You guys DO realize there's not a single 1GW datacenter in operation, right? Yet somehow we're going to build from 30 - 100(!) of them in the next decade?
The problem here is power. There's nowhere (in the US at least) where you can just "hook up" to a GW of power. That's enough power for a city the size of *Denver* btw. The rosy projections are great if you're fundraising, but I'm sensing a distinct lack of reality here.
There's not much doubt that China has this option ... rough order of 1GW/day electric capacity growth there (most of it intermittent but still!). Since US sees this as an all-in strategic matter, money will be printed and a physical way will also be found even, if it impoverishes the rest of the economy (which we're seeing early warning signs of already, even though the meat of the buildout is several years away).
Natgas standalone units (ie typ used as peaker) exist on the order of 30MW and 100MW. These too have multi-year backlog on their manufacture, but presumably it's not as bad as the backlog on grid upgrades. US siting preferences could arguably shift after decommisioned plants blessed with unused grid connection are exhausted - toward locations near substantial natgas pipelines - so especially at Gulf coast and perhaps PA.
A strange piece. I agree with much of it directionally, but feel opposed to the evidence presented.
Fab capex overhang is true, but it's hard to guess based off the numbers in the post alone. A reader can't intrinsically *know* that Fab 18B + 21A won't be *enough* for all upcoming 3nm demand by Rubin, from the provided evidence alone.
How would one know if Jensen's making a mistake — or the right decision — in investing in e.g. OpenAI rather than TSMC?
Speaking of capex: how can a reader tell that capex invested into GEV today will cash out as new turbines before 2030? How many elements up the supply chain need to be shook awake for that doubling in cost to produce turbines in time?
There's naive "slide-deck" math throughout. Reference example:
> If AI truly lives up to its promise, then it’s in the reference class (at the very least) of white-collar wages, which are $10s of trillions of dollars a year.
White-collar AI agents can't make up anywhere near the same amount of revenue as their human counterparts do, because of deflationary effects (↓margins) and slow uptake by the human economy (↓demand).
Similar argumentation against the labor bottleneck (which workers are sufficiently efficient?), against flexible loads (isn't that what batteries are for?)
An uneducated reader will also not be informed about the ongoing NIMBY/anti-datacenter movements, which are currently sufficiently large to significantly (>10% at minimum) stymie AI datacenter construction in the US.
But if the goal of the article was *not* to be US centric, then comparing China vs the US alone is also strange, given the significant ongoing, and potential future AI buildouts internationally, which may significantly affect the perceived balance of compute power in the long run.
We're not saying Jensen is making a mistake by investing in OAI instead of TSMC. Or that there's not enough fab capacity for 3 nm demand. But that if there were such a shortage, Nvidia could subsidize expanding production for years on end with a single year of earnings.
To address the general point about napkin math. Yes of course, there's always going to be many reasons that a first pass estimate on big questions could be wrong. I don't think the correct response to that is to never try to think about complicated questions. Or torture the audience with a million hedges.
For many of your examples, you're pointing out a way our estimates could be directionally wrong on one side. But you could just as easily point out how we could be wrong on the other side. White collar AI agents will have lower margins, and in the short term, uptake might reduce demand, as you say. But over the long run, you should expect demand for white collar labor to go up. For the same reason that value of human labor goes up with population. This is the no lump of labor fallacy. We figure out new productive things to do with labor, density and agglomeration drives up labor's value, etc. I expect similar things to happen to AI.
The goal of this article is not to give a comprehensive tree of ways any Fermi estimate could be wrong. It's just to do, as you say, some slide math to help us start thinking about these questions.
The current price of knowledge work is dependent on human economic agents putting in significant effort to obtain fair wages. So far, this is not true; no AI lab cares that it could negotiate a much larger price against specific consumers based on their value-add.
Arguably, aligned AI agents should never negotiate against consumer surplus. In that world, humans produce trivial amounts of physical labor for absurd amounts of cheap knowledge work.
> The goal of this article is ... to help us start thinking about these questions.
Frankly, I'd assumed the premises were guaranteed to be in the right direction, on account of your roommate. Most of my comment was written under the assumption all of the points would be deeply familiar to Dylan, and that it would take not much time to extract good answers.
