The talk about AI water usage exists because in Karen Hao’s book Empire of AI, she got water usage wrong by three orders of magnitude.[1] For data centers that use water cooling, it's almost always non-potable water, and it's usually returned to the body of water it was pumped out of (though slightly warmer).
If you're worried about water consumption, things like golf courses and farm subsidies for water-intensive crops are much bigger issues.
> For longer queries or requests that rely on AI agents or reasoning, the footprint is higher. Epoch AI estimated that a long query (7,500 words of input) could consume about 2.5 Wh, and a very long one (75,000 words) could consume 40 Wh. Those are far higher than the simple text query.
One thing that's missing from almost all of these energy estimates is the impact of "agentic" queries, which were rare a year ago and increasingly common today.
I've asked questions of GPT-5.6 Pro that take 19 minutes to return an answer, as the model runs hundreds of search and site navigation tool calls (and a bit of executed Python code too) trying to get to a credible answer.
I'm confident that 19 minutes involves a lot more than 75,000 words worth of tokens.
With coding agents we can at least get an accurate token count. I'm already at 204,000 tokens of GPT-5.6 Sol today and I haven't started work yet.
I ran a couple of prompts to reverse-engineer my ~/Library/Application Support/Firefox/Profiles/*/sessionstore-backups/recovery.jsonlz4" file - Codex figured out how to lz4 decompress the file, then I had it interpret some base64 encoded binary data in the JSON, that session is here: https://gisthost.github.io/?986cb066dd8502cf0200a5cc6fdeb4c6
With the new Ultracode modes in Claude Code and Codex it's been taken to the next level. I mean, multi agent systems have been available for a long time, but the newest models seem to be much more RL'd for them.
Both GPT-5.6 and Fable will happily run 3+ hours straight off a relatively simple goal in that mode, burn through millions of tokens in the process.
Not saying it's great bang for your buck, but just that those modes are there and being pushed by the companies.
Something I havent heard alot of people talk about is meat. I dont think you can be against data centers, but also buy cheap meat. I understand food != compute, and thats a fact. BUT the same people who are protesting against AI data centers are the same people who fly every year, eat steak in every other dinner, and generally indulge in activities that are bad for the planet (and theres no acknowledgement).
and flying. beef and flying are the two most impactful personal choices you can make (nearly everything else has a 'over a decade'/difficult-life-tradeoff factor). beef and flying are easy.
If you make such claims please back this up with evidence? I, for one, would expect a strong correlation between skepticism against AI data centers, vegetarianism and the preference for alternative travel methods to flying. Your comment sounds a lot like whataboutism to me.
> I, for one, would expect a strong correlation between skepticism against AI data centers, vegetarianism and the preference for alternative travel methods to flying.
And then you'd be equally wrong as the commenter you replied to. Without data it's just speculation.
ch4s3 | 5 hours ago
DenisM | 5 hours ago
* 5% of electricity in the United States is used to power data centers. For AI-focused ones specifically, it’s probably around 2%.*
I think it’s massive.
gruez | 4 hours ago
So what about the other 3%? That's not "massive"? Or it gets a pass because it's not AI? The whole debate over datacenter resource consumption feels like a rorschach test. Either you think the technology is transformative and therefore the resource consumption is justified, or you think it's slop and therefore any resource consumption can't be justified.
dundarious | 4 hours ago
I also don't think it's particularly conclusive about anything though. It doesn't on its face make me think anything like "it must be stopped", but it's remarkable.
sroussey | 4 hours ago
AI is hardly used at all? Basically every company and consumer in the country is using it, transitively if not directly.
dundarious | 4 hours ago
sroussey | 4 hours ago
Or reading almost anything these days. :(
Advertising they see. Using a credit card, writing a check.
Beam forming wireless signals.
It’s everywhere. Not all AI is an LLM.
dundarious | 3 hours ago
sroussey | 3 hours ago
Who exactly is the 2% figure? Can’t all be HN people on side projects using Claude code.
dundarious | 3 hours ago
As for capex implying high usage, look at xAI's capex and how it turns out most of it is now being rented out to Google, etc. High capex does not imply capex that has already found its customers and margins. Meta is an extraordinarily bad example to use, given the failure of the whole metaverse/VR thing that is now their name, but is completely vestigial.
Have a good afternoon.
0xffff2 | 2 hours ago
gruez | 4 hours ago
Speak for yourself. For "most people", "streaming" is basically TV for them, and the average american watches 2.6 hours per day[1]. It's probably safe to say that at least half of that is "streaming" of some sort, even more if you include other types of "streaming" like music streaming. Moreover nowadays even if they're not plopped in front of a TV, they're watching short form videos.
[1] https://www.bls.gov/news.release/atus.nr0.htm
dundarious | 4 hours ago
gruez | 4 hours ago
dundarious | 3 hours ago
> [AI] that most people hardly use at all
and while ambiguous, I meant that quantifying the number of people that use it, results in a not large number. I've had discussions in other threads about how to talk about transitive use, etc., so as with everything at all times, there are certainly ways to skin the cat that justify or invalidate the statement. One has to rely upon good judgment to decide which way is best, and reasonable people may disagree.
