This is a great start, especially considering that most of us are frozen by uncertainty looking at scenarios that could go in such wildly different directions.
Having said that, does this analysis take the exponentials involved seriously enough? For the sake of argument, if inference costs continue to go down 10X per year, does that imply that our assumptions about GPUs / energy will be 4-5 orders of magnitude off by 2030? Will o1-pro level inference cost $2 per year instead of $2000?
I usually shy away from such extreme exponential forecasts, but this seems like a situation where they might apply, especially because all the AI companies are going to be using the best models internally for AI development and training data creation.
I had to use some starting point and assumptions. Otherwise the problem with assuming exponentials are that they swamp everything else ( eg GPUs become 1000000x cheaper), which is also why I chose the power constraints so there's a fixed starting point.
There is going to be an interesting trade off of Moore’s Law, or its related cousin on chip efficiency vs the cost of electricity in a rising demand environment. If you want a moonshot, I’d throw all the ‘AGI’ at solving the power generation issue - although the simple short term solution would be ‘drill baby drill’.
It will be interesting to see what happens once power demands start causing increased energy costs for other industries and consumers.
The comment on Phd monetisation is interesting, because of course the whole huge industry of computing is based on PhD level mathematics (from Turing to bitcoin) - most of which had no commercial utility for years or decades. Or in the case of group theory, a couple of centuries.
Your math seems incorrect and has a very large error. You quote $192-$576 for 64 GPU-hours and then state the query takes 5.3 GPU hours and will cost $1000. Would it not cost 5.3/64x ($192-$576) which is $16-$48?
I find the ai power usage projection a bit unconvincing. Current ai throws everything it's got at problems. That's just not smart. There's no need for say the history of baseball to be used on a plumbing problem.
The human brain operates at like couple of percent activation, most of your brain is quiescent even when you're mentally active. Systems of neurons are switched on as needed then switched off. This is a result of competition driving efficiency.
I expect to see much lighter models used for domain specific problem. One thing that ai can not do very well, or maybe at all, currently is detect when it needs additional resources. That will come.
There was a time in early computing when people boasted about the size of computers and how much power it used. The space program started developing silicon chips to get computers that weighed kilos, not tons, and that conversation changed.
It early days in ai. Power is more important than efficiency. That will change.
This is another way of saying we'll massively increase the power efficiency of how we run AI. It's possible, and I account for some of it, but you have to choose some number and work the math out.
Isn't it too soon to make such a prediction while we're still in so much flux? Appears to me (soon after trying out DeepThink which was impressively quick) that costs & capabilities will continue to evolve. My base assumption is no longer that OpenAI or Google et al will continue to dominate - they could but this is no longer certain. OTOH thinking about what it might mean is useful of course - Ethan had a post with a slightly different emphasis that I'm sure you've read.
A quick back of the envelope calculation is about 1e15s worked in the US by 2030*. Round up your estimate of 2030 AI power use to 200TWh, divide by 500W = 1.44e15s, which is 44% more hours of AI work than human work.
---
"Even assuming it takes 1000x effort in some domains and at 1% success rate, that's still 400 breakthroughs."
This isn't a simple calculation to make, but one can get a much better idea of how many such extremely hard problems one can solve by applying Rasch measures of intelligence (the basis for item response theory and computer adaptive testing). The crucial variable is the difficulty of the problems one defines as breakthroughs. Difficulty is measured on the same scale as intelligence using Rasch measures. This difficulty can be better estimated for problems that have already been solved, and at least a lower bound can be estimated for unsolved problems. I suggested this in a application for an AI psychometrician position posted on the old SL4 list back in 2005 or -6, but the AI eval field seems to still have not gotten up to speed on 1960s-80s test theory.
This is a good start. People can debate over the nitty gritty but the absolute limit of how many advanced GPUs can be fabbed and power output available puts a hard limit on available compute in 2030. It's unlikely that you're off by much there, of course assuming no big geopolitical conflict or economic correction. What remains is everyone's personal timelines/bets for how efficient and effective AGI will be.
