Thanks for the article. I'm wondering whether it makes sense to lump all AI together here though — the increase in power is mainly coming from GenAI, but the energy-saving benefits seem to be in other areas of Deep Learning. I can't really see how GenAI ends up as a net positive on the climate now (which ofc doesn't necesarilly mean it's worth it to stop using it).
Excellent point. My impression is that GenAI's having a lot of spillover effects in other areas of AI, like pouring massive amounts of money into developing better GPUs etc. and bringing more researcher talent into the space, but definitely agree that I'd be surprised if chatbots were having much of a positive climate effect specifically. Busy for the next few days but would like to look into the spillover effects more. If they're not there this weakens my case.
The (possibly overhyped) early signs of AI contributing to scientific research are all based on GenAI, and that's still what the AI industry is betting on. OpenAI's IMO gold medal-placing model was notable for not just being GenAI - all the models in this space are - but for being an LLM, ultimately a chatbot using a strange and curt form of language. Depending on your definition, even AlphaFold is generative.
Disease-curing AI and anime-fanfic-writing AI mostly differ in their application. It's more about how much energy is put into each use than how much energy is put into the research. The research benefits all.
It also seems like Jevons' paradox is very unlikely to apply to AI, at least in terms of power efficiency. Power is only ~10% of the cost of GPU compute, so as far as price signals go it's pretty minor. So Jevons' paradox would only apply to the extent that people are curtailing usage due to carbon footprint concerns, which I expect is a pretty small share of people
Thank you for this perspective! While it is possible for a net good, what about the individual communities facing new environmental issues because of where the data centers are being built?
Definitely need to take things on a case by case basis, I don't support any and all data center buildouts. I'm pretty skeptical about water issues (https://andymasley.substack.com/p/i-cant-find-any-instances-of-data) but seems like a few specific data centers have polluted local communities. Definitely need strong enviro regs, seems like some of the big labs have had a lot of undue influence over local governments.
Thank you - this could have been a really dense explanation but you’ve made it accessible even to a curious non academic like me. Important topic too. Bravo
Thanks for writing the article! I do want to note on a point of technicality - your example on solar panel is not true Jevon's paradox.
A better example would be due to drop in solar price => we end up consuming 1000x times more energy, which means that while the marginal emission for solar is small, the total emission would still be significant.
I think your main point would still stand though - we care *on net* if something prevent more emission than it creates.
Does it make sense to say though that because the energy used in building each solar panel has gone way down, the total energy we spend on building solar panels has gone way up? That does seem like a pure example of Jevons' Paradox to me, but I might be misunderstanding it?
I agree on the efficiency point. The cost argument is murkier because while the price of a module goes down, the gains are offset by other parts of the adoption process such as project management, labor-driven installation costs, and permitting. You see this K-shaped divergence in costs between developing countries (especially parts of Africa) where energy costs are falling and solar adoption is rising, while countries like Germany are not seeing any reduction in overall energy costs.
Hmmm yes! I think there's a missing link that I was trying to point at: energy used in building solar panel goes down -> appetite for solar panels increase -> total energy spent goes up.
To be fair to critics, as others have pointed out, the generative AI-based science and research would have to a lot of the heavy lifting to cancel out the (heavy quotation marks) "frivolous" and "inefficient" uses of AI.
On the other hand, a world in which the predicted amount of energy goes to any kind of generative AI is going to be a very different world regardless! The trillions we're talking about here far exceed the cost of just "really good ChatGPT" and "really good Midjourney" - both which might already be reaching peak utility for many casual users. Those trillions spent must somehow approximate trillions in value, because there literally isn't enough venture capital to keep moonshotting AI for some vague future promise. And that value could consist of more leisure time, greater quality of life, lowered costs of goods and services (education, healthcare...), reduced demand for labor for both corporations and individuals, or just plain economic growth. I'm sure I'm forgetting a few here.
An additional 200 Mt is obviously not a good thing, but it's not planet-destroying in the greater scheme of things, and we might just decide that it is "worth it". The projected amounts imply that we would indeed decide that it is. This doesn't help the anti-Jevons argument directly, but many of the spinoffs here absolutely do.
For instance, changes in employment patterns affect how many cars are on the road, how many planes are in the air, and how many buildings need to be heated and cooled. Economic growth affects consumption patterns (boosting e.g. heatpumps and solar). Etc.
