When the Great Recession happened in 2008, to most people it came out of the blue. "No one could have seen this coming," said the cable TV channels. In reality, there were plenty of people calling it out in 2006-2008.
This is the best piece that I have seen explaining what to watch for to see it all happen again with AI.
I love Substack because incredible people take the time to share their insights, and there is an audience of people who read it. This is one of those pieces that is a great investment of your time and effort to read and to understand. Because, sooner or later, this will be a key factor in an economic recession.
Groundbreaker relies on first-semester calculus here.
The basic principle is this. Think of using your car. The distance you travel, your speed, and how fast you are breaking (or stomping on the gas) are three different things, right? But they are related. Speed measures how far you can go in an increment of time. High speed means that you are covering ground really fast; but it doesn't mean that you necessarily are covering a lot of ground (because that depends on how long you keep it up).
You can be going fast, but suddenly a stoplight changes color and you stomp on the brake. You were going fast, but only covered a short distance, right? So even though your speed was high, the distance was short. That is because your acceleration was great. What?? How can your acceleration be great when you slowed down? Because it was "negative" acceleration.
I had a typo as I wrote that paragraph to say that "suddenly a stoplight changes color and you stomp on the gas." In that case, the acceleration is significant and in the "positive" direction. You have positive acceleration. This increases your speed. And you keep going, so you increase your distance as well.
The article's point is to understand what this means for the AI industry. In the case of AI, distance correlates to the total number of data centers and compute. Speed corresponds to how fast data centers are being added. And acceleration corresponds to changes in how fast data centers are being added. When data centers are not being added as fast, that will upset the apple cart of how their financing works. And this is inevitable.
F**k f**king substack’s ios app. I merely glance at my unposted comment the wrong way and it slides off the page into oblivion.
Sigh…typing for the third time - a friendly counterpoint: I think ‘the Big Short’ has skewed memories of the Great Recession.
There was nothing sudden about it at all by my recollection. Subprime delinquencies dominated headlines starting in 2006. Lenders and hedge funds failed throughout 2007. Bear Stearns was on its death bed for a solid year mid 07 to mid 08z Anyone surprised by Lehman and Merrill bankruptcies in Sept 2008 was seriously not paying attention…
Most people don't pay attention, then or now. That's what scares me about AI, actually. People who don't think won't understand that it doesn't think, either. AI is Sean Hannity with Tucker Carlson's demeanor. (Ha. I like that one.)
Re no one could see this coming. That is ChatGPT’s version of the housing crash. Which just goes to show it is a ginormous internet word salad regurgitation machine.
Trump calls everybody "communists", but under Trump, the government is Intel's largest shareholder (and he strong-arms other companies to buy Intel products). The government is also a major shareholder of MP Materials and of US Steel and gets a cut of Nvidia's sales.
Very insightful article. I especially like the second-derivative framework — the idea that markets often break not when growth stops, but when the acceleration of growth disappoints.
My only challenge with applying that framework too aggressively to AI capex is that it risks treating a multi-dimensional phenomenon as a single-variable equation.
Yes, AI infrastructure spending growth will eventually decelerate. Mathematically, it has to. No exponential curve keeps accelerating forever. But a slowdown in the rate of investment growth does not necessarily imply a slowdown in the surface area of impact.
A useful comparison is exactly the one you used - the subprime mortgage crisis. That was largely a one-dimensional leverage cycle: more credit → more housing demand → higher home prices → more collateral value → more credit. Once the underlying assumption (housing prices only go up) broke, the feedback loop reversed. The second derivative told you almost everything because the system depended heavily on one dominant variable: the continuous expansion of mortgage credit.
AI capex feels fundamentally different.
The current investment wave is not only chasing one asset class or one revenue stream; it is building a new computational layer. The second-order and third-order effects are spread across productivity, software, robotics, science, healthcare, energy, education, manufacturing, defense, and entirely new categories we probably have not imagined yet.