Regardless, it's also patently obvious then that by 2030 or even 2035 short of an AGI led manufacturing/industrial revolution the world won't have equitably distributed compute whatsoever. There will be US and China with access to ludicrous amounts of power and compute, as well as the Middle Eastern Petrostates and a few European countries (Norway, perhaps Iceland) who might also be able to host sufficient amounts of compute.
Countries who don't have these datacenters to augment productivity and output will be left behind, much like what we saw (or didn't see as we weren't alive then) with the Industrial Revolution.
This is really great and tracks with an Artifact I built that looked at estimated about of “digital brains” that would be deployed in workforce over next few years. It jibes with the trajectory that you laid out.
I feel like I’ve seen broader based analyses on whole substitutions. However, it would be cool to see is a deeper dive into a sector by sector rollout of how AI might be deployed. I.e education still feels like it won’t get moved in near term (teacher salaries are Lindy).
I find this Fermi estimate quite useful - some numbers might be off for power but the analysis holds. I also think we have no way yet of projecting or measuring an AI-agent or 'uneven-AGI' progress over time in a given set of tasks or a role. a 90% capable agent is likely at least asymptotically getting to higher productivity + performance over a few months since the rate and variance of tasks is likely to evolve slower than the agent capability (my assumption) and efficiency improvements especially with more agile smaller models. The big question in my mind is the rate of change required on the receiving side of this set of innovations. If I can make available few or few tens of near 100% capable agents today to any large enterprise in a non-coding workflow, I don't think the fabric exists to make those agents productive. As an engineer, I am comfortable with power-constrained, or silicon-constrained scenarios for AI. I worry though that we have not yet begun to develop frameworks for autonomous-agent interaction and 'work'.
Rohit, your analysis of AGI’s resource constraints is critical, but it’s worth revisiting an even deeper challenge: memory and experiential learning. Two years ago, I raised this concern in response to one of your posts, and it remains as relevant today. Current AI lacks the dynamic memory needed to build cumulative understanding over time, a limitation that hinders its ability to adapt meaningfully across evolving contexts or retain a stable identity. This bottleneck doesn’t just constrain technical progress; it reshapes how society must prepare for an AI-driven future.
While AGI will dominate structured and repeatable tasks, human skills in improvisation, repair, and relational expertise are poised to become even more valuable. Leadership, creativity, and navigating social complexity—abilities rooted in emotional intelligence and adaptability—could define success in an AI-rich world. These trends are already visible, as the most impactful AI implementations enhance human strengths rather than attempt to replace them entirely.
The challenge is not merely managing this transition but designing systems that thrive on stress and setbacks. History teaches us that resilience comes from learning, adaptation, and growth in the face of disruption. To succeed, we must build frameworks—spanning technology, organizations, and society—that not only tolerate change but actively improve through it. The future lies in synergy: a partnership where AI amplifies human wisdom and creativity, fostering systems that are both adaptive and enduring.
" I could imagine a job where you do certain tasks enough that they're teachable to an AI, collect data with sufficient fidelity, adjust their chains of thought, adjust the environment within which they learn, and continually test and adjust the edge cases where they fail. A constant work → eval → adjust loop."
This will be the most probably practical way in the future of AGI.
Really appreciated the detailed breakdown of power demand and FLOP bottlenecks. Also, o-1 did a great job with the matrix of scenarios.
I agree with @Jim Birch that specialization will be increasingly important in the future. So much of finding the right answer is having and appropriately considering the right context, and the more context you need, the more your inference cost goes up, especially with a big general model.
The cost of testing/ensuring AI is giving contextually appropriate answers probably makes the efficiency gains low (or even negative) early on but much higher once successfully implemented. Though I do wonder how much human supervision will be required for adaptation to new tasks later on.
Loved this essay. Unrelated question , today my wife asked me , if AGI takes away most of jobs who would buy all the things these tech companies create ? When no one has money to buy anything who would advertise on meta and googles of the world ? Any thoughts ?
This is a great start, especially considering that most of us are frozen by uncertainty looking at scenarios that could go in such wildly different directions.
Having said that, does this analysis take the exponentials involved seriously enough? For the sake of argument, if inference costs continue to go down 10X per year, does that imply that our assumptions about GPUs / energy will be 4-5 orders of magnitude off by 2030? Will o1-pro level inference cost $2 per year instead of $2000?
