Great read. Great take. Way to hang those balls out. I see the timeline as somewhere between tomorrow and January '28. CapEx borrowing vs. Treasury borrowing is a car crash in the offing. Sustained 4.8% 10 year Treasuries will kill companies with huge refinancing obligations. We're almost there. A few AI winners. A nation full of empty wallets.
The music is going to stop, there are not enough chairs to go around and the fed is so leveraged, no bailouts this time so everything will unwind and markets will reprice. Will be interesting to see who is left standing and who will be sacrificed.?
Totally agree (which given my track record should make us all uneasy! LOL)
At any rate, it's hard to know if you're "early" or "wrong".
I spent last week & weekend doing a deep dive (for me) into AI bubble thesis. I can't convince myself it's anything else.
The "AI Field of Dreams (Build it and they will come)" needs enormous profitable inference demand to justify the planned compute build. It has to come from consumers or business. If I do the numbers I can't get there either way. Its' so far short it's incredible. And the way to force it there is to assume: 1) consumers spend more on AI than on most of their entertainment combined or 2) business AI use explodes with agentic systems, which comes hand in hand with massive job losses. And even inference volume isn't enough, it has to be profitable volume. And inference prices are falling like rocks out of Galileo's leaning tower.
Already there's cultural AI revulsion taking form -- "AI slop" backlash and corporate employee resistance. Business use of AI at Field of Dreams level isn't a slam dunk.
It's technical, but LLMs have severe limitations as algorithms for artificial general intelligence (AGI). In short, they're mistake prone at a mathematical level that's impossible to fix. For fun and consumer use, whatever. For serious business use? Watch out.
That's not to say they're useless. I know enough about LLMs and programming to understand the tech-world fanboy energy. It's incredible to see your computer "do things" -- read/write files, web surf, "reason", code, write good prose -- on its own. Smart people (defined narrowly) can extrapolate their enthusiasm beyond reason. Or they can overthink and under-realize: Isaac Newton himself lost money in S. Sea Bubble.
To me, it all comes back to profitable inference rising to levels that I just can't see with any lucid thinking. From consumers? From business? It just doesn't pencil out, IMHO. When does that dream break down? I don't know but I agree it will.
I'm curious how you calculated possible business based income? We at my work have greatly leveraged the AI for SW/Firmware development . We are using the mid-tier levels at aprox $100 month. I use only one 2nd tier model. My co-part is using multiple, and has them check each other outputs. I think I could make a business case for $1000 month, per engineering person. At this basis, I am not sure how I would scale that upto aggregate business case with any kind of confidence? I have talked with our general manager on how she could utilize AI better? It is not clear for in that case, which means it is not a valuable of a business case. I agree, direct consumer income - likely irrelevant. The hyperscalers, like google and meta will likely be their own biggest customer of their compute. The AI needed to power the algorithm behind the media platforms.
great questions. I do coding myself (not a developer, semi retired "serious" amateur) and AI coding support is incredible. I can do things I'd never dreamed -- but I've also read lots of LLM-Hate by "real" developers. Hard to say. I just did a Google 5-day seminar on agentic coding, so it's not all Hate. There's adoption too.
Hard to forecast how it all sifts out in developer land.
My "Deep Dive" was this: I listened to lots of podcasts, saw compute build committments & associated revenue requirements quantified and defended/attacked (one source is uber-bear Ed Zitron who I don't doubt has done the work but his manner can be bit hypercritical, but there are others), I read stuff, analytically queried various scenarios (ironically) extensively using my Claude Pro subscription. I'm not saying LLMs aren't incredible, I am saying I don't trust the Bull thesis at the use-case & profitable inference level.
Would I bet my life I'm right? No. I wouldn't. Would I bet my investments, in the sense of "don't buy the top". Yes.
Spent several days on this. My take away was the "Profitable Inference Revenue" needed to support the bullish build is just not feasible unless AI inference grows to some unrealistic large fraction of consumer spending or if business use displaces so many workers with agentic AI that society revolts.
This isn't a 60-day research project. It's a semi-retired personal project, reading and thinking and relying on data provided by those who do claim to have done a rigorous deep dive. It would take me weeks to months to recreate the data sets I am to some extent relying on. So it's just a best guess on my part.
I'm not saying the bulls can't be right eventually. But to the macro points in QTR's post, to my tea leaf reading of use cases and profit needs, to tech bubble history (per QTRs review of that). All of it says to me "don't buy the hype or the top".
