AI energy draw from Chicago datacenters to rise ninefold US energy provider Exelon has calculated that power demand from datacenters in the Chicago area is set to increase ninefold, in more evidence that AI adoption is will put further strain electricity supplies. The utility giant revealed there are about 25 datacenter projects planned in the area around Chicago that would consume …
Both technologies, fossil and nuclear, have their own set of advantages and disadvantages.
Often the people being pro more nuclear see it as a replacement of fossil and most importantly coal. Reality may be a bit different. The current push of the datacenter and even more so AI warehouse owners is to add energy generation in order to cover their truly humongous needs for electricity. Chances are real that part of the cost for building those nuclear power plants will directly or indirectly be paid by taxpayers, be it in the form of direct subsidies for the power plant, extra investments in grid if needed, free government insurance in the unlikely case of a massive fallout, final storage of waste and given the likeliness of megacorps to duck costs the decommission of the power plant when their datacenter owner lets the subdivision go bankrupt just before the end of its profitable lifetime.
The question then is: do we chose to limit the combined amount of energy datacenters and AI warehouses can consume or do we let big tech companies of whom many have a strong history of showing very few respect for anti trust laws, taxation laws, privacy laws or even maintaining their own services and products for more then two years if they become unprofitable become owners and operators of nuclear power plants that only add copious amounts of power generation for few things other then the temporary benefit of the companies building them?
In addition, there is a hidden load. Big tech companies like MS are very strongly pushing the AI PC. I read here: the total energy consumption of AI training plus inference will be so massive that making that clear would raise very strong concerns and opposition. So these AI companies want to move the bulk of the energy consumption away from their bills and scrutiny: on the local computers of the clients.
If AI companies, for example, would build 20% of current global electricity generation capacity to satisfy their needs with both nuclear and renewables, they might claim that their business is sustainable. That however would be greenwashing, as it would hide the massive amount of energy and resources needed to build and operate their chips and equipment and the equipment needed to fill the semiconductor fabs, plus the massive and far larger amount of electricity needed to run / inference the AI models on the computers of the client. In that way, what optically seems like a net-zero contribution of AI (with even the promise it will help invent techniques to reduce CO2 emissions) could hide a massive global increase of unsustainable energy and resource use.
I think if nuclear is to have more wide acceptance in the general public, that it will need to go hand in hand with a strong global plan to minimize total (fossil, nuclear, renewable) energy consumption. Building more nuclear without a very strong plan only guarantees more nuclear power generation and its set of advantages and disadvantages, not an actual reduction in fossil energy generation. It could be argued that if no nuclear power plants are build to deal with this very quickly rising demand in power that that will lead to massive increases in fossil fuel consumption. I say that is only partially true: increasing power generation so quickly as what a major AI and datacenter boom demands with fossil would be near impossible, very impractical, so expensive and so controversial it would severely inflate the AI boom / hype.
A few misconceptions here:
1. IT, cloud infra or AI companies will never invest in nuclear energy. they will just contract PPAs. Maybe invest in storage solutions. But nuclear is simply off the table. It's not their core business or domain of expertise. Nuclear energy requires in incredibly large amount of very diverse engineering competences which are very thin on the ground. ROE is negative and very long term. Some of the largest giant industry leaders went bust, needed bailouts or are in bad shape: Westinghouse, Hitashi, Toshiba, Areva/Framatome. It's impossible to insure. Waste management is a nightmare.
2. Energy consumption decrease is a pipe dream. It will not happen either. Unless a big population decrease takes place. But self-managed energy consumption decrease is unlikely. Not the way mankind works.
The vast majority of waste from nuclear power plants (depleted rods) is stored on site at the nuclear power plant in isolated pools of water enshrouded in concrete which are underground. Nuclear power doesn't produce anywhere near the amount of waste you think it does. Look at a Google Maps view of any nuclear power plant and you'll see the structures used to house waste.
Are you one of these people that thinks waste from a nuclear power plant is billions of barrels of green glowing slime? It's not...the only waste is the fuel rods when when stood vertically take up much less space than you'd think.
As with any kind of waste the perception is worse than the reality...it is possible to recycle spent fuel rods, there are various battery technologies etc that have been developed using waste products from nuclear power plants...but the perception the general public has prevents any of them going mainstream...and of course, big business hates some of the products...because they last...for decades.