Ten years from now, there is a reasonable chance we look back at today’s “AI use cases” the same way we look at early internet websites — useful demonstrations, but nowhere close to the final economic architecture.
The risk is real: overinvestment, misallocated capital, and unrealistic timelines can absolutely create painful corrections. But the harder question is not simply “will AI capex acceleration slow?”
It is: “how much economic surface area gets created before and after that slowdown happens?”
The second derivative may capture the investment cycle.
Are you saying "this time is different"? The point of the author's analysis is that this is, for all practical purposes, a credit bubble a la 2008. The homes built at that time eventually found long term owners. It was the transition from pre-2008 to post 2008 that was painful. AI use cases may indeed explode just as internet use cases did, but the precondition to that massive adoption maybe a dramatic drop in token pricing precipitated by the bubble bursting. It's the mechanics of a potential burst we are pondering here--and the entire focus of the analysis.
Yes, that could absolutely happen — and I agree the transition mechanics matter. A capex cycle can be directionally right in the long run and still create enormous pain if capital allocation gets ahead of near-term monetization.
My point is slightly different: the analysis feels incomplete if it only models the liability side of the equation and not the asset being created.
In 2008, the world discovered that a significant portion of the “asset” backing the credit expansion was mispriced housing demand. The houses remained useful, yes, but the marginal economic productivity unlocked by an extra subdivision was relatively limited.
AI infrastructure is a different question. The debate is whether we are overbuilding empty capacity or whether we are building a new productivity layer. If the second is true, then some portion of this capex has to eventually appear on the world’s balance sheet as productivity expansion, lower costs, faster innovation cycles, new businesses, and new economic output.
Almost nobody knows the magnitude or timing of that yet — and that uncertainty is my point.
If AI capex were purely a 2008-style credit bubble with no offsetting productivity asset, you could argue it should have broken already because the gap between investment and immediate cash flow is historically extreme. The fact that it has persisted this long is itself information: markets may be pricing something beyond the traditional bubble template.
The question is not “is this time different?”
The better question is: “what variable exists in this cycle that was missing in previous cycles?”
For AI, the answer may be the scale and speed of the productivity feedback loop. Whether that is large enough to justify the current capex spend is the trillion-dollar question…
MLPortfolios you have the right question. I want to add a third position to you and Krishna's debate, if I may
The bear watches the credit structure and says it breaks on deceleration. The bull watches the productivity feedback loop and says the missing variable justifies the capex. Both are right on their own terms.
But both assume a knowable revenue ceiling. The bear needs it to price credit risk. The bull needs it to justify the multiple. Neither asks whether the ceiling itself is the assumption that breaks.
AlphaFold had no knowable revenue ceiling before it ran. The discoveries it made possible were not in any model because they had not been made yet. If AGI-adjacent systems do that across multiple domains simultaneously, the feedback loop you describe cannot be sized using current instruments.
Krishna is right that the precondition for massive adoption may be the bubble bursting. The dotcom infrastructure survived. The financing around it did not. Survivors bought stranded assets for pennies.
Both can be true simultaneously. Credit structure breaks on deceleration while long-run productive value remains unbounded. That is the position the retail investor inside this through an index fund has no framework to navigate.
Thanks Sovereign. I really like this framing and find it quite illuminating as It deftly reconciles what initially looked like opposing views.
Where I’d extend your idea is this: perhaps the real bottleneck isn’t estimating the revenue ceiling, but measuring the rate at which the ceiling itself expands?
Historically, we’ve valued businesses on relatively stable addressable markets. AI may be changing the market while simultaneously participating in it. Every productivity gain increases humanity’s capacity to create the next productivity gain. That’s a recursive system, not a linear one.
I also agree that a bubble and a productivity revolution are not mutually exclusive. The dot-com analogy is instructive: the financing layer can implode while the infrastructure layer compounds for decades. Credit can be catastrophically wrong while technology is directionally right.
What fascinates me is the missing state variable.