I usually shy away from such extreme exponential forecasts, but this seems like a situation where they might apply, especially because all the AI companies are going to be using the best models internally for AI development and training data creation.
Things could get much weirder.
I had to use some starting point and assumptions. Otherwise the problem with assuming exponentials are that they swamp everything else ( eg GPUs become 1000000x cheaper), which is also why I chose the power constraints so there's a fixed starting point.
I think the math is wrong. It costs $200-$300 per inference run
There is going to be an interesting trade off of Moore’s Law, or its related cousin on chip efficiency vs the cost of electricity in a rising demand environment. If you want a moonshot, I’d throw all the ‘AGI’ at solving the power generation issue - although the simple short term solution would be ‘drill baby drill’.
It will be interesting to see what happens once power demands start causing increased energy costs for other industries and consumers.
The comment on Phd monetisation is interesting, because of course the whole huge industry of computing is based on PhD level mathematics (from Turing to bitcoin) - most of which had no commercial utility for years or decades. Or in the case of group theory, a couple of centuries.
Your math seems incorrect and has a very large error. You quote $192-$576 for 64 GPU-hours and then state the query takes 5.3 GPU hours and will cost $1000. Would it not cost 5.3/64x ($192-$576) which is $16-$48?
I find the ai power usage projection a bit unconvincing. Current ai throws everything it's got at problems. That's just not smart. There's no need for say the history of baseball to be used on a plumbing problem.
The human brain operates at like couple of percent activation, most of your brain is quiescent even when you're mentally active. Systems of neurons are switched on as needed then switched off. This is a result of competition driving efficiency.
I expect to see much lighter models used for domain specific problem. One thing that ai can not do very well, or maybe at all, currently is detect when it needs additional resources. That will come.
There was a time in early computing when people boasted about the size of computers and how much power it used. The space program started developing silicon chips to get computers that weighed kilos, not tons, and that conversation changed.
It early days in ai. Power is more important than efficiency. That will change.
This is another way of saying we'll massively increase the power efficiency of how we run AI. It's possible, and I account for some of it, but you have to choose some number and work the math out.
Good start to get one's head around this issue. Like the flow of arguments!
Isn't it too soon to make such a prediction while we're still in so much flux? Appears to me (soon after trying out DeepThink which was impressively quick) that costs & capabilities will continue to evolve. My base assumption is no longer that OpenAI or Google et al will continue to dominate - they could but this is no longer certain. OTOH thinking about what it might mean is useful of course - Ethan had a post with a slightly different emphasis that I'm sure you've read.
This was exactly my own thinking reading this post. It's based on old, pre DeepSeek methods.
I'd love to see how this can be updated
A quick back of the envelope calculation is about 1e15s worked in the US by 2030*. Round up your estimate of 2030 AI power use to 200TWh, divide by 500W = 1.44e15s, which is 44% more hours of AI work than human work.
---
"Even assuming it takes 1000x effort in some domains and at 1% success rate, that's still 400 breakthroughs."
This isn't a simple calculation to make, but one can get a much better idea of how many such extremely hard problems one can solve by applying Rasch measures of intelligence (the basis for item response theory and computer adaptive testing). The crucial variable is the difficulty of the problems one defines as breakthroughs. Difficulty is measured on the same scale as intelligence using Rasch measures. This difficulty can be better estimated for problems that have already been solved, and at least a lower bound can be estimated for unsolved problems. I suggested this in a application for an AI psychometrician position posted on the old SL4 list back in 2005 or -6, but the AI eval field seems to still have not gotten up to speed on 1960s-80s test theory.
More on the relationship of difficulty to ability determines probability of correct answers: https://enonh.substack.com/i/149185059/on-w-scores
*(=280e9 hrs 2030; FRED B4701C0A222NBEA last data for 2022 was about 260e9 hrs, 243e9 in 2015)
Excellent as always!
Thank you!
This is a good start. People can debate over the nitty gritty but the absolute limit of how many advanced GPUs can be fabbed and power output available puts a hard limit on available compute in 2030. It's unlikely that you're off by much there, of course assuming no big geopolitical conflict or economic correction. What remains is everyone's personal timelines/bets for how efficient and effective AGI will be.
