We are making great strides on AI adoption, but new investment is largely AGI focused. This is surprising, since luminaries have been walking back their timelines. And future science may be at risk.
In your interview with Goldman Sachs you made an observation about the flow of dollars in the AI universe: "One observation that a lot of people have made is, if a dollar comes in at the top, Nvidia keeps $1.20 today. So Nvidia is capturing a lot of the value in the supply chain today." I'm not following... how is NVDA capturing $1.20 out of every $1? Is there leverage? Is this based on a FV vs NPV calculation? Thanks in advance.
This may be a silly question but if the data used in these models are biased - what outcome are you expecting and how do you say the outcome is trustworthy?
Yours is a clear-eyed analysis of the state of GenAI.
I feel very strongly that artificial general intelligence will someday be possible, but the current GenAI models are not on the right road, or even the right continent to achieve it. The incumbents know this. Two facts stand out, for example, in the GPT-4 technical report. First, the clear statement that these models are token guessers and second, that their progress is slowing. There is a figure that shows that going from GPT-3.5 to GPT-4 took 10,000 times as much compute to reduce the error on their chosen measure by 1 bit. According to that same curve, the next 1-bit reduction will require 100 quadrillion times as much compute. Satya Nadella has similarly noted that "intelligence" is proportional to the log of compute. That does not sound promising for scaling our way to general intelligence, even if scaling were enough.
The market for GenAI is projected to rise to $400 billion by 2031, but the capex is projected to be $4 trillion. As you note, there may not be enough of a market to justify these expenditures, and there may not be enough power to run them, even if they are built.
The loudest voices in this space predict that general intelligence is just around the corner, a prediction that has been made every year since at least 1956. Near as I can tell, these loud voices do not even know what it would take to achieve general intelligence, let alone know how anyone could actually build it. Here are a few suggestions (and the book from which it is excerpted). https://thereader.mitpress.mit.edu/ai-insight-problems-quirks-human-intelligence/
Readers may also find interesting an alternative approach to AI, based on representing facts rather than words as its basic units (https://www.reliath.ai/). This model uses a fraction of the compute, a fraction of the energy, needed to run transformer-based models, and it is incapable of hallucinating.
Global gross wage costs(not including non-wage labor costs) are roughly $55T vs global GDP of $110T. Lets assume that AI agents(which is not AGI) lead to some monetary gain that is combination of one-time productivity boost(PB) and wage cost reduction(WCR). This would be total monetary gain of $55T * PB% + $110T * WCR%. Obviously best case is all gain is from PB - for example 5% increase from $110T is $5.2T monetary gain with no change in wage costs. Lets get some rather arbitrary base case: 1% WCR and one-time 3% PB on average. That would lead to total monetary gain of $55T*1% + $110T*3% - roughly equal to $3.6T annual monetary gain. I assume no growth in wage cost and GDP, which will make the monetary gain even bigger. I am assuming one-time PB, which is worse case than compounding PB, and makes the monetary gain appear smaller. I am assuming that not all jobs will be affected by AI that is why put only 3% PB on aggregate.
This is rather simplified, maybe oversimplified, but it tries to illustrate that if AI agents are remotely useful and permeate the service economy, the annual monetary gains will be in the trillions. I fail to see how this is not a real possibility and therefore does not justify even $1T annual spending in the next 5-10 years.
When you say the bet is on reaching "AGI", I'm not sure what you exactly by that.
But if the question is what capability would we need to gain that would pay off this CapEx investment, I'd say continual learning on self-generated data seems to be the frontier. As it would break past current limitations on quantity and quality of existing data.
Thanks for sharing! I still remember when your article about the $600B gap was the first of its kind, back in 2024.
More and more are speaking up about the AI bubble that we are in, what it [CAPEX investment in data center] means, how severe this gap is... but what I feel was lacking is the now-what. What should we do as individuals/investors? What should we do as lawmakers? What should we do as a society? Would appreciate your thought leader here.
In your interview with Goldman Sachs you made an observation about the flow of dollars in the AI universe: "One observation that a lot of people have made is, if a dollar comes in at the top, Nvidia keeps $1.20 today. So Nvidia is capturing a lot of the value in the supply chain today." I'm not following... how is NVDA capturing $1.20 out of every $1? Is there leverage? Is this based on a FV vs NPV calculation? Thanks in advance.