Thanks for the reply and thoughts. Reading your response, made me think maybe working it backwards is a reasonable approach - what percentage total business expense could be dedicated to AI purchases? Is that large enough to sustain the datacenters? Re-reading- I guess that is what you did. Looks unreasonable. I agree - "Profitable Inference Revenue" is likely the key metric in the long run.
Even if your timing isn't perfect the principles underlying your arguments are valid. I was in my 20s in 2007, had recently bought my first house, and expected life to be permanently good. My dad, who has worked with state DOTs for decades, was visiting and telling me he expected the financial shit to hit the fan within the next year. He said there were too many "weird things" going on with financing in the mortgage markets and in state governments. I thought he was being a downer; everything Dad predicted came to pass within in the next year. Fundamentals (like whether or not a business generates profit) eventually matter.
Good analysis and likely truer than we know. However, El Trumpo stands ready to infuse the data center/A/i buildout with trillions of taxpayer dollars should the 1,000 foot tall tsunami start to trickle down to a two foot beach breaker. This is his baby as he has declared that we must beat China at all costs in the A/i race. Stupid and silly but that doesn't stop a megalomaniac.
I don't believe A/i is the wave of the future as it is severely overhyped. It will never be sentient and will revert to being nothing more than a glorified Internet answer man complete with fake everything and highly censored as a major propaganda machine. I would buy long term puts on NVDA going out to the end of 2027/early 2028. As I cannot stomach the upward trend into exponential obscurity.
The gov is expected to borrow $739 billion for the final 2026 quarter (July-September) and another $628 billion for the first quarter (Oct-Dec) of fiscal 2027. The excuse is not enough revenues. Tax revenues are not quite up to snuff and that imaginary tariff windfall is actually now negative. It would seem the economy is not so vibrant as the gov paints it to be and the federal deficit will blow past $40 trillion by the end of August.
I see the gov and its Fed being squeezed on many fronts...rising deficits with rising interest rates (to refinance the borrowing), increasing war costs (El Trumpo is demanding $1.5 trillion for the military in 2027...about an increase of $400 billion), higher energy costs, a stifled or falling economy (less tax revenues), rising social security and medicare payouts (country getting older and sicker), more money for data centers, possible rescue of failing industries/investments (to keep the ball rolling), and a rising tide in falling confidence in the government.
Another consideration: Moore’s Law. Moore's Law is the empirical observation that the number of transistors on a dense integrated circuit doubles approximately every two years, leading to exponential growth in computing power and efficiency. Doubling the number of transistors quadruples the computer’s power. While the 1965 “law” specifically referred to transistors, it can more broadly be applied to computing power.
Speculation: What if next month China announces they’ve developed and manufactured an AI-specific chip that runs 5X the calculations in a clock cycle on 1/4 of the power draw? Instantly, all the Nvidia chips in AI racks across the world, plus all the unused chips sitting in warehouses waiting to be installed, become more-or-less obsolete. Nobody’s going to buy an IBM 286 PC/AT when a Dell Pentium III Core 2 Duo is half the price.
very plausible that will happen since tech obsolescence is ALWAYS a risk.
But "inference" can run on existing chips for many use cases. So older chips may not go obsolete and (price depending) it could be profitable compute.
IMHO use cases are key. does AI get embedded everywhere in everything so data centers are running lots of invisible inference load (your refrig, your car, your HVAC, your robot, your apps)? All that won't need cutting edge chips.
I think this is already happening in some applications. Voice-rec ordering at a fast-food drive-thru, for example, does not need to have the entire history of the world as its database for completing an order.
Of the big guys, these hyperscalers - I believe some of these will not survive. Or be severely damaged from the current boom cycle. In a year or so, someone will be getting infrastructure (ie compute) at cents on the dollar. It reminds me of Motorola/Iridium saga with satellite phones from 1990s/early 2000s. I have seen that play out in smaller area many times, unique tech, requires capex to build the infrastructure. One complete, the business case has shifted and we know the story ends - chapter 11. Infrastructure sold for a bargain. The buyer, with the lower costs, can usually then turn a profit.
I have actively tried to limit my exposure to the mag7 companies. I have sold off my SPY and QQQ holdings. I have moved some into RSP. I had planned on QEW but never executed. I have a "dividend growth" ETF that is still large cap heavy with mag7. But that is the balance I am taking at the moment.
I have some cash (5%), and also maintaining my precious metals monetary hedge (20%).
But, that all being said, I currently have 30% of my portfolio in Nebius (NBIS). This was a recent move, I still have one tranche that is negative when I started DCA into it. I do not have any CRWV, IREN or BRUN. I have my reasons for NBIS vs the others. I have read Michael Burry thesis and have watched Jim Chanos interview who specifically called out Nebius as a silly company. I would say I am fully informed and not just a 30 year old blind AI bull that has never experienced a significant downturn in the market.
