So-called artificial intelligence (AI) is all the rage right now between your grandma asking ChatGPT how to code in Python or influencers making videos without having to hire extras, but one growing concern is where the power is going to come from for the data centers. The MIT Technology Review team did a deep dive on what the current situation is and whether AI is going to kill us all (with carbon emissions).
Probably of most interest to you, dear hacker, is how they came up with their numbers. With no agreed upon methods and different companies doing different types of processing there were a number of assumptions baked into their estimates. Given the lack of information for closed-source models, Open Source models were used as the benchmark for energy usage and extrapolated for the industry as a whole. Unsurprisingly, larger models have a larger energy usage footprint.
While data center power usage remained roughly the same from 2005 to 2017 as increases in efficiency offset the increase in online services, data centers doubled their energy consumption by 2023 from those earlier numbers. The power running into those data centers is 48% more carbon intensive than the US average already, and expected to rise as new data centers push for increased fossil fuel usage, like Meta in Louisiana or the X data center found to be using methane generators in violation of the Clean Air Act.
Technology Review did find “researchers estimate that if data centers cut their electricity use by roughly half for just a few hours during the year, it will allow utilities to handle some additional 76 gigawatts of new demand.” This would mean either reallocating requests to servers in other geographic regions or just slowing down responses for the 80-90 hours a year when the grid is at its highest loads.
If you’re interested in just where a lot of the US-based data centers are, check out this map from NREL. Still not sure how these LLMs even work? Here’s an explainer for you.
Nothing is going to kill us all. I can’t believe people really don’t get embarrassed by predictions of the apocalypse… As if you don’t still sound crazy just because you put “AI” in there instead of idunno demons or whatever.
Just stop screwing around with big dumb pinwheels and solar panels and hook up the datacenters to a bunch of modernized nuclear plants. AI isn’t going anywhere, massive computation isn’t going anywhere, and we have a solution that fixes it all… Why aren’t we using it? At a certain point it starts to feel a bit suspicious. Why is all the world still burning a bunch of coal in the 21st century like it’s the 19th, when we have nuclear energy? And we’re messing with windmills? What are we, Dutch?! And all the combined nuclear disasters have killed about as many people as choke to death on rubber duckies. Makes no sense.
Is the world gonna end? Is human life doomed? Okay cool let’s use this solution we have that works right now as a complete system, and doesn’t need decades of R&D and thirteen other complementary technologies added to cover “night time” and “not a windy day.” Let’s do that. Let’s fix it and move on, eh?
Well said
If we had the political will to get off of fossil fuels we would’ve switched to renewables with storage already, since they’re much faster to build than nuclear and even cheaper. The barriers to solving the fossil CO2 problem aren’t technological.
I also expect that AI as we currently know it will mostly go away once investors tire of throwing money at an unprofitable enterprise with only lousy sci-fi promises to show for it, and with cryptocurrency mining mostly in the rearview, the need for massive(ly wasteful) computation as we’ve known it for the past decade or so will go away as well.
That is what they always say in theory, but in practice it hasn’t borne fruit. There are caveats and hidden charges with renewables so far, which I have mentioned, and which FREQUENTLY feature here in articles on HaD. But multiple western countries and even countries in the rest of the world have had heavily-nuclear grids in the recent past, without excessive cost or danger, and they recently dismantled them because of panicky nonsense. All the damage at Fukushima was water damage. The tidal waves really messed up the coast. But human and environmental cost of the nuclear plant? Virtually nil. Far more have been harmed by coal and gas every year since then, and if the apocalyptic importance of switching NOW is to believed (I believe it personally, to some extent) then the choice is clear. We do nuclear the way we did in the late 20th century and maybe work on renewables once we have a bit more airstrip and time to work with.
Exactly. That’s what I’ve been saying, but probably not the way you imply. We have nuclear worked out fine, but it has political hurdles… However renewables have huge technological problems thus far (not to say they are forever-doomed, but they need work) and at the moment the focus on those rather than nuclear seems to refer back to:
Nuclear is something which must be subsidized by the state. It’s not very attractive commercially. What IS attractive commercially is spinning up huge amounts of (very carbon-hungry!) industry to build windmills and whirligigs and huge battery arrays and solar panels that halve in efficiency every five years and need to be replaced at cosmic scales. There IS a green-industrial complex forming. Already going. This can be true while also admitting the importance of green technology.
I tend to agree, I think that the hype is overblown and another AI winter will come (despite a lot of it being truly impactful so far)… However, extremely compute-intensive work is NOT going away, even if we have another AI crash. The future will have a lot of computation happening at scale for one reason or another.
Apparently nuclear plants and fuel just grow on trees.