One aspect that may be nice to add, if possible, would be the energy-(in)efficiency of the current codes that were programmed by humans vs the energy-efficiency of newer/improved codes that were (co-)programmed with/by AI. With faster and more energy efficient codes, I expect that the net carbon emissions from running all these codes in various industries and academic fields could also lean towards something like +x carbon -y carbon, where x is much smaller than y.
I read the Nature Communication paper. "The mechanism of how AI could best be used to reduce energy consumption and carbon emissions in buildings is also not clear."
So "Trust me bro, AI will work."
One factor to consider with heating and cooling is if you turn them off over night, then you need bigger systems to bring the temperature up (or down) in the morning. My parent's church discovered this when they went with the low bid. If they turn the AC off in the summer then it takes /days/ to cool off the hall, so they run it 7 days a week.
It's also worth noting that they stipulate that achieving these reductions will require substantial government intervention to force (or encourage) retrofits.
In my class on AI, I've been talking about Jevons' Paradox mainly in the section that I talk about employment. Some people think that if AI reduces the number of employees needed for a given amount of output, then the total number of employees will decrease, but I point to historic examples of automation where the total number of employees decreased, increased, or stayed the same, due to interactions with Jevons' Paradox.
My concern is the same one that I feel is facing all industries: the lure of private equity. We have seen how private equity is capturing housing and making it unaffordable, as well as capturing all other companies and profit sources and making them unaffordable. These have the effects of lowering the quality and availability of both. Local housing being priced sky high means nobody benefits, energy prices soaring means less money to actually maintain the grid. Is there an opposite to the Jevon’s Paradox that encompasses private equity? Because that’s the future I’m worried about.
It is the old Sustainable Development ploy that was used or continues to be used i.e. our new methods of creating environmental impacts are more efficient, so we can create less pollution per part but more pollution in total.
It is the total load on the environment that creates the impacts.
As long as you continue to put more carbon in the atmosphere, more ozone depletors, more particulate, more VOC's, etc, etc. - do not expect different results - which is the definition of insanity.
Are you talking about opportunity costs i.e. instead of 10 I put in 1?
Is it still one more unit than than before?
Carbon capture is still not a net positive benefit in life cycle terms as far as I know.
It can slow things down but not the overall impact over the long term.
I am ignoring the physics/cycles of atmospheric carbon - considering just the total load on the atmosphere which if it continually increases cannot improve climate change options under current global emission levels.
I'm not talking about carbon capture? Just "All else equal, if you want 100 units of energy, and one thing emits 100 units of carbon to get it, and the other emits 10, we should use the 10, even though it still emits."
I am all in favour of efficiency in energy use and reduction in impacts by newer technologies.
Still, we are all facing the conundrum of how to mitigate long term environmental impacts while still keeping our economies going to maintain the life style we are used to.
The equilibrium point is still far off and it will be a race to the finish.
I know people who bought solar panels, connected it to the grid, and had a huge energy credit in their bill they could only access it if they used more energy. So they bought a second car that was electric and bought resistant heating for corners of the house that were cold in the winter. They would have probably not done that have they not installed solar panels. That seems like a waste of resources to me. I have not seen any data saying that solar panels are replacing fossil fuels, they are just adding to our energy mix, so in that case are solar panels really good for the environment?
I feel like solar panels work similar to recycling. It makes us feel good in using more resources, even though very few items are actually recycled correctly.
I mean an electric car powered by solar also probably nets out to way fewer emissions. This mostly seems like a problem with the grid's local policies rather than the panels themselves. Ultimately we need to get away from fossil fuels and making cheap abundant green energy is how we do that, and sometimes that'll cause people to consume more.
The end goal is to use less fossil fuels, but I have not seen any data that fossil fuel use is decreasing. I remember that 2010 was the year where everyone was flabbergasted with the low price of solar and windmills. We thought surely in the next 15 years fossil fuel use would go down, but it has not happened.
I guess solar getting super cheap just seems like a necessary but not sufficient condition to move away from them, so it still seems like they net out to good in expectation even if at the moment they're not directly preventing emissions.
Thanks for the article. I'm wondering whether it makes sense to lump all AI together here though — the increase in power is mainly coming from GenAI, but the energy-saving benefits seem to be in other areas of Deep Learning. I can't really see how GenAI ends up as a net positive on the climate now (which ofc doesn't necesarilly mean it's worth it to stop using it).