Just read the interview and have the same question.
This may be a silly question but if the data used in these models are biased - what outcome are you expecting and how do you say the outcome is trustworthy?
Yours is a clear-eyed analysis of the state of GenAI.
I feel very strongly that artificial general intelligence will someday be possible, but the current GenAI models are not on the right road, or even the right continent to achieve it. The incumbents know this. Two facts stand out, for example, in the GPT-4 technical report. First, the clear statement that these models are token guessers and second, that their progress is slowing. There is a figure that shows that going from GPT-3.5 to GPT-4 took 10,000 times as much compute to reduce the error on their chosen measure by 1 bit. According to that same curve, the next 1-bit reduction will require 100 quadrillion times as much compute. Satya Nadella has similarly noted that "intelligence" is proportional to the log of compute. That does not sound promising for scaling our way to general intelligence, even if scaling were enough.
The market for GenAI is projected to rise to $400 billion by 2031, but the capex is projected to be $4 trillion. As you note, there may not be enough of a market to justify these expenditures, and there may not be enough power to run them, even if they are built.
The loudest voices in this space predict that general intelligence is just around the corner, a prediction that has been made every year since at least 1956. Near as I can tell, these loud voices do not even know what it would take to achieve general intelligence, let alone know how anyone could actually build it. Here are a few suggestions (and the book from which it is excerpted). https://thereader.mitpress.mit.edu/ai-insight-problems-quirks-human-intelligence/
Readers may also find interesting an alternative approach to AI, based on representing facts rather than words as its basic units (https://www.reliath.ai/). This model uses a fraction of the compute, a fraction of the energy, needed to run transformer-based models, and it is incapable of hallucinating.
David Cahn's writing is new to me. This is a comment on how most AI bloggers take AGI as a given possibility.
Does a tree falling in the woods make a sound?
Does a useless technology make a ripple in the real world?
The challenge - imagine you are an AGI. What could you do that would make life worth living?
AGI probably won’t be developed until at least the 2050s:
https://www.youtube.com/watch?v=3yEQaHvQxlE
Global gross wage costs(not including non-wage labor costs) are roughly $55T vs global GDP of $110T. Lets assume that AI agents(which is not AGI) lead to some monetary gain that is combination of one-time productivity boost(PB) and wage cost reduction(WCR). This would be total monetary gain of $55T * PB% + $110T * WCR%. Obviously best case is all gain is from PB - for example 5% increase from $110T is $5.2T monetary gain with no change in wage costs. Lets get some rather arbitrary base case: 1% WCR and one-time 3% PB on average. That would lead to total monetary gain of $55T*1% + $110T*3% - roughly equal to $3.6T annual monetary gain. I assume no growth in wage cost and GDP, which will make the monetary gain even bigger. I am assuming one-time PB, which is worse case than compounding PB, and makes the monetary gain appear smaller. I am assuming that not all jobs will be affected by AI that is why put only 3% PB on aggregate.
This is rather simplified, maybe oversimplified, but it tries to illustrate that if AI agents are remotely useful and permeate the service economy, the annual monetary gains will be in the trillions. I fail to see how this is not a real possibility and therefore does not justify even $1T annual spending in the next 5-10 years.
Good one! Thanks
Great post David
Just another Wall Street bubble with the minds being twisted into more leverage.
Gotta hope that continual learning / "learning from experience" (ie self generated data) starts to work.
https://substack.com/@dwarkesh/note/p-175283310?r=84yrx
https://arxiv.org/pdf/2508.05619
"https://storage.googleapis.com/deepmind-media/Era-of-Experience /The Era of Experience Paper.pdf"
When you say the bet is on reaching "AGI", I'm not sure what you exactly by that.
But if the question is what capability would we need to gain that would pay off this CapEx investment, I'd say continual learning on self-generated data seems to be the frontier. As it would break past current limitations on quantity and quality of existing data.
Great piece
The “deus ex machina” phrase for these dynamics is perfection.
Quaint, so fitting
Thanks for sharing! I still remember when your article about the $600B gap was the first of its kind, back in 2024.
More and more are speaking up about the AI bubble that we are in, what it [CAPEX investment in data center] means, how severe this gap is... but what I feel was lacking is the now-what. What should we do as individuals/investors? What should we do as lawmakers? What should we do as a society? Would appreciate your thought leader here.