Markets are exceptionally good at pricing cash flows. They’re much less equipped to price new possibility space (I am one of those who strongly believe that new technology should be priced more like options rather than using using traditional tools such as DCF) —economic activity that doesn’t yet exist because the tools to create it don’t yet exist.
If that’s the regime we’re entering, then both the bull and the bear are working with incomplete state vectors. The debate shifts from “Who is right?” to “What variables are our models still missing?”
That, to me, is the more interesting research problem.
MLPortfolios your options point is right. DCF requires a terminal state. Options pricing requires only a distribution of possible states and a timeframe. If the ceiling expands recursively, the terminal state is undefined by construction and DCF cannot reach it.
The missing state variable point is where I would go further. Markets are not just less equipped to price new possibility space. They are structurally incapable using current instruments, because those instruments require knowable cash flows in knowable timeframes. That is not a calibration problem. It is a category error.
Which brings it to the practical question for the investor sitting inside this right now. If both bull and bear are working with incomplete state vectors, and the market is pricing the asset using instruments that cannot reach the relevant variable, then conviction framework matters more than the valuation model.
You cannot know the ceiling. You can know what you believe about the recursive system, document it, and define the conditions under which that belief would be falsified. That is a different kind of analysis from DCF. It is the only one that works when the state vector is incomplete.
Alas, we converge. This has been my point all along.
If the state vector is incomplete, then your positioning in the AI capex trade inevitably becomes conviction-based rather than valuation-based.
But here’s the question almost no one asks: where does conviction come from?
My observation is that conviction is rarely created by reading. It is created by experience.
If you’re only consuming articles about AI, your conviction is largely borrowed from other people’s models. It will naturally be fragile because every new headline can change your mind.
If, however, you’re a daily power user building with AI, watching it fail, improve, surprise you, and reshape your own productivity, your conviction becomes grounded in first-hand evidence. It doesn’t make you right—it simply means your priors come from direct interaction rather than second-hand narratives.
Ironically, this is probably why the AI debate is so polarized. Many of the strongest bulls are heavy users. Many of the strongest bears evaluate AI primarily through financial statements, capex charts, and valuation multiples. Both are observing something real—but through different lenses.
The truth, as usual, is probably somewhere in between.
My advice to anyone navigating this trade is simple: first identify your bias. Bullish, bearish, or indifferent. Then actively seek evidence that could falsify your position. If you’re bullish, spend time with the best bearish arguments. If you’re bearish, spend time building real things with AI.
When we look back a decade from now, I suspect both bulls and bears will have been right—and wrong—in different ways.
The biggest winners may not be those who picked a side. They may be those who understood the asymmetry, managed their risk accordingly, and positioned themselves to benefit regardless of which part of the thesis unfolded first…
I see the innovation curve as a echo of the high-speed internet build out in the early 2000s - fiber optic cable was the game changer for high speed internet but the economics played out different.
I actually like this fibre analogy as it is probably the closest historical precedent.
The difference, though, I think, is that fiber primarily reduced the cost of moving information. AI has the potential to reduce the cost of producing intelligence itself and the distinction between both matters more than we think.
When bandwidth became abundant, we didn’t just browse websites faster, rather we got streaming, cloud computing, social media, mobile apps, and entire industries that barely existed before.
If intelligence follows a similar cost curve, today’s use cases may end up looking like email did in the early internet era i.e. important, but we maybe looking at only a tiny glimpse of what the infrastructure may ultimately enable.
Ironically, that doesn’t invalidate the economics either. The fiber build out was painful for many investors, but transformative for society. AI may follow a similar path where the infrastructure outlasts the financing cycle.
The hardest part, as investors, is separating the timing of returns from the permanence of the technology. History suggests those two are often very different.
Overbuild the houses and no one will use them, their marginal value drops to 0. Overbuild the AI compute layer and the marginal value of the chips, while still drops, doesn't go to 0. People will always find a way to use it. So I think while there's merits to compare with 2008, it's not complete analysis to not look at the demand part of the equation. Real compute demand, as you say, in different industries. Right now it's mostly consumers and coding, but there would be more overtime. So only looking at second derivatives is really over-simplifying.