> By the way, if AI turns out to be a bubble and we’re much further from AGI than Silicon Valley thinks, what lasting value gets built? You could tell a story about how the dot-com bubble paved the way for all the value the internet has generated. What’s the equivalent for this AI buildout?
Now, I have no idea what the hell I'm talking about here, but what about biotech?
It's not doing so hot right now, but my impression is that one of the big problems in biotech is that we don't really understand a lot of biological systems - especially ones in complicated organisms like humans. So development is heavily weighted towards experimental work, but bio experiments are a huge pain in the ass, and even if you do hit on a process that works, it probably fails half the time. What we really need is the equivalent of Ansys or LTSpice for biology, so you can validate designs in silico and work on much faster cycles. But of course bio systems are much, much more complex, so this would use a lot of compute. I'm not sure how much of these problems you can solve with lots of FLOPs, though, e.g. see neuroscience where we can't even simulate the 302 neuron worm. There's also what Deepmind has been doing rather than trying to just run physics sims directly - building simulations of these systems by hand does not seem tractable.
The GPUs deprecate, yes. But if AI collapses, the lifespans of existing GPUs are immediately extended. Nvidia probably stops releasing new chips every 6 months, and there is no longer enough demand to run every chip at 70% utilization 24/7.
If demand for current AI collapses, you have a big population of GPUs which are basically worthless, which are going to flood secondary markets all at once, and my impression is that biotech stuff whether startups or research is a big part of this secondary market. Right now AI applications are very profitable per watt, so compute is expensive and doesn't get applied to bio so much. There are surely some other less profitable compute applications that are promising here, like climate modelling, but biotech seems like the most important field.
Thinking about who is showing up to these liquidation auctions for e.g. Stargate, most of them will be there to repurpose the GPUs (or hoard them, or something). Unlike past bubbles there is not much value in scrapping H100s in bulk. My impression is that a similar thing happened with dark fibre.
What problems are there? The low precision operations AI chips are good at are maybe not so useful for high precision sims. Data is definitely still a big problem, but there are areas where we have some data, like the UK biobank. Maybe bio stuff is just really, really irreducible. The chips will still be quite expensive to run even if the cost is just power. And H100s are pretty hard to set up in a university closet, but I think this is not so much of a problem, it looks more like the university department leasing cloud compute from the lowest bidder at the liquidation auction.
Could someone with actual expertise tell me why I'm wrong?
I don’t think it’s just biotech who would be able to use these firesale GPUs. While training absolutely needs the latest cutting edge GPU’s and depreciate very fast, the reality is that inference can use legacy GPU for some time. Even in the semiconductor world 100+ node processes still gets used to this day.
The current generation AI is hardly being used in the real world - it going to take time a long time for companies to integrate and change their processes around these capabilities. Could be 5-10 years before a typical F1000 company integrates today technology. When that time comes, there will be a dark data center somewhere with H100, fully depreciated, that can be turned on.
“If the 27M Software engineers worldwide are all on super charged $1000/month AI agent plans that double their productivity (for 10-20% of their salary), that would be $324B revenue already.”
There is zero chance of this happening. Even the most expensive current coding plans are 200 USD, 1/5th of that. And now there is very strong competition on the coding side: Kimi K2, GLM 4.6, Qwen Coder: all of which are open source.
The 1,000 USD per month argument only makes sense if you assume OpenAI and Anthropic etc have some proprietary technology they can earn huge margins on. The reality is that it is rapidly becoming highly commoditised.
Can somebody help me out with a very dumb question (I must be missing something)... for the dynamic you described in the first section about NVDA's prodigious profits compared to TSMC's capex costs for its fabs, why are you referring to this as a "fab capex overhang"? To me that word implies that the fab capex is very high. Shouldn't we call it a "fab capex underhang" or something since what you've discovered is that the fab capex is quite low? Again I think I'm missing something obvious and I'd appreciate if anybody can help me out here.
Regarding the Duke paper on load flexibility (Tyler Norris et al.), it’s worth clarifying that “transmission capacity” remains a bottleneck even if the study shows that we have enough “capacity” (a.k.a. power) or “resources” to make the findings true.