Not true at all, the solar stuff now lasts decades without really declining meaningfully – usually a bit of steep decline in the first year down to something close but often still over the sticker rating and then decades of creeping down so slowly you don’t have to replace them at all. HOWEVER they are being replaced rather rapidly as the performance of new panels keeps on getting better and political will and funding to create new solar arrays is constantly lagging behind – so it is more profitable to scrap panels that have well over 90% of their rated output left, and would stay that way for many years as the newer panel either works better in lower lighter, has higher peak power, has some newer coating that makes it better at selfcleaning etc. As its just too costly and impossible to put the new better panels somewhere else.
Much the same thing with Wind turbines too, a great deal of those have been replaced because making an existing turbine larger and better is possible, getting planing permision and funding to build a whole new windfarm on the other hand…
I do agree Nuclear power should be used more though, its a technology that has been ready, reliable and very very safe (if you aren’t more interested in making weapons than electricity) for many many decades. And being ‘small’ central nodes of immense power is a good drop in replacement for the coal and gas powerstations, and thus if you can site your new nuclear generator at the same sort of the location you can save hugely in the infrastructure required.
“If we had the political will to get off of fossil fuels we would’ve switched to renewables with storage already, since they’re much faster to build than nuclear and even cheaper. ”
When nuclear plants were being built in the 60s/70s where was renewable tech? How efficient were solar panels? How was battery storage?
Imagine if we’d continued to invest in nuclear technology, we would potentially have fusion now, and certainly liquid metal reactors, small modular ones, low cost power; and all this green washing “renewable” nonsense would have fallen by the wayside.
I have a theory that nuclear reactors aren’t a good fit for datacentres because their load varies throughout the day. Latency encourages you to use the nearest datacentre along with everyone else around you so all the virtual desktops demand in New York should end up in the same datacentre and in use roughly 9-5. Video streaming should be coming from a local cache to reduce data transfer costs and that will be used heavily when people get home. Even calls to AI services the article mentions will be more frequent in the day than at night.
So what about sizing nuclear to cover the daytime maximum? Then overnight you can export power at a low demand time when it’s not valuable to do so.
Just like our grids nuclear can provide the base load cost effectively but the rest could be better served by something else like stored renewables or fossil fuel.
This is just a theory though, I need to check spot pricing on AWS EC2 for patterns. If it’s true maybe our daytime data processing that isn’t latency sensitive should be taking place in another country’s night time.
Big Tech Backs Small Nuclear
Google and Amazon invest in small modular reactors to power data centers
10 Dec 2024
I’m sure they do to cover the base load. That 40% of capacity they will always consume even overnight same as grid scale consumption. It doesn’t change the fact that building a nuclear power plant to use for 10 minutes a day isn’t cost effective.
“… should be taking place in another country’s night time.”
Is this true? The equipment is not cheap. Perhaps utilization is mandatory to recoup costs, like an airliner?
Yeah, it’s not there are people in towns and cities begging and protesting to put a stop to the use of the AI data centers due to the actual pollution that’s slowing killing the residents. /s
We’ve got people begging companies to stop doing 5G roll out because it’s causing them health issues and killing them.
What’s you’re point?
You mean the same people who were told 5G caused COVID versus a set of people who are actively – with proof – being poisoned because companies want the profits but bot to oay workers for labor?
“not to pay” = ” bot to oay”
My life for a damn edit button!!!
Unless it’s getting colder (isn’t looking like it is) Nuclear Powerplants) are getting shut down, like now in France and Switzerland. Doesn’t sound like a solution to me
Because nuclear is very expensive compared to solar and wind (which are cheapest and quickest to market), they require fuel extraction, refining and transportation and also have their own waste problem much like carbon fuels. Solar and wind also have resource extraction related costs in terms of raw material extraction, refining and manufacturing, but minimal upkeep and no fuel cost.
It is worth noting that not all uses of ML or NN run on massive server farms. Some run on your computer and do not require internet, and therefore use no more power than the maximum draw of your computer’s power supply. It’s also worth noting that not all ML uses scraped/unconsensually shared data.
I’m all for energy efficiency, and I also happen to hate AI, but still. Energy demand is not a good reason to not do things. If we refuse to move forward with technology because of the energy usage, we’re doomed to stay on this rock until nuclear winter or meteor strike.
Energy demand can be a good reason not to do things if the demand:utility ratio is bad enough. Compare using energy to run pumps for a municipal water system vs. sending a few celebrities on a joyride almost into space for example. Using it to train and run stochastic parrots that have vanishingly few practical uses (separating investors from their money indefinitely doesn’t count) is on the “suborbital amusement park ride” side of the spectrum.
On the topic of going into space, I expect that the laws of physics as we currently know them, the distance between stars, and sane risk management practices will keep us stuck on this rock for the foreseeable future. We should care for it accordingly.
NEO detection & defense, nuclear disarmament, and building outposts near Earth’s poles are a lot easier, less resource-intensive and more humane than trying to build a self-sustaining colony on an airless radiation-scorched toxic dustball, red or gray.