Excellent point. My impression is that GenAI's having a lot of spillover effects in other areas of AI, like pouring massive amounts of money into developing better GPUs etc. and bringing more researcher talent into the space, but definitely agree that I'd be surprised if chatbots were having much of a positive climate effect specifically. Busy for the next few days but would like to look into the spillover effects more. If they're not there this weakens my case.
The (possibly overhyped) early signs of AI contributing to scientific research are all based on GenAI, and that's still what the AI industry is betting on. OpenAI's IMO gold medal-placing model was notable for not just being GenAI - all the models in this space are - but for being an LLM, ultimately a chatbot using a strange and curt form of language. Depending on your definition, even AlphaFold is generative.
Disease-curing AI and anime-fanfic-writing AI mostly differ in their application. It's more about how much energy is put into each use than how much energy is put into the research. The research benefits all.
Here's another example of generative AI benefits for science, New tool makes generative AI models more likely to create breakthrough materials, https://news.mit.edu/2025/new-tool-makes-generative-ai-models-likely-create-breakthrough-materials-0922
It also seems like Jevons' paradox is very unlikely to apply to AI, at least in terms of power efficiency. Power is only ~10% of the cost of GPU compute, so as far as price signals go it's pretty minor. So Jevons' paradox would only apply to the extent that people are curtailing usage due to carbon footprint concerns, which I expect is a pretty small share of people
[10% ballpark is for training compute, and comes from Fig 5 of Epoch's paper here: https://arxiv.org/html/2405.21015v1]
Thank you for this perspective! While it is possible for a net good, what about the individual communities facing new environmental issues because of where the data centers are being built?
Definitely need to take things on a case by case basis, I don't support any and all data center buildouts. I'm pretty skeptical about water issues (https://andymasley.substack.com/p/i-cant-find-any-instances-of-data) but seems like a few specific data centers have polluted local communities. Definitely need strong enviro regs, seems like some of the big labs have had a lot of undue influence over local governments.
Thank you - this could have been a really dense explanation but you’ve made it accessible even to a curious non academic like me. Important topic too. Bravo
Thanks for writing the article! I do want to note on a point of technicality - your example on solar panel is not true Jevon's paradox.
A better example would be due to drop in solar price => we end up consuming 1000x times more energy, which means that while the marginal emission for solar is small, the total emission would still be significant.
I think your main point would still stand though - we care *on net* if something prevent more emission than it creates.
Good point, I should add a note there. Appreciate the clarification!
Does it make sense to say though that because the energy used in building each solar panel has gone way down, the total energy we spend on building solar panels has gone way up? That does seem like a pure example of Jevons' Paradox to me, but I might be misunderstanding it?
I agree on the efficiency point. The cost argument is murkier because while the price of a module goes down, the gains are offset by other parts of the adoption process such as project management, labor-driven installation costs, and permitting. You see this K-shaped divergence in costs between developing countries (especially parts of Africa) where energy costs are falling and solar adoption is rising, while countries like Germany are not seeing any reduction in overall energy costs.
Hmmm yes! I think there's a missing link that I was trying to point at: energy used in building solar panel goes down -> appetite for solar panels increase -> total energy spent goes up.
To be fair to critics, as others have pointed out, the generative AI-based science and research would have to a lot of the heavy lifting to cancel out the (heavy quotation marks) "frivolous" and "inefficient" uses of AI.
On the other hand, a world in which the predicted amount of energy goes to any kind of generative AI is going to be a very different world regardless! The trillions we're talking about here far exceed the cost of just "really good ChatGPT" and "really good Midjourney" - both which might already be reaching peak utility for many casual users. Those trillions spent must somehow approximate trillions in value, because there literally isn't enough venture capital to keep moonshotting AI for some vague future promise. And that value could consist of more leisure time, greater quality of life, lowered costs of goods and services (education, healthcare...), reduced demand for labor for both corporations and individuals, or just plain economic growth. I'm sure I'm forgetting a few here.
An additional 200 Mt is obviously not a good thing, but it's not planet-destroying in the greater scheme of things, and we might just decide that it is "worth it". The projected amounts imply that we would indeed decide that it is. This doesn't help the anti-Jevons argument directly, but many of the spinoffs here absolutely do.
For instance, changes in employment patterns affect how many cars are on the road, how many planes are in the air, and how many buildings need to be heated and cooled. Economic growth affects consumption patterns (boosting e.g. heatpumps and solar). Etc.
Very nice article. Thanks for writing it.
