It looks like wishful thinking. There were reasons why USA for decades was buying Russian nuclear fuel and not building new nuclear power plants. Meanwhile, Russia and China were building them all over the world. Restoring competences and building lots of new stations in just 10 years seems not realistic.
The U.S. was buying Russian nuclear fuel because it comes from weapons grade uranium blended down to low enriched. The U.S. made a deal with Russia to reduce their stockpile of HEU and provide them with some cash. This distorted the market for uranium mining and enrichment in the U.S.
I doubt it will happen in 10 years, but I think there’s a good chance it will happen.
The economics of new nuclear improve significantly if you’re willing to build many at once, and repeal a few key low benefit but high cost regulations.
The power requirements of AI can only really be met with new baseload power. You don’t want to leave your expensive silicon idle when the sun isn’t shining, and while batteries might help with say to day fluctuations, they don’t solve the problems of seasonal cycles.
The French did not achieve exonomies of scale when they built out nuclear. And that was in a much friendlier regulatory environment than today.
Power is a massive component of the cost of AI. They might pay extra for 99.9% reliability, but are not going to pay significantly more for more 9's IMO.
Anybody can predict anything. This seems extremely unlikely, unless we magically start building a whole lot of plants today, and they somehow are able to be finished in 10 years. The author of this blog seems completely out of touch.
It's a measure of the failure of nuclear power that a blogger making wild predictions is estimating that the country with the most nuclear plants installed will be getting less than half their new power from it in a decade.
This seems to boil down to AI uses lots of power and some of that currently comes from nuclear so we’re going to definitely ramp up nuclear energy production. It in no way addresses the cost, regulatory, or public acceptance issues. It also completely ignores things like grid storage that could make renewables a viable alternative. I don’t know if nuclear energy production is going to increase and after reading this I’m absolutely no closer to a deeper understanding of the issue.
Nobody needs a kind of AI that consumes that much energy. There's no way that's commercially viable and carries its own weight. Also, the whole discussion is taking place in the context of present day GenAI where nobody has found a marketable use case for it. This whole bubble would be long bust by the time construction puts its boots on the ground.
A better justification for more nuclear energy is to electrify transportation and residential heating/cooling while navigating the climate catastrophe, don't you think?
> Nobody needs a kind of AI that consumes that much energy.
Rather: nobody needs to spend that much on current quality AI.
Say for the sake of argument that the accelerationists are correct about the timescale for future capabilities (I hope they're wrong in the short term, but run with it):
If we had an AI that could drive robots — cars, androids, surgical machines, industrial equipment — at least as well as any human, but the compute meant that it took 10 kW to do so, at electricity prices of $0.1/kWh that AI is cost-competitive with $1/h human labour.
If that happens and there's enough compute but not enough energy, energy prices rise to match human wages (so at least a factor of ten for this arbitrary example).
To avoid energy price rises, and for current economic output, that would require the USA to have about 170 million (workforce) * 10 kW = 1.7 TW ≈ 3.9 times current US electrical production.
(For the avoidance of doubt, 10 kW is a completely made up number with no real justification beyond it being simultaneously a relatively small number for an isolated valuable industrial device while also much more than a human brain).
> A better justification for more nuclear energy is to electrify transportation and residential heating/cooling while navigating the climate catastrophe, don't you think?
Not really, the reason for nuclear rather than renewables is more about the relative cost compared to time-shifting power generation with storage, and (most*) transportation already needs storage.
Heating and cooling is likewise, with good insulation, something that can work fine as time-shifting — the apartment I own back in the UK has storage heaters, everywhere I've lived (outside university) has/had hot water tanks.
The other reason for nuclear reactors is so you can maintain a nuclear arsenal, which sadly seems to be the only way for nations to stop other nuclear powers from messing with them.
* trains and trams can get rail/overhead line power, but cars, aircraft can't do that; trucks can be adapted that way, but only to reduce storage not eliminate it and even then it's just experimental**; shipping technically can use nuclear but ports don't want to accept such vessels and nobody wants pirates stealing nuclear reactors
Perhaps the biggest issue isn't public acceptance, but simply we stuck at building fission plants, even when we intend to.
I was pro nuclear power sixteen years sixteen years ago (these days I'm more ambivalent). In that time the UK has built maybe three quarters of a nuclear power plant.
Nuclear waste sucks but carbon in the atmosphere is way worse. But we're incapable of increasing nuclear production. Meanwhile renewables have gotten way cheaper and storage is beginning to look feasible.
At the risk of looking like [0] guy, i suggest this is incredibly far from the truth.
First of all, because building a new nuclear plant in 10 years is all but impossible. Therefore, since there isn't many of them under construction now already, not many will be built within 10 years regardless of demand.
Secondly, if demand comes from AI, it is used for training AI models and is thus not very time-sensitive. If those data centres can buy electricity very cheaply, they will happily do it half of the time, especially as GPUs wear out from constant use, and have a longer "moral" lifespan due to obsolescence than physical, so they won't be losing much from keeping them off for part of the day if that buys them much cheaper power - and buying it NOW, not in 10 years. In this case, AI can even help deployment of renewables because they can feed extra "free" electricity during peak production hours, to the data centres, getting something rather than nothing for it, thus reducing "duck curve" problem of curtailment.
It looks like wishful thinking. There were reasons why USA for decades was buying Russian nuclear fuel and not building new nuclear power plants. Meanwhile, Russia and China were building them all over the world. Restoring competences and building lots of new stations in just 10 years seems not realistic.