I am not arguing against this article, I think Chris is really addressing the AI boom well.
The AI related stocks, memory, photonics/fiber are extremely volatile - definitely not for the faint of heart. >10% swings in a day are not uncommon. Wild times.
I’m not sure if you wrote this before the OpenAI executive team walked, but this seems even more of a precarious situation now than perhaps it was before.
So far from what I see on the current internet so-called AI is a bad joke as far as the "intelligence" aspect is concerned. Looks to me like a "tale full of sound and fury - signifying nothing." (Hat tip to Charles Hugh Smith as well; he calls the product "AI Slop"). Even if we're wrong there are huge hidden costs in the entire AI universe that are only now surfacing - for one thing the social costs of dumping huge electric power demands onto the existing grid and expecting existing consumers to pay for the buildout costs on their current monthly bills from already overstrained family budgets. That's not going over well in any sense! I'm with Chris on this one - I say - after the Midterms - if the Looney Left Dems win big then watch out below!
Excellent essay and fun read, though one in which I differ on the timing. I see the crash occurring sometime between later this year and 2034, around the time social security payouts tank and decrease the seniors that are collecting by about 24%.
A word of caution QTR, don't let those big balls dangle too far or steadily increasing S&P 500 earnings will smash into them with hurricane force rather than a cool refreshing breeze.
Excellent article! I forwarded it to my family’s Financial Advisor. He recently told me I’m holding too much cash and that the AI boom won’t end anytime soon.
I told him I’ll have the last laugh in twelve months or less. I. Immediately had him purchase the ETF: $HDGE
to hedge what small amount of equities I do own in my retirement account.
I’m 60 and I was around in March 2020 when the dotcom implosion occurred.
This period were in the very similar…. Your article is very timely!
Great read. Great take. Way to hang those balls out. I see the timeline as somewhere between tomorrow and January '28. CapEx borrowing vs. Treasury borrowing is a car crash in the offing. Sustained 4.8% 10 year Treasuries will kill companies with huge refinancing obligations. We're almost there. A few AI winners. A nation full of empty wallets.
Well said…
The music is going to stop, there are not enough chairs to go around and the fed is so leveraged, no bailouts this time so everything will unwind and markets will reprice. Will be interesting to see who is left standing and who will be sacrificed.?
Totally agree (which given my track record should make us all uneasy! LOL)
At any rate, it's hard to know if you're "early" or "wrong".
I spent last week & weekend doing a deep dive (for me) into AI bubble thesis. I can't convince myself it's anything else.
The "AI Field of Dreams (Build it and they will come)" needs enormous profitable inference demand to justify the planned compute build. It has to come from consumers or business. If I do the numbers I can't get there either way. Its' so far short it's incredible. And the way to force it there is to assume: 1) consumers spend more on AI than on most of their entertainment combined or 2) business AI use explodes with agentic systems, which comes hand in hand with massive job losses. And even inference volume isn't enough, it has to be profitable volume. And inference prices are falling like rocks out of Galileo's leaning tower.
Already there's cultural AI revulsion taking form -- "AI slop" backlash and corporate employee resistance. Business use of AI at Field of Dreams level isn't a slam dunk.
It's technical, but LLMs have severe limitations as algorithms for artificial general intelligence (AGI). In short, they're mistake prone at a mathematical level that's impossible to fix. For fun and consumer use, whatever. For serious business use? Watch out.
That's not to say they're useless. I know enough about LLMs and programming to understand the tech-world fanboy energy. It's incredible to see your computer "do things" -- read/write files, web surf, "reason", code, write good prose -- on its own. Smart people (defined narrowly) can extrapolate their enthusiasm beyond reason. Or they can overthink and under-realize: Isaac Newton himself lost money in S. Sea Bubble.
To me, it all comes back to profitable inference rising to levels that I just can't see with any lucid thinking. From consumers? From business? It just doesn't pencil out, IMHO. When does that dream break down? I don't know but I agree it will.