As a first-order study, the original paper uses a lumped model, treating all power sources as one lump and all load as one lump. This is an idealization that allows imaginary transmission from arbitrary locations to arbitrary locations in the country. Obviously, that’s not actually possible.
But that’s its own matter. Separately, I’ve found this a useful portrait of buildout and look forward to more from Romeo.
Sam Altman wants a gigawatt a week. That’s not a metaphor. That’s an energy empire plan.
What we’re watching isn’t a software boom. It’s an infrastructure race. AI is pushing the internet from “cloud” to “power grid.” GPUs are the new oil rigs.
The irony? The same companies preaching efficiency are now forced to build the most resource-intensive industry since steel. But that’s how revolutions start — by overbuilding first, optimizing later.
We’re past the age of “apps.” We’ve entered the age of “assets.”
Whoever controls the watts, wins the intelligence.
Watts this watts that… what’s the real objective? Reaching the asset age across all productivity nodes will require heavy power grid and cloud usage. I just hope we have toast for all the toasters…
Great piece, this is a fascinating breakdown of the buildout's moving parts. You charted an "AI Winter" scenario crashing around 2029, and I think you're right to be cautious. I'd like to add a few factors that might even pull that timeline forward to 2027/28.
My main concern is the sheer scale of the investment. You're talking about $400B/year in US AI CapEx. This creates two huge, linked risks:
The AGI "Due Date": That $400B is being spent now with the assumption that AGI or, at the very least, transformative economic utility is coming soon. If it's delayed, or the path to it proves harder than thought, the entire financial justification collapses.
Slow Adoption vs. Impatient Investors: Even if the tech works, customers (especially large enterprises) won't adopt it overnight. As commenter Anthony Rivadeneira pointed out, it could take 5-10 years for a typical F1000 to re-tool its processes. Investors, however, are impatient and will want to see revenue catch up to that massive CapEx much faster. If they don't, the funding will dry up quickly.
Secondly, you covered bottlenecks like labor and the grid, but I think two others are being underestimated:
Social Unrest: Commenter Sherman mentioned local NIMBYism. I think this could balloon into a much larger societal issue. If white-collar displacement becomes rapid and visible, we could see a "21st-century Luddite" movement with serious political pressure to halt or slow the buildout.
Water for Cooling: This wasn't mentioned, but cooling these massive GW-scale facilities requires an enormous amount of water. This puts them in direct competition with cities and agriculture, especially in the hot, dry regions often chosen for these projects. This could become a critical physical and political bottleneck.
These factors together—the massive financial overhang needing immediate justification, slow real-world adoption, and new social/physical bottlenecks—make me think that "AI winter" scenario is highly plausible, and could hit even sooner than 2029.
Great comment, I've noticed no one has has mentioned the "Ai winter" description might aptly be named "Ai early autumn". It seems to me things could have way more downside than outlined
Very interesting read. Wanted to share some more information on Europe's side.
Our datacenter-power demand is forecast to double to about 36 GW by 2030 in the major markets.
We’ve got major firms like Siemens Energy and ABB ready to build large-scale power systems and grid gear, so the “upstream supply chain” isn’t purely a U.S./China story.
On semiconductors, the fab in Dresden (with TSMC, Infineon etc) is a solid sign we’re not entirely behind the boom either. And ofc ASML is a key player there, they are scaling their production (though they have challenges scaling their suppliers production)
But obviously there are SO MANY thing that needs work:
A big bottleneck is permitting and siting. You can build the best fab or turbine factory, but if it takes 3–5 years to tie it into the grid, you lose the hot-moment.
Labor for all this isn’t cheap or infinite. Europe needs a pipeline of heavy-equipment builders, electricians, EPC crews. we might have a bigger labor problem than the US
Standardisation. If every site needs custom racks, cooling, power-design, we’ll still be doing “car in your driveway” style buildouts (one off) instead of “module out of factory” style. The blog post nails that and that is a problem we have had in europe for so long (some trains in europe have 4 different transformer energy systems on them because the grid is not standardised across countries)
My bet: Europe can play the “get power and infrastructure fast” card, while the U.S. and China hammer on leading-edge chips, we focus on “make 100 MW/10-year sites everywhere, off-grid or grid-flexible”.