Regardless, it's also patently obvious then that by 2030 or even 2035 short of an AGI led manufacturing/industrial revolution the world won't have equitably distributed compute whatsoever. There will be US and China with access to ludicrous amounts of power and compute, as well as the Middle Eastern Petrostates and a few European countries (Norway, perhaps Iceland) who might also be able to host sufficient amounts of compute.
Countries who don't have these datacenters to augment productivity and output will be left behind, much like what we saw (or didn't see as we weren't alive then) with the Industrial Revolution.
This is really great and tracks with an Artifact I built that looked at estimated about of “digital brains” that would be deployed in workforce over next few years. It jibes with the trajectory that you laid out.
I feel like I’ve seen broader based analyses on whole substitutions. However, it would be cool to see is a deeper dive into a sector by sector rollout of how AI might be deployed. I.e education still feels like it won’t get moved in near term (teacher salaries are Lindy).
https://chatgpt.com/share/6799c383-0e7c-800f-8f0e-9226c024b32d
Thank you! I will check this out.
great read.
I find this Fermi estimate quite useful - some numbers might be off for power but the analysis holds. I also think we have no way yet of projecting or measuring an AI-agent or 'uneven-AGI' progress over time in a given set of tasks or a role. a 90% capable agent is likely at least asymptotically getting to higher productivity + performance over a few months since the rate and variance of tasks is likely to evolve slower than the agent capability (my assumption) and efficiency improvements especially with more agile smaller models. The big question in my mind is the rate of change required on the receiving side of this set of innovations. If I can make available few or few tens of near 100% capable agents today to any large enterprise in a non-coding workflow, I don't think the fabric exists to make those agents productive. As an engineer, I am comfortable with power-constrained, or silicon-constrained scenarios for AI. I worry though that we have not yet begun to develop frameworks for autonomous-agent interaction and 'work'.
Rohit, your analysis of AGI’s resource constraints is critical, but it’s worth revisiting an even deeper challenge: memory and experiential learning. Two years ago, I raised this concern in response to one of your posts, and it remains as relevant today. Current AI lacks the dynamic memory needed to build cumulative understanding over time, a limitation that hinders its ability to adapt meaningfully across evolving contexts or retain a stable identity. This bottleneck doesn’t just constrain technical progress; it reshapes how society must prepare for an AI-driven future.
While AGI will dominate structured and repeatable tasks, human skills in improvisation, repair, and relational expertise are poised to become even more valuable. Leadership, creativity, and navigating social complexity—abilities rooted in emotional intelligence and adaptability—could define success in an AI-rich world. These trends are already visible, as the most impactful AI implementations enhance human strengths rather than attempt to replace them entirely.
The challenge is not merely managing this transition but designing systems that thrive on stress and setbacks. History teaches us that resilience comes from learning, adaptation, and growth in the face of disruption. To succeed, we must build frameworks—spanning technology, organizations, and society—that not only tolerate change but actively improve through it. The future lies in synergy: a partnership where AI amplifies human wisdom and creativity, fostering systems that are both adaptive and enduring.
Mike
" I could imagine a job where you do certain tasks enough that they're teachable to an AI, collect data with sufficient fidelity, adjust their chains of thought, adjust the environment within which they learn, and continually test and adjust the edge cases where they fail. A constant work → eval → adjust loop."
This will be the most probably practical way in the future of AGI.
Whether what kind of definition it is.
Yeah it's going to be interesting!
Exactly. Your idea is the most practical one.
Others are here and there.
Really appreciated the detailed breakdown of power demand and FLOP bottlenecks. Also, o-1 did a great job with the matrix of scenarios.
I agree with @Jim Birch that specialization will be increasingly important in the future. So much of finding the right answer is having and appropriately considering the right context, and the more context you need, the more your inference cost goes up, especially with a big general model.
The cost of testing/ensuring AI is giving contextually appropriate answers probably makes the efficiency gains low (or even negative) early on but much higher once successfully implemented. Though I do wonder how much human supervision will be required for adaptation to new tasks later on.
Loved this essay. Unrelated question , today my wife asked me , if AGI takes away most of jobs who would buy all the things these tech companies create ? When no one has money to buy anything who would advertise on meta and googles of the world ? Any thoughts ?