The U.S. was buying Russian nuclear fuel because it comes from weapons grade uranium blended down to low enriched. The U.S. made a deal with Russia to reduce their stockpile of HEU and provide them with some cash. This distorted the market for uranium mining and enrichment in the U.S.
I doubt it will happen in 10 years, but I think there’s a good chance it will happen.
The economics of new nuclear improve significantly if you’re willing to build many at once, and repeal a few key low benefit but high cost regulations.
The power requirements of AI can only really be met with new baseload power. You don’t want to leave your expensive silicon idle when the sun isn’t shining, and while batteries might help with say to day fluctuations, they don’t solve the problems of seasonal cycles.
The French did not achieve exonomies of scale when they built out nuclear. And that was in a much friendlier regulatory environment than today.
Power is a massive component of the cost of AI. They might pay extra for 99.9% reliability, but are not going to pay significantly more for more 9's IMO.
Why do they improve, ie. what are the relevant significant cost factors, and why were those low benefit regulations written in the first place?
Anybody can predict anything. This seems extremely unlikely, unless we magically start building a whole lot of plants today, and they somehow are able to be finished in 10 years. The author of this blog seems completely out of touch.
It's a measure of the failure of nuclear power that a blogger making wild predictions is estimating that the country with the most nuclear plants installed will be getting less than half their new power from it in a decade.
This seems to boil down to AI uses lots of power and some of that currently comes from nuclear so we’re going to definitely ramp up nuclear energy production. It in no way addresses the cost, regulatory, or public acceptance issues. It also completely ignores things like grid storage that could make renewables a viable alternative. I don’t know if nuclear energy production is going to increase and after reading this I’m absolutely no closer to a deeper understanding of the issue.
Nobody needs a kind of AI that consumes that much energy. There's no way that's commercially viable and carries its own weight. Also, the whole discussion is taking place in the context of present day GenAI where nobody has found a marketable use case for it. This whole bubble would be long bust by the time construction puts its boots on the ground.
A better justification for more nuclear energy is to electrify transportation and residential heating/cooling while navigating the climate catastrophe, don't you think?
> Nobody needs a kind of AI that consumes that much energy.
Rather: nobody needs to spend that much on current quality AI.
Say for the sake of argument that the accelerationists are correct about the timescale for future capabilities (I hope they're wrong in the short term, but run with it):
If we had an AI that could drive robots — cars, androids, surgical machines, industrial equipment — at least as well as any human, but the compute meant that it took 10 kW to do so, at electricity prices of $0.1/kWh that AI is cost-competitive with $1/h human labour.
If that happens and there's enough compute but not enough energy, energy prices rise to match human wages (so at least a factor of ten for this arbitrary example).
To avoid energy price rises, and for current economic output, that would require the USA to have about 170 million (workforce) * 10 kW = 1.7 TW ≈ 3.9 times current US electrical production.
(For the avoidance of doubt, 10 kW is a completely made up number with no real justification beyond it being simultaneously a relatively small number for an isolated valuable industrial device while also much more than a human brain).
> A better justification for more nuclear energy is to electrify transportation and residential heating/cooling while navigating the climate catastrophe, don't you think?
Not really, the reason for nuclear rather than renewables is more about the relative cost compared to time-shifting power generation with storage, and (most*) transportation already needs storage.
Heating and cooling is likewise, with good insulation, something that can work fine as time-shifting — the apartment I own back in the UK has storage heaters, everywhere I've lived (outside university) has/had hot water tanks.
The other reason for nuclear reactors is so you can maintain a nuclear arsenal, which sadly seems to be the only way for nations to stop other nuclear powers from messing with them.
* trains and trams can get rail/overhead line power, but cars, aircraft can't do that; trucks can be adapted that way, but only to reduce storage not eliminate it and even then it's just experimental**; shipping technically can use nuclear but ports don't want to accept such vessels and nobody wants pirates stealing nuclear reactors
** Literally the only example I've encountered, quick google suggests it's still only in testing: https://youtu.be/_3P_S7pL7Yg?si=IWKWIVAKbjHD0eL2
Perhaps the biggest issue isn't public acceptance, but simply we stuck at building fission plants, even when we intend to.
I was pro nuclear power sixteen years sixteen years ago (these days I'm more ambivalent). In that time the UK has built maybe three quarters of a nuclear power plant.
Nuclear waste sucks but carbon in the atmosphere is way worse. But we're incapable of increasing nuclear production. Meanwhile renewables have gotten way cheaper and storage is beginning to look feasible.
How on earth did a blog post with five paragraphs devoid of any content get voted to the front page?
So to be clear this is just a blog throwing out unfounded predictions?
Correct
At the risk of looking like [0] guy, i suggest this is incredibly far from the truth.
First of all, because building a new nuclear plant in 10 years is all but impossible. Therefore, since there isn't many of them under construction now already, not many will be built within 10 years regardless of demand.
Secondly, if demand comes from AI, it is used for training AI models and is thus not very time-sensitive. If those data centres can buy electricity very cheaply, they will happily do it half of the time, especially as GPUs wear out from constant use, and have a longer "moral" lifespan due to obsolescence than physical, so they won't be losing much from keeping them off for part of the day if that buys them much cheaper power - and buying it NOW, not in 10 years. In this case, AI can even help deployment of renewables because they can feed extra "free" electricity during peak production hours, to the data centres, getting something rather than nothing for it, thus reducing "duck curve" problem of curtailment.
[0] https://external-preview.redd.it/nv9Rv16O4tFC4rWm443AW3oBdyK...
That reminded me of Information Batteries (2021) https://raghavan.usc.edu/papers/infobatteries-eir21.pdf
I hope it's not just US but the world!