Very insightful article. No doubt the macro trend is growth.
A few assumptions we should challenge.
± How can you optimize compute better? Training and large scale inference are batch workloads. There is capacity in existing data center and grid capacity, how can we optimize? illuma.cloud
± If you assume OpenAI revenue growth comes from advertising, would that be at the expense of Google and Meta? Or is the assumption media/ad budget from companies shift from offline to GenAI?
± Will fine tuning and small language models give enough performance for many task with a fraction of the compute requirements?
Meta’s “Prometheus” glasses are referred this way as an idle name call. Orion (the glasses) seem to be lacking from this specific article. Interesting point you raise.
There is one problem: The human consumers don't have the money to give the huge investments enough returns. Just asked yourself and your friends: How much they have to spend on OpenAI so that OpenAI will be profitable?
How about fusion, ITER and Tokamak- will it solve the energy problem for ever - though the wait is long but perhaps can bring an end to energy crisis for AI..
You guys DO realize there's not a single 1GW datacenter in operation, right? Yet somehow we're going to build from 30 - 100(!) of them in the next decade?
The problem here is power. There's nowhere (in the US at least) where you can just "hook up" to a GW of power. That's enough power for a city the size of *Denver* btw. The rosy projections are great if you're fundraising, but I'm sensing a distinct lack of reality here.
There's not much doubt that China has this option ... rough order of 1GW/day electric capacity growth there (most of it intermittent but still!). Since US sees this as an all-in strategic matter, money will be printed and a physical way will also be found even, if it impoverishes the rest of the economy (which we're seeing early warning signs of already, even though the meat of the buildout is several years away).
Natgas standalone units (ie typ used as peaker) exist on the order of 30MW and 100MW. These too have multi-year backlog on their manufacture, but presumably it's not as bad as the backlog on grid upgrades. US siting preferences could arguably shift after decommisioned plants blessed with unused grid connection are exhausted - toward locations near substantial natgas pipelines - so especially at Gulf coast and perhaps PA.
A strange piece. I agree with much of it directionally, but feel opposed to the evidence presented.
Fab capex overhang is true, but it's hard to guess based off the numbers in the post alone. A reader can't intrinsically *know* that Fab 18B + 21A won't be *enough* for all upcoming 3nm demand by Rubin, from the provided evidence alone.
How would one know if Jensen's making a mistake — or the right decision — in investing in e.g. OpenAI rather than TSMC?
Speaking of capex: how can a reader tell that capex invested into GEV today will cash out as new turbines before 2030? How many elements up the supply chain need to be shook awake for that doubling in cost to produce turbines in time?
There's naive "slide-deck" math throughout. Reference example:
> If AI truly lives up to its promise, then it’s in the reference class (at the very least) of white-collar wages, which are $10s of trillions of dollars a year.
White-collar AI agents can't make up anywhere near the same amount of revenue as their human counterparts do, because of deflationary effects (↓margins) and slow uptake by the human economy (↓demand).
Similar argumentation against the labor bottleneck (which workers are sufficiently efficient?), against flexible loads (isn't that what batteries are for?)
An uneducated reader will also not be informed about the ongoing NIMBY/anti-datacenter movements, which are currently sufficiently large to significantly (>10% at minimum) stymie AI datacenter construction in the US.
But if the goal of the article was *not* to be US centric, then comparing China vs the US alone is also strange, given the significant ongoing, and potential future AI buildouts internationally, which may significantly affect the perceived balance of compute power in the long run.
We're not saying Jensen is making a mistake by investing in OAI instead of TSMC. Or that there's not enough fab capacity for 3 nm demand. But that if there were such a shortage, Nvidia could subsidize expanding production for years on end with a single year of earnings.
To address the general point about napkin math. Yes of course, there's always going to be many reasons that a first pass estimate on big questions could be wrong. I don't think the correct response to that is to never try to think about complicated questions. Or torture the audience with a million hedges.