I don't see why local inference is a bad thing...if anything it's better than using a cloud service. Inference doesn't require much energy anyway...it's the training that uses the most energy and nobody has the kit at home to perform training to the level of OpenAI etc...it will be possible eventually train an AI at home comparable to AI being trained today, but it's decades away at this point.
Over longer time scales, AI won't actually be a business anymore, not the way we know it...AI will follow the same curve we see for all technology...eventually it will plateau and improvements will be incremental...like smart phones...we had a period of huge innovation (roughly 10-15 years ago) and we've reached a point where the improvements are incremental and don't really matter. How fast does a phone need to be to run WhatsApp etc? The problem for AI in the long run isn't power, it isn't hardware...it's data...there is only so much data that it is possible to collect and train an AI with...especially within reasonable time frames...most of the leg work has been done in that regard...digitising old content and data etc...that's been going on for decades...the barrier now, for the most part, is rights and ownership...unless you want to embark on creating an entirely new digitised knowledgebase...spending billions etc.
Power is not a problem long term, it's a problem now because of current hardware and it's limitations. In 10 years time, to reach the same amount of compute today you will need a fraction of the energy. As I say, we will reach a point where hardware is sufficiently powerful enough and the results from training are sufficiently good enough that improvements will become incremental at best...what we're more likely to see in the area that requires tons of power, is ever more specialised AI...but that won't be mainstream for the most part. Who on earth needs their own specialised rocket science AI at home? Not many folks I'd wager...but an AI that is adept at providing information on how to cook a healthy meal? Help them manage their grocery shop? Do their tax returns? Hell yeah...everyone...but there is a plateau with that kind of knowledge...there are no new frontiers for an AI to discover in basic accountancy, once a point has been reached...that's it...the difference between a 250b parameter model and a 500b parameter model for accountancy might be virtually nothing...but the energy requirement for the 500b one vs 250b might be massive.
Hinkley point C is 3.2GW; so 2x would be 6.4GW rather than just 5GW to be fair.
However, yes it's a stupid amount of power to be using for something that's not really required. The simple answer would be something like the cigarette taxes; tax the electricity bill to 500% to encourage a new interest in not wasting huge amounts of power. And if it is profitable at that sort of price then it will at least pay for the power plants and infrastructure required.
By that point it's too late, the climate is ruined even further. Better to prevent it happening in the first place by making them criminally responsible for their emissions.
It's a good place to start because the public will be accepting but I fear that we need to regard all energy use as theft from the future. That's not to say primitivism is needed but we have to start justifying every Joule consumed.
That's a lot of cooling needed in the summer months. I was there in the summer of '88, aka among some friends, the "Hot and Tired" summer. That extreme seems to now be the norm. Will the gear be able to create enough heat in the whole datacenter for Chicago's cold snaps?
To produce 5Gw electric from nuclear requires a few extra Gw of thermal* output from the reactors, that's quite a lot of spare heat being removed.
In fact, electrical generation is just what we happen to use the heat for, the raw heat could just as easily be directed into a city wide heating system.
*Hinkley-C is around 9Gw of heat to produce 6.2Gw electric.
We sure could use better ways to convert nuclear energy into electricity IMHO, more direct and less medieval than excitement → heat → steam → turbine → alternator → electricity. We need something smarter, maybe excitement → metamaterial → electricity. Getting there will require concentration, intellectual focus, and possibly genius as well, which last week's Higgs obit reminds us is likely to be in increasingly short supply.
At least there's hope of reducing the amount of dog turds on French sidewalks by having folks swap their organic quadrupeds for robodogs (but how will we power them all?)!
Let's consider the possible sources of useful energy.
Kinetic? Nuke to Kinetic? (Nope!)
Gravitational Potential? (Nope!)
Chemical? (Nope!)
Radiance / light? (Nope!)
Sound? (sub-division of kinetic anyway, and Nope!)
Nuclear? (we're trying to get the energy OUT of that form and into something more useful to us, so NOPE!)
Are you detecting a running theme here? Thermal isn't going anywhere. You can alter the material selections but water has so many advantages for so many reasons. It's not particularly hazardous, very well understood, dense and abundant in most places.
I'm with you on not wasting power on frivolities though screaming into the hurricane about it is a losing battle. Bitcoin, "AI", yada...
Too embroiled in the question of how to build a thing, did they stop to think if they should?
The very word Sabotage derives from the wooden shoes thrown in to stop the machines. The idea of what should be national scale projects being frittered away on datacentres is potty.
If I have become a tech luddite, so be it.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
