Excellent analysis. Berkshire Hathaway recently bought 10B worth of Google shares at 350/share in private placement. Surely they are aware of everything mentioned here. Are they betting that this will all work out for them in the long run? Or, did the new CEO succumb to AI FOMO?
Very thorough and well founded analysis. Another interesting aspect is the evaluation of the limits of possible AI demand. My fear is that AI may turn out be a giant machine that converts energy and matter into waste heat, slop, shitty code, and porn. There will be some real value of course, but the EROI of the whole system may turn out to be negative, limiting demand and usefulness. On the other hand, if AI turns out to be ultimately useful, it will lead to severe unemployment,, which in turn will decrease consumer demand and feed back into limited end demand - corporations can save labor costs, but who will they sell their products to, if they destroy consumer demand in the process?
you must not be using AI correctly - in my experience it has changed the game in terms of being able to ship out products/projects/analysis. Things that would have taken weeks done in days. It is already useful and I am still employed. The unemployment argument I think becomes stronger when AI can think for itself - a lot of the usefulness I get comes from what I already know and the quality of the questions and prompts I direct. But the ability of AI to self direct projects and apps ultimately is a question of whether you get AGI and better off just hiring an Agent than a skilled person who can leverage AI. For now we are in the latter - and I can guarantee it is making life more efficient.
Hi Peter. AI will help the corporations. But it will cause massive unemployment for the workers. People do not understand that tech is thier enemy. Not thier friend.
Notice that consumers constantly need to buy more tech equipment due to obsolescence? Notice that consumers need to constantly learn new tech?
This should be a big tip off. Tech is no friend of the people. MSFT takes over your computer, and charges you for the privilege.
People don't care. Tech is sexy. You don't want to talk about a ball bearings company at the party.
Fear not. AI is useful now, and getting better, no unemployment necessary.
My software business is using AI to accelerate output, not remove employees. Granted a software business is an early adopter and many LLMs are tuned for code, but these tools have applications in other businesses to accelerate output.
I've reduced the work I paid my contract lawyer for by two-thirds, using AI. I imagine he can use it to get 3x the work done without having to hire a junior attorney.
Coding and testing has become much more efficient. Junior coders +Claude become as capable as senior coders (less getting stuck, less calling seniors in for how-to questions). Senior coders have doubled output (pull requests/commits). AI writes the unit tests (annoying but necessary work senior codes don't like to do). AI records and reruns QA test cases.
We can also get projects done which were impossible to tackle before without these logical machines. E.g. refactor a large complex code set (pay technical debt) and validate outputs to a hash - achieving 100% translation (impossible before considering human error), translate a set of HTML and Word docs into JSON docs, again with 100% fidelity (multiple rounds of human tuning were required).
We are leveraging AI to offload support cases - handling the easy and annoying ones which could have been solved is users bothered reading the documentation. We are now moving it into Sales and Marketing - not for writing, although that can be useful in some specialized SEO/GEO cases, but more for process automations, data validation and pattern recognition.
The AI tools are best leveraged with skilled operators, not as self-directing robots, and those operators become more efficient at their work. It's worthwhile to invest in AI skills/prompt training, and AI works best when there is a deterministic validation, lest AI run off at the mouth.
"because the end-demand that pays for it is bounded."
I think this is probably the crux of the matter- where that boundary level lies. I also can't emphasize more the issue you point out about how peak earnings in cyclicals drive down the PE to "cheap" levels- a lesson I learned in 2000 bubble. The problem is how growth companies are cyclicals in disguise at some point.
Great analysis. The point about “cheap” valuations being supported by peak-cycle earnings deserves more attention.
Where I differ slightly is the 2008 comparison. The collateral today isn’t empty housing stock. It’s infrastructure that businesses are actively using and paying for.
That doesn’t remove the risk. It just changes the failure mode. A slowdown in demand, lower GPU pricing, or weaker utilisation rates could create a painful shakeout for heavily leveraged neoclouds.
Feels less like a technology bubble and more like a capital allocation cycle that could end with a few winners and a lot of casualties.
Excellent analysis. Berkshire Hathaway recently bought 10B worth of Google shares at 350/share in private placement. Surely they are aware of everything mentioned here. Are they betting that this will all work out for them in the long run? Or, did the new CEO succumb to AI FOMO?
I like Google. It's not how high can the price go? It's how not bad the price goes down.