One aspect that may be nice to add, if possible, would be the energy-(in)efficiency of the current codes that were programmed by humans vs the energy-efficiency of newer/improved codes that were (co-)programmed with/by AI. With faster and more energy efficient codes, I expect that the net carbon emissions from running all these codes in various industries and academic fields could also lean towards something like +x carbon -y carbon, where x is much smaller than y.
You continue to do great, important, patient work in this area. Thank you.
Hard to deal with the Doomers who find the cloud in every silver lining.
https://www.mattball.org/2025/01/few-things-are-less-important-than.html
I read the Nature Communication paper. "The mechanism of how AI could best be used to reduce energy consumption and carbon emissions in buildings is also not clear."
So "Trust me bro, AI will work."
One factor to consider with heating and cooling is if you turn them off over night, then you need bigger systems to bring the temperature up (or down) in the morning. My parent's church discovered this when they went with the low bid. If they turn the AC off in the summer then it takes /days/ to cool off the hall, so they run it 7 days a week.
It's also worth noting that they stipulate that achieving these reductions will require substantial government intervention to force (or encourage) retrofits.
This is interesting!
In my class on AI, I've been talking about Jevons' Paradox mainly in the section that I talk about employment. Some people think that if AI reduces the number of employees needed for a given amount of output, then the total number of employees will decrease, but I point to historic examples of automation where the total number of employees decreased, increased, or stayed the same, due to interactions with Jevons' Paradox.
My concern is the same one that I feel is facing all industries: the lure of private equity. We have seen how private equity is capturing housing and making it unaffordable, as well as capturing all other companies and profit sources and making them unaffordable. These have the effects of lowering the quality and availability of both. Local housing being priced sky high means nobody benefits, energy prices soaring means less money to actually maintain the grid. Is there an opposite to the Jevon’s Paradox that encompasses private equity? Because that’s the future I’m worried about.
It is the old Sustainable Development ploy that was used or continues to be used i.e. our new methods of creating environmental impacts are more efficient, so we can create less pollution per part but more pollution in total.
It is the total load on the environment that creates the impacts.
As long as you continue to put more carbon in the atmosphere, more ozone depletors, more particulate, more VOC's, etc, etc. - do not expect different results - which is the definition of insanity.
What about if each unit of carbon I put into the atmosphere is done to cancel out 10 other units of carbon?
Are you talking about opportunity costs i.e. instead of 10 I put in 1?
Is it still one more unit than than before?
Carbon capture is still not a net positive benefit in life cycle terms as far as I know.
It can slow things down but not the overall impact over the long term.
I am ignoring the physics/cycles of atmospheric carbon - considering just the total load on the atmosphere which if it continually increases cannot improve climate change options under current global emission levels.
I'm not talking about carbon capture? Just "All else equal, if you want 100 units of energy, and one thing emits 100 units of carbon to get it, and the other emits 10, we should use the 10, even though it still emits."
I am all in favour of efficiency in energy use and reduction in impacts by newer technologies.
Still, we are all facing the conundrum of how to mitigate long term environmental impacts while still keeping our economies going to maintain the life style we are used to.
The equilibrium point is still far off and it will be a race to the finish.
I know people who bought solar panels, connected it to the grid, and had a huge energy credit in their bill they could only access it if they used more energy. So they bought a second car that was electric and bought resistant heating for corners of the house that were cold in the winter. They would have probably not done that have they not installed solar panels. That seems like a waste of resources to me. I have not seen any data saying that solar panels are replacing fossil fuels, they are just adding to our energy mix, so in that case are solar panels really good for the environment?
I feel like solar panels work similar to recycling. It makes us feel good in using more resources, even though very few items are actually recycled correctly.
I mean an electric car powered by solar also probably nets out to way fewer emissions. This mostly seems like a problem with the grid's local policies rather than the panels themselves. Ultimately we need to get away from fossil fuels and making cheap abundant green energy is how we do that, and sometimes that'll cause people to consume more.
The end goal is to use less fossil fuels, but I have not seen any data that fossil fuel use is decreasing. I remember that 2010 was the year where everyone was flabbergasted with the low price of solar and windmills. We thought surely in the next 15 years fossil fuel use would go down, but it has not happened.
I guess solar getting super cheap just seems like a necessary but not sufficient condition to move away from them, so it still seems like they net out to good in expectation even if at the moment they're not directly preventing emissions.
👍