I'm curious how you calculated possible business based income? We at my work have greatly leveraged the AI for SW/Firmware development . We are using the mid-tier levels at aprox $100 month. I use only one 2nd tier model. My co-part is using multiple, and has them check each other outputs. I think I could make a business case for $1000 month, per engineering person. At this basis, I am not sure how I would scale that upto aggregate business case with any kind of confidence? I have talked with our general manager on how she could utilize AI better? It is not clear for in that case, which means it is not a valuable of a business case. I agree, direct consumer income - likely irrelevant. The hyperscalers, like google and meta will likely be their own biggest customer of their compute. The AI needed to power the algorithm behind the media platforms.
great questions. I do coding myself (not a developer, semi retired "serious" amateur) and AI coding support is incredible. I can do things I'd never dreamed -- but I've also read lots of LLM-Hate by "real" developers. Hard to say. I just did a Google 5-day seminar on agentic coding, so it's not all Hate. There's adoption too.
Hard to forecast how it all sifts out in developer land.
My "Deep Dive" was this: I listened to lots of podcasts, saw compute build committments & associated revenue requirements quantified and defended/attacked (one source is uber-bear Ed Zitron who I don't doubt has done the work but his manner can be bit hypercritical, but there are others), I read stuff, analytically queried various scenarios (ironically) extensively using my Claude Pro subscription. I'm not saying LLMs aren't incredible, I am saying I don't trust the Bull thesis at the use-case & profitable inference level.
Would I bet my life I'm right? No. I wouldn't. Would I bet my investments, in the sense of "don't buy the top". Yes.
Spent several days on this. My take away was the "Profitable Inference Revenue" needed to support the bullish build is just not feasible unless AI inference grows to some unrealistic large fraction of consumer spending or if business use displaces so many workers with agentic AI that society revolts.
This isn't a 60-day research project. It's a semi-retired personal project, reading and thinking and relying on data provided by those who do claim to have done a rigorous deep dive. It would take me weeks to months to recreate the data sets I am to some extent relying on. So it's just a best guess on my part.
I'm not saying the bulls can't be right eventually. But to the macro points in QTR's post, to my tea leaf reading of use cases and profit needs, to tech bubble history (per QTRs review of that). All of it says to me "don't buy the hype or the top".
Could I be wrong? Sure.
Thanks for the reply and thoughts. Reading your response, made me think maybe working it backwards is a reasonable approach - what percentage total business expense could be dedicated to AI purchases? Is that large enough to sustain the datacenters? Re-reading- I guess that is what you did. Looks unreasonable. I agree - "Profitable Inference Revenue" is likely the key metric in the long run.
Even if your timing isn't perfect the principles underlying your arguments are valid. I was in my 20s in 2007, had recently bought my first house, and expected life to be permanently good. My dad, who has worked with state DOTs for decades, was visiting and telling me he expected the financial shit to hit the fan within the next year. He said there were too many "weird things" going on with financing in the mortgage markets and in state governments. I thought he was being a downer; everything Dad predicted came to pass within in the next year. Fundamentals (like whether or not a business generates profit) eventually matter.
That's even more impressive when you take into account the limited information available to us in 2007. Hats off to your dad.
Good analysis and likely truer than we know. However, El Trumpo stands ready to infuse the data center/A/i buildout with trillions of taxpayer dollars should the 1,000 foot tall tsunami start to trickle down to a two foot beach breaker. This is his baby as he has declared that we must beat China at all costs in the A/i race. Stupid and silly but that doesn't stop a megalomaniac.
I don't believe A/i is the wave of the future as it is severely overhyped. It will never be sentient and will revert to being nothing more than a glorified Internet answer man complete with fake everything and highly censored as a major propaganda machine. I would buy long term puts on NVDA going out to the end of 2027/early 2028. As I cannot stomach the upward trend into exponential obscurity.
The gov is expected to borrow $739 billion for the final 2026 quarter (July-September) and another $628 billion for the first quarter (Oct-Dec) of fiscal 2027. The excuse is not enough revenues. Tax revenues are not quite up to snuff and that imaginary tariff windfall is actually now negative. It would seem the economy is not so vibrant as the gov paints it to be and the federal deficit will blow past $40 trillion by the end of August.
I see the gov and its Fed being squeezed on many fronts...rising deficits with rising interest rates (to refinance the borrowing), increasing war costs (El Trumpo is demanding $1.5 trillion for the military in 2027...about an increase of $400 billion), higher energy costs, a stifled or falling economy (less tax revenues), rising social security and medicare payouts (country getting older and sicker), more money for data centers, possible rescue of failing industries/investments (to keep the ball rolling), and a rising tide in falling confidence in the government.
Which is well deserved.
Another consideration: Moore’s Law. Moore's Law is the empirical observation that the number of transistors on a dense integrated circuit doubles approximately every two years, leading to exponential growth in computing power and efficiency. Doubling the number of transistors quadruples the computer’s power. While the 1965 “law” specifically referred to transistors, it can more broadly be applied to computing power.