We’ll be stuck on this rock due to Starlink debris.
Starlink may expand to 42,000 satellites. The Chinese are also planning a massive 12,000 satellite network. Amazon’s Project Kuiper plans a 3,236 satellite network.
Starlinks are in low earth orbit, ranging from 340km to 550km up. Debris there will clear in years to decades due to atmospheric drag. While a Starlink collision cascade could make low earth orbit unusable for a while (a very, very bad thing), it won’t be what keeps us stuck here.
They were saying we have 12 years until we’re doomed… and that was 6 years ago. How’s that going?
i’d say we’re ahead of schedule?
When I was in grade school we were told we were headed into an ice age…. Then we were told nyc and florida would be underwater by the year 2000. If they really wanted us to believe in climate change they would be less hyperbolic and more realistic.
Also they always say weather is getting worse due to climate change but the us forest service says there aren’t more wildfires than the past and noaa tells us that hurricanes aren’t any stronger and we get the same average yearly precipitation. Then the dhs tells us there are nearly 400 arson cases every year by environmental activists.
Why can’t anyone be upfront and honest about what climate change really means?
seems like you are relying on low quality information sources. of course you can cherry pick something (or let someone else cherry pick something) and then over-generalize (or let someone else over-generalize) and hand wave (or stare blankly at someone else’s waving hands) and then say that no one knows anything or no one is saying anything true. but i prefer to dig, and for that, the tradition of scientific publishing is pretty good.
so…this 1981 article is pretty upfront and honest, and imo it hasn’t really been obsoleted at all:
https://www.giss.nasa.gov/pubs/docs/1981/1981_Hansen_ha04600x.pdf
1) we can all look at how accurate climate models such as rcp8.5
2) Al gore in 2006 – nyc and florida will be underwater by 2013… Vice president. and this is who is controlling environmental policy?
3)I thought noaa.gov was high quality information? “There is no strong evidence of century-scale increasing trends in U.S. landfalling hurricanes or major hurricanes. Similarly for Atlantic basin-wide hurricane frequency”
https://www.ncei.noaa.gov/access/monitoring/wildfires/ytd/0
https://cpo.noaa.gov/increasing-wildfire-area-burned-amid-a-fire-deficit-in-north-american-forests/
It seems like the media/pllicymakers are saying one thing while the science says another
4) Gonna have to admit I was wrong about the arson. It is about 20-30 arsons and bombings from just 2 evnvironmental/animal rights groups.
https://www.dhs.gov/sites/default/files/publications/OPSR_TP_TEVUS_Bombing-Arson-Attacks_Environmental-Animal%20Rights-Extremists_1309-508.pdf
haha when i said “you’re relying on low quality information sources”, i did not mean “citations needed.” i have no doubt of the abundance of low quality information out there. you don’t need to tell me “Al Gore was wrong.”
since you expressed a desire for upfront and correct information, i gave you some. but i am thinking, maybe you don’t want correct information, you’d rather debunk low quality information.
there’s plenty of it out there for you to debunk so i wish you happy hunting!
If your solution to global power problems is not nuclear you are wasting everyone’s time.
Agreed, use nukes to decrease population and problem is solved, bwahaha:)
“We don’t know who struck first, us or them. But we do know it was us that scorched the sky”.
In Europe we are currently shuting down nuclear power stations or at least reducing their output due to cooling problems due to lack of or too warm water. For every thermal plant regardles of fueling you need cooling to keep the carnot cycle efficient… Think also of that before promoting nuclear power as a general solution!
IMHO, not AI, though it sure IS one of the contributing factors.
“We’ll go down in history as the first society that wouldn’t save itself because IT WAS’T COST-EFFECTIVE” (Kurt Vonnegut). Because it did not generate short-term profit, so the hell with all the solutions that exist already, profit is paramount.
Translation – presently quite a LOT of our tax money (aside from the needed/critical stuffs) goes down the drain financing all kinds of politicians’ pet/entertainment projects of mostly no merit to the average Sam. Political Circus we cannot leave and are stuck with.
Since HaD just loves to censor anyway, how about we first fix HaD by getting rid of the nuclear industry lobbyist eh.
Alternatively, how about you ask a cut and finance HaD – and a small country as well – from the profits?
Nuclear is clean and safe. What’s your problem?
It is neither … It is also not reliable (unexepected downtimes due to “suddenly” required maintenance, downtimes due to lack of cooling during summer), and it is expensive (building, maintenance, destruction, waste disposal).
None of these are unique to nuclear. All infrastructure elements, including coal-fired power stations, need continuous monitoring and maintenance. In fact, if you have more nuclear power stations you have more redundancy and more resilience in the case of unexpected downtime (which, ironically, is totally expected). It’s also cleaner and safer than coal, for example, which emits more nuclear particles than any nuclear power station.
Wait until they discover how much energy does crypto mining draw.