For many of your examples, you're pointing out a way our estimates could be directionally wrong on one side. But you could just as easily point out how we could be wrong on the other side. White collar AI agents will have lower margins, and in the short term, uptake might reduce demand, as you say. But over the long run, you should expect demand for white collar labor to go up. For the same reason that value of human labor goes up with population. This is the no lump of labor fallacy. We figure out new productive things to do with labor, density and agglomeration drives up labor's value, etc. I expect similar things to happen to AI.
The goal of this article is not to give a comprehensive tree of ways any Fermi estimate could be wrong. It's just to do, as you say, some slide math to help us start thinking about these questions.
Re: lump-of-labor-fallacy
The current price of knowledge work is dependent on human economic agents putting in significant effort to obtain fair wages. So far, this is not true; no AI lab cares that it could negotiate a much larger price against specific consumers based on their value-add.
Arguably, aligned AI agents should never negotiate against consumer surplus. In that world, humans produce trivial amounts of physical labor for absurd amounts of cheap knowledge work.
In the longer run, when not just white-collar work, but all work is solved, I buy Cunningham’s (https://tecunningham.github.io/posts/2025-09-19-transformative-AI-notes.html#if-ai-can-do-everything-then-wages-will-fall.) frame of the issue; in a resource-bottlenecked regime, all wages should fall.
> The goal of this article is ... to help us start thinking about these questions.
Frankly, I'd assumed the premises were guaranteed to be in the right direction, on account of your roommate. Most of my comment was written under the assumption all of the points would be deeply familiar to Dylan, and that it would take not much time to extract good answers.
Sorry if that guess was too presumptive.
> By the way, if AI turns out to be a bubble and we’re much further from AGI than Silicon Valley thinks, what lasting value gets built? You could tell a story about how the dot-com bubble paved the way for all the value the internet has generated. What’s the equivalent for this AI buildout?
Now, I have no idea what the hell I'm talking about here, but what about biotech?
It's not doing so hot right now, but my impression is that one of the big problems in biotech is that we don't really understand a lot of biological systems - especially ones in complicated organisms like humans. So development is heavily weighted towards experimental work, but bio experiments are a huge pain in the ass, and even if you do hit on a process that works, it probably fails half the time. What we really need is the equivalent of Ansys or LTSpice for biology, so you can validate designs in silico and work on much faster cycles. But of course bio systems are much, much more complex, so this would use a lot of compute. I'm not sure how much of these problems you can solve with lots of FLOPs, though, e.g. see neuroscience where we can't even simulate the 302 neuron worm. There's also what Deepmind has been doing rather than trying to just run physics sims directly - building simulations of these systems by hand does not seem tractable.
The GPUs deprecate, yes. But if AI collapses, the lifespans of existing GPUs are immediately extended. Nvidia probably stops releasing new chips every 6 months, and there is no longer enough demand to run every chip at 70% utilization 24/7.
If demand for current AI collapses, you have a big population of GPUs which are basically worthless, which are going to flood secondary markets all at once, and my impression is that biotech stuff whether startups or research is a big part of this secondary market. Right now AI applications are very profitable per watt, so compute is expensive and doesn't get applied to bio so much. There are surely some other less profitable compute applications that are promising here, like climate modelling, but biotech seems like the most important field.
Thinking about who is showing up to these liquidation auctions for e.g. Stargate, most of them will be there to repurpose the GPUs (or hoard them, or something). Unlike past bubbles there is not much value in scrapping H100s in bulk. My impression is that a similar thing happened with dark fibre.
What problems are there? The low precision operations AI chips are good at are maybe not so useful for high precision sims. Data is definitely still a big problem, but there are areas where we have some data, like the UK biobank. Maybe bio stuff is just really, really irreducible. The chips will still be quite expensive to run even if the cost is just power. And H100s are pretty hard to set up in a university closet, but I think this is not so much of a problem, it looks more like the university department leasing cloud compute from the lowest bidder at the liquidation auction.
Could someone with actual expertise tell me why I'm wrong?
I don’t think it’s just biotech who would be able to use these firesale GPUs. While training absolutely needs the latest cutting edge GPU’s and depreciate very fast, the reality is that inference can use legacy GPU for some time. Even in the semiconductor world 100+ node processes still gets used to this day.