Speculation: What if next month China announces they’ve developed and manufactured an AI-specific chip that runs 5X the calculations in a clock cycle on 1/4 of the power draw? Instantly, all the Nvidia chips in AI racks across the world, plus all the unused chips sitting in warehouses waiting to be installed, become more-or-less obsolete. Nobody’s going to buy an IBM 286 PC/AT when a Dell Pentium III Core 2 Duo is half the price.
very plausible that will happen since tech obsolescence is ALWAYS a risk.
But "inference" can run on existing chips for many use cases. So older chips may not go obsolete and (price depending) it could be profitable compute.
IMHO use cases are key. does AI get embedded everywhere in everything so data centers are running lots of invisible inference load (your refrig, your car, your HVAC, your robot, your apps)? All that won't need cutting edge chips.
I don't see it happening but I may be wrong.
I think this is already happening in some applications. Voice-rec ordering at a fast-food drive-thru, for example, does not need to have the entire history of the world as its database for completing an order.
Of the big guys, these hyperscalers - I believe some of these will not survive. Or be severely damaged from the current boom cycle. In a year or so, someone will be getting infrastructure (ie compute) at cents on the dollar. It reminds me of Motorola/Iridium saga with satellite phones from 1990s/early 2000s. I have seen that play out in smaller area many times, unique tech, requires capex to build the infrastructure. One complete, the business case has shifted and we know the story ends - chapter 11. Infrastructure sold for a bargain. The buyer, with the lower costs, can usually then turn a profit.
I have actively tried to limit my exposure to the mag7 companies. I have sold off my SPY and QQQ holdings. I have moved some into RSP. I had planned on QEW but never executed. I have a "dividend growth" ETF that is still large cap heavy with mag7. But that is the balance I am taking at the moment.
I have some cash (5%), and also maintaining my precious metals monetary hedge (20%).
But, that all being said, I currently have 30% of my portfolio in Nebius (NBIS). This was a recent move, I still have one tranche that is negative when I started DCA into it. I do not have any CRWV, IREN or BRUN. I have my reasons for NBIS vs the others. I have read Michael Burry thesis and have watched Jim Chanos interview who specifically called out Nebius as a silly company. I would say I am fully informed and not just a 30 year old blind AI bull that has never experienced a significant downturn in the market.
I am not arguing against this article, I think Chris is really addressing the AI boom well.
The AI related stocks, memory, photonics/fiber are extremely volatile - definitely not for the faint of heart. >10% swings in a day are not uncommon. Wild times.
FYI - for Iridium background found this https://www.iridiummuseum.com/timeline/
The $5B constellation was purchased for $25M. 0.5 cents per $1 build.
It already started.
Just like the amplifiers in Spinal Tap, T-Bill yields are going to 11.
I’m not sure if you wrote this before the OpenAI executive team walked, but this seems even more of a precarious situation now than perhaps it was before.
So far from what I see on the current internet so-called AI is a bad joke as far as the "intelligence" aspect is concerned. Looks to me like a "tale full of sound and fury - signifying nothing." (Hat tip to Charles Hugh Smith as well; he calls the product "AI Slop"). Even if we're wrong there are huge hidden costs in the entire AI universe that are only now surfacing - for one thing the social costs of dumping huge electric power demands onto the existing grid and expecting existing consumers to pay for the buildout costs on their current monthly bills from already overstrained family budgets. That's not going over well in any sense! I'm with Chris on this one - I say - after the Midterms - if the Looney Left Dems win big then watch out below!
Brilliant. Needed an article like this to reset, rethink and make decisions that means you can survive this coming debt fueled dreams.
Perfect analysis, all time best ball hanging article.
"...all time best ball hanging article"
Excellent essay and fun read, though one in which I differ on the timing. I see the crash occurring sometime between later this year and 2034, around the time social security payouts tank and decrease the seniors that are collecting by about 24%.
A word of caution QTR, don't let those big balls dangle too far or steadily increasing S&P 500 earnings will smash into them with hurricane force rather than a cool refreshing breeze.
Excellent article! I forwarded it to my family’s Financial Advisor. He recently told me I’m holding too much cash and that the AI boom won’t end anytime soon.
I told him I’ll have the last laugh in twelve months or less. I. Immediately had him purchase the ETF: $HDGE
to hedge what small amount of equities I do own in my retirement account.
I’m 60 and I was around in March 2020 when the dotcom implosion occurred.
This period were in the very similar…. Your article is very timely!
Great disclaimer…lol
"Transformative technology and catastrophic overinvestment have always been perfectly capable of occupying the same room." - Nailed it!!