DenisM | 4 hours ago
How did you read that in my comment?
everdrive | 4 hours ago
eddyg | an hour ago
https://www.andymasley.com/visuals/ai-prompt-footprint/
lopis | 5 hours ago
StrauXX | 5 hours ago
lopis | 5 hours ago
Aurornis | 4 hours ago
Many are surprised to learn that data centers use much less water than golf courses or pools. We hear about it so much in the news it sounds like it's the overwhelming largest consumption problem around.
axus | 4 hours ago
rick_dalton | 4 hours ago
fxwin | 4 hours ago
Most people just hear <gigantic number of m³ or kWh> and their brain goes "damn, that must be a lot, i don't use nearly that much at home"
sroussey | 4 hours ago
paulhebert | 3 hours ago
(I think golf courses - and some pools - are quite wasteful and would support tax incentives around them.)
sroussey | 3 hours ago
vidarh | 4 hours ago
AI focused data centres also could, but require a higher amount of power and cooling per rack, so it's less practical. The point is not that they don't use more than a typical office, but that they don't use the world-ending amounts people assume they use.
The operators should still be made to pay for the externalities, and regulated to ensure they don't make life miserable for people around them.
But that doesn't mean we need to stop building them.
sroussey | 4 hours ago
Water is too expensive and wasteful to just run through pipes and send out to the ocean just for cooling. Power though… the reason bank buildings are loved for data centers is that they are on two power lines so if one goes out the other is available.
raychis | 4 hours ago
xnx | 4 hours ago
For example: 7% to 10% of all global electricity is consumed by air conditioning. It's not hard to think of relatively small interventions that could save 1.5% of global energy use annually.
ggreer | 4 hours ago
paulhebert | 3 hours ago
The more our energy use increases the harder it is to do so.
RetroTechie | 2 hours ago
Does seem like a problem worth worrying about.
shipman05 | 4 hours ago
I want to see a pro-growth approach that makes the mega corps assume the costs of their negative externalities. I want to see an increased build out of renewable and nuclear energy sources. I want to see environmental protections in place to limit impacts on aquifers and other water sources.
Democracies don't handle nuance well.
jjulius | 4 hours ago
Especially when modern discourse revolves around 140-character hot takes.
bakies | 4 hours ago
penr0se | an hour ago
But uninstalling the app is never a bad choice regardless
runarberg | 4 hours ago
I think it does the discourse a disservice to take these two as equal.
fellowmartian | 4 hours ago
bakies | 4 hours ago
thereisnospork | 4 hours ago
[0]I've seen people whinging about building solar plants in the desert, windmills off shore, data centers in silicon valley...
palmotea | 4 hours ago
jagged-chisel | 3 hours ago
palmotea | 2 hours ago
I like it!
reticulates | 4 hours ago
thegrim33 | 4 hours ago
barbazoo | 3 hours ago
Or are those just the two ends of the spectrum? I don't know anyone on either extreme of the spectrum.
glaslong | 18 minutes ago
It's too much easier and cheaper to drive one town over and bribe a local politician to let you pollute the river with impunity.
Maybe if we had some sort of Environmental Protection Agency or Federal Trade Commission that could enforce more broadly it would be a different story, and every hyperscaler DC could come with a guarantee of boosting clean energy and water reclamation. That would be a cool thing to have.
DemocracyFTW2 | 11 minutes ago
Failing democracies don't handle nuance well.
lordnacho | 4 hours ago
shipman05 | 4 hours ago
Our World in Data is based in the UK. Electric showers like this one https://www.mirashowers.co.uk/p/showers/shop-electric-shower... are more common there than they are in the US.
chrisgd | 4 hours ago
How much water do they use? Is it potable to begin with? Is it usable after? Will they drive prices up. These are one time builds with minimal jobs created after for a tech that could end my career. If I lose my job, I don’t also want to have no water and higher bills.
ggreer | 4 hours ago
If you're worried about water consumption, things like golf courses and farm subsidies for water-intensive crops are much bigger issues.
1. https://blog.andymasley.com/p/empire-of-ai-is-wildly-mislead...
simonw | 4 hours ago
One thing that's missing from almost all of these energy estimates is the impact of "agentic" queries, which were rare a year ago and increasingly common today.
I've asked questions of GPT-5.6 Pro that take 19 minutes to return an answer, as the model runs hundreds of search and site navigation tool calls (and a bit of executed Python code too) trying to get to a credible answer.
I'm confident that 19 minutes involves a lot more than 75,000 words worth of tokens.
With coding agents we can at least get an accurate token count. I'm already at 204,000 tokens of GPT-5.6 Sol today and I haven't started work yet.
DenisM | 4 hours ago
simonw | 4 hours ago
helloplanets | 4 hours ago
Both GPT-5.6 and Fable will happily run 3+ hours straight off a relatively simple goal in that mode, burn through millions of tokens in the process.
Not saying it's great bang for your buck, but just that those modes are there and being pushed by the companies.
Abh1Works | 4 hours ago
cagenut | 4 hours ago
barbazoo | 3 hours ago
Just stop. If you fund yourself starting a sentence that way, just close the browser.
DWe1 | 3 hours ago
RetroTechie | 3 hours ago
And then you'd be equally wrong as the commenter you replied to. Without data it's just speculation.
eddyg | an hour ago
kreyenborgi | 56 minutes ago
DenisM | 3 hours ago
This puts a number to my intuitive feeling that things move an order of magnitude faster this time.