The current generation AI is hardly being used in the real world - it going to take time a long time for companies to integrate and change their processes around these capabilities. Could be 5-10 years before a typical F1000 company integrates today technology. When that time comes, there will be a dark data center somewhere with H100, fully depreciated, that can be turned on.
“If the 27M Software engineers worldwide are all on super charged $1000/month AI agent plans that double their productivity (for 10-20% of their salary), that would be $324B revenue already.”
There is zero chance of this happening. Even the most expensive current coding plans are 200 USD, 1/5th of that. And now there is very strong competition on the coding side: Kimi K2, GLM 4.6, Qwen Coder: all of which are open source.
The 1,000 USD per month argument only makes sense if you assume OpenAI and Anthropic etc have some proprietary technology they can earn huge margins on. The reality is that it is rapidly becoming highly commoditised.
I pay ~1000/mo on cursor rn!
> Even the most expensive current coding plans are 200 USD, 1/5th of that.
(Just pointing out that Devin has a $500/month plan)
Can somebody help me out with a very dumb question (I must be missing something)... for the dynamic you described in the first section about NVDA's prodigious profits compared to TSMC's capex costs for its fabs, why are you referring to this as a "fab capex overhang"? To me that word implies that the fab capex is very high. Shouldn't we call it a "fab capex underhang" or something since what you've discovered is that the fab capex is quite low? Again I think I'm missing something obvious and I'd appreciate if anybody can help me out here.
I was also confused. I asked ChatGPT for help--
Me: Would it still be accurate (if wordy) to call it an "overhang of fab-capex-demand" ?
If you want maximum clarity, consider one of these instead:
“Demand‑induced fab capex overhang.”
Keeps their noun (“capex overhang”) but makes the demand causality explicit.
“Overhang of demand for fab capacity.”
Focuses on the thing buyers ultimately want (capacity), not the accounting category (capex).
“Upstream capacity‑investment overhang.”
Puts the emphasis where the essay does: a large, still‑to‑arrive build‑out upstream because WTP ≫ capacity cost today.
[WTP = Willingness To Pay]
Funny, I also asked ChatGPT. I guess we both experienced "ChatGPT query overhang" there.
Yeah, overhang means different things in different contexts.
There's an increasingly large pile of money that could be spent on fabs, and this capacity is "overhanging" the actual investments.
This use follows from e.g., https://www.lesswrong.com/posts/N6vZEnCn6A95Xn39p/are-we-in-an-ai-overhang
I see. Thank you!
Thanks for putting this together.
Regarding the Duke paper on load flexibility (Tyler Norris et al.), it’s worth clarifying that “transmission capacity” remains a bottleneck even if the study shows that we have enough “capacity” (a.k.a. power) or “resources” to make the findings true.
As a first-order study, the original paper uses a lumped model, treating all power sources as one lump and all load as one lump. This is an idealization that allows imaginary transmission from arbitrary locations to arbitrary locations in the country. Obviously, that’s not actually possible.
But that’s its own matter. Separately, I’ve found this a useful portrait of buildout and look forward to more from Romeo.
Sam Altman wants a gigawatt a week. That’s not a metaphor. That’s an energy empire plan.
What we’re watching isn’t a software boom. It’s an infrastructure race. AI is pushing the internet from “cloud” to “power grid.” GPUs are the new oil rigs.
The irony? The same companies preaching efficiency are now forced to build the most resource-intensive industry since steel. But that’s how revolutions start — by overbuilding first, optimizing later.
We’re past the age of “apps.” We’ve entered the age of “assets.”
Whoever controls the watts, wins the intelligence.
Watts this watts that… what’s the real objective? Reaching the asset age across all productivity nodes will require heavy power grid and cloud usage. I just hope we have toast for all the toasters…
Great piece, this is a fascinating breakdown of the buildout's moving parts. You charted an "AI Winter" scenario crashing around 2029, and I think you're right to be cautious. I'd like to add a few factors that might even pull that timeline forward to 2027/28.
My main concern is the sheer scale of the investment. You're talking about $400B/year in US AI CapEx. This creates two huge, linked risks:
The AGI "Due Date": That $400B is being spent now with the assumption that AGI or, at the very least, transformative economic utility is coming soon. If it's delayed, or the path to it proves harder than thought, the entire financial justification collapses.