Very thorough and well founded analysis. Another interesting aspect is the evaluation of the limits of possible AI demand. My fear is that AI may turn out be a giant machine that converts energy and matter into waste heat, slop, shitty code, and porn. There will be some real value of course, but the EROI of the whole system may turn out to be negative, limiting demand and usefulness. On the other hand, if AI turns out to be ultimately useful, it will lead to severe unemployment,, which in turn will decrease consumer demand and feed back into limited end demand - corporations can save labor costs, but who will they sell their products to, if they destroy consumer demand in the process?
you must not be using AI correctly - in my experience it has changed the game in terms of being able to ship out products/projects/analysis. Things that would have taken weeks done in days. It is already useful and I am still employed. The unemployment argument I think becomes stronger when AI can think for itself - a lot of the usefulness I get comes from what I already know and the quality of the questions and prompts I direct. But the ability of AI to self direct projects and apps ultimately is a question of whether you get AGI and better off just hiring an Agent than a skilled person who can leverage AI. For now we are in the latter - and I can guarantee it is making life more efficient.
Hi Peter. AI will help the corporations. But it will cause massive unemployment for the workers. People do not understand that tech is thier enemy. Not thier friend.
Notice that consumers constantly need to buy more tech equipment due to obsolescence? Notice that consumers need to constantly learn new tech?
This should be a big tip off. Tech is no friend of the people. MSFT takes over your computer, and charges you for the privilege.
People don't care. Tech is sexy. You don't want to talk about a ball bearings company at the party.
Fear not. AI is useful now, and getting better, no unemployment necessary.
My software business is using AI to accelerate output, not remove employees. Granted a software business is an early adopter and many LLMs are tuned for code, but these tools have applications in other businesses to accelerate output.
I've reduced the work I paid my contract lawyer for by two-thirds, using AI. I imagine he can use it to get 3x the work done without having to hire a junior attorney.
Coding and testing has become much more efficient. Junior coders +Claude become as capable as senior coders (less getting stuck, less calling seniors in for how-to questions). Senior coders have doubled output (pull requests/commits). AI writes the unit tests (annoying but necessary work senior codes don't like to do). AI records and reruns QA test cases.
We can also get projects done which were impossible to tackle before without these logical machines. E.g. refactor a large complex code set (pay technical debt) and validate outputs to a hash - achieving 100% translation (impossible before considering human error), translate a set of HTML and Word docs into JSON docs, again with 100% fidelity (multiple rounds of human tuning were required).
We are leveraging AI to offload support cases - handling the easy and annoying ones which could have been solved is users bothered reading the documentation. We are now moving it into Sales and Marketing - not for writing, although that can be useful in some specialized SEO/GEO cases, but more for process automations, data validation and pattern recognition.
The AI tools are best leveraged with skilled operators, not as self-directing robots, and those operators become more efficient at their work. It's worthwhile to invest in AI skills/prompt training, and AI works best when there is a deterministic validation, lest AI run off at the mouth.
Exactly Peter. This will happen.
Very well done - clearly articulated
"because the end-demand that pays for it is bounded."
I think this is probably the crux of the matter- where that boundary level lies. I also can't emphasize more the issue you point out about how peak earnings in cyclicals drive down the PE to "cheap" levels- a lesson I learned in 2000 bubble. The problem is how growth companies are cyclicals in disguise at some point.
Well done. Probably one of the most important pieces I've read in the last few years. I'm sharing it with my readers.
Very thought provoking!
excellent piece
Wow! Groundbreaking is really smart! Very scary article!
Very compelling!
Solid.
Great article. Here is a perspective from a few weeks ago. We will only build a small fraction of announced AI data centers.
https://needsofthemany98.substack.com/p/the-two-bubbles-america-is-building?r=gwg0&utm_campaign=post&utm_medium=web
Thank you.
The fantastic writing ..
Thanks so much to sharing it here..
Critical thinking. I agree on the circular revenue but do they really usd GPUs as collateral ? That sounds crazy.
Great analysis. The point about “cheap” valuations being supported by peak-cycle earnings deserves more attention.
Where I differ slightly is the 2008 comparison. The collateral today isn’t empty housing stock. It’s infrastructure that businesses are actively using and paying for.
That doesn’t remove the risk. It just changes the failure mode. A slowdown in demand, lower GPU pricing, or weaker utilisation rates could create a painful shakeout for heavily leveraged neoclouds.
Feels less like a technology bubble and more like a capital allocation cycle that could end with a few winners and a lot of casualties.
Great analysis. However without an estimate of earnings inflation through the 3 channels, it is difficult to use this information in a practical sense