Slow Adoption vs. Impatient Investors: Even if the tech works, customers (especially large enterprises) won't adopt it overnight. As commenter Anthony Rivadeneira pointed out, it could take 5-10 years for a typical F1000 to re-tool its processes. Investors, however, are impatient and will want to see revenue catch up to that massive CapEx much faster. If they don't, the funding will dry up quickly.
Secondly, you covered bottlenecks like labor and the grid, but I think two others are being underestimated:
Social Unrest: Commenter Sherman mentioned local NIMBYism. I think this could balloon into a much larger societal issue. If white-collar displacement becomes rapid and visible, we could see a "21st-century Luddite" movement with serious political pressure to halt or slow the buildout.
Water for Cooling: This wasn't mentioned, but cooling these massive GW-scale facilities requires an enormous amount of water. This puts them in direct competition with cities and agriculture, especially in the hot, dry regions often chosen for these projects. This could become a critical physical and political bottleneck.
These factors together—the massive financial overhang needing immediate justification, slow real-world adoption, and new social/physical bottlenecks—make me think that "AI winter" scenario is highly plausible, and could hit even sooner than 2029.
Great comment, I've noticed no one has has mentioned the "Ai winter" description might aptly be named "Ai early autumn". It seems to me things could have way more downside than outlined
Very interesting read. Wanted to share some more information on Europe's side.
Our datacenter-power demand is forecast to double to about 36 GW by 2030 in the major markets.
We’ve got major firms like Siemens Energy and ABB ready to build large-scale power systems and grid gear, so the “upstream supply chain” isn’t purely a U.S./China story.
On semiconductors, the fab in Dresden (with TSMC, Infineon etc) is a solid sign we’re not entirely behind the boom either. And ofc ASML is a key player there, they are scaling their production (though they have challenges scaling their suppliers production)
But obviously there are SO MANY thing that needs work:
A big bottleneck is permitting and siting. You can build the best fab or turbine factory, but if it takes 3–5 years to tie it into the grid, you lose the hot-moment.
Labor for all this isn’t cheap or infinite. Europe needs a pipeline of heavy-equipment builders, electricians, EPC crews. we might have a bigger labor problem than the US
Standardisation. If every site needs custom racks, cooling, power-design, we’ll still be doing “car in your driveway” style buildouts (one off) instead of “module out of factory” style. The blog post nails that and that is a problem we have had in europe for so long (some trains in europe have 4 different transformer energy systems on them because the grid is not standardised across countries)
My bet: Europe can play the “get power and infrastructure fast” card, while the U.S. and China hammer on leading-edge chips, we focus on “make 100 MW/10-year sites everywhere, off-grid or grid-flexible”.
Worth considering that GPUs improvements are less significant each generation, and they physically last longer than 3 years
Very insightful article. No doubt the macro trend is growth.
A few assumptions we should challenge.
± How can you optimize compute better? Training and large scale inference are batch workloads. There is capacity in existing data center and grid capacity, how can we optimize? illuma.cloud
± If you assume OpenAI revenue growth comes from advertising, would that be at the expense of Google and Meta? Or is the assumption media/ad budget from companies shift from offline to GenAI?
± Will fine tuning and small language models give enough performance for many task with a fraction of the compute requirements?
Great write up thank you! Small note is meta’s “Prometheus” instead of Orion (glasses)?
I think the question to ask here, is rather: Orion as a foundational principle, or as a base AI functionality ?
Another egregious, almost philanthropical question pondered here by Mike. Lots of fruitful discussion to be had!
Meta’s “Prometheus” glasses are referred this way as an idle name call. Orion (the glasses) seem to be lacking from this specific article. Interesting point you raise.
When are you going to have a Pod with Dylan Patel? So much looking forward to that!
There is one problem: The human consumers don't have the money to give the huge investments enough returns. Just asked yourself and your friends: How much they have to spend on OpenAI so that OpenAI will be profitable?
https://open.substack.com/pub/feisal/p/screaming-about-bubbles?r=9pv9y&utm_medium=ios
How about fusion, ITER and Tokamak- will it solve the energy problem for ever - though the wait is long but perhaps can bring an end to energy crisis for AI..