Exactly. I think you’ve put your finger on the variable that’s easiest to overlook (and explained it so simply): demand isn’t static.
An oversupplied housing market and an oversupplied compute market don’t necessarily behave the same way.
An empty house creates very little new demand simply because it exists. Compute is different. As its marginal cost falls, entrepreneurs, researchers and businesses find entirely new ways to consume it. History is full of examples where cheaper computation didn’t just satisfy existing demand—it created demand that didn’t previously exist.
That doesn’t mean AI capex can’t overshoot. It absolutely can. Some projects will destroy capital, and some companies will almost certainly overbuild.
But I struggle with analyses that treat compute as if it’s just another cyclical asset. Compute is a general-purpose input into innovation itself. That changes the demand equation in ways our traditional bubble frameworks don’t fully capture.
To me, that’s the missing state variable. The question isn’t only “How much compute have we built?” It’s also “What new economic activity does cheaper, more abundant compute unlock?”
That’s a much harder variable to model—and probably the one that matters most…
That's very eloquently put. I think that's where bulls and bears differ. Bulls assume new high value use cases will spring up right on time to absorb extra compute supply. Bears treat total demand as deterministic, and once supply is overbuilt, the whole thing collapses. I think both are wrong, and we will be experiencing both compute under-supply (as of now), and compute oversupply (maybe in 2027), but then demand will catch up again because who doesn't need more intelligence? Right now it doesn't seem anyone has an answer yet.
What strikes me is that everyone wants a static answer to what is fundamentally a dynamic problem.
Compute can be undersupplied in one regime, oversupplied in the next, and then undersupplied again as new applications emerge. Those states aren’t contradictory—they’re part of the same feedback loop.
Perhaps the real mistake is asking, “Will there be oversupply?” There almost certainly will.
The better question is, “How quickly does intelligence create the next wave of demand for intelligence?”
That’s the variable I don’t think we’ve measured well yet. And until we do, both the bull and bear cases will remain incomplete—which is exactly what makes this one of the most fascinating investment problems of our generation…
Chock full o' LLM gobbledygook. I couldn't make it past the 50% mark, all the "it's not this... it's that" phrases and techno-econ jargon were making me nauseous. Which is too bad because I thought it had some good ideas and views on how this AI financing vicious cycle is playing out. The article just needs some serious human touch and de-Dickens'ing.
Yeah glad someone else noticed. May well contain some truth but also surely contains bs. Ai slop, not written by a human author who understands things.
In 2006, the only growing sector of the economy was homebuilding. There is no inherent value creation; we use already-created value to invest in a home. If no value is being created, all you can do is to borrow from the future. Entropy says that this can't extend forever.
There is some value being created by AI. What that value is will probably play an important role in how long this can be sustained.
Relatively low value compounding, though. Building housing doesn't raise productivity much (which is ultimately what drives growth) and is the difference between housing and AI (or the 2000 boom).
Most "value creation" is terminal consumer goods rather than investment - about 3x from GDP figures.
Still, it's not exactly the case that housing is a pure consumer good. More housing in more productive cities in particular can create compounding returns as it allows workers to produce more, just like regular old technology or capital investments.
I think you are quite correct. About the vulnerability, the mechanism and the trigger (OpenAI). I do not know how to answer the timing question. And I need conviction on the timing to make this a tradable premise.
So I am watching carefully. Sentiment is shifting, has shifted. OpenAI is the canary.
This article is why a reminder of everything I hate about AI and how users use it to write. There are a good number of points that stand strong in isolation in the article. However, they are misapplied without sufficient specific context. Overfitting is the term of use. The second derivative idea is very useful to spot blow off tops in markets but the limit is that it generally requires rapid feeback in the form of price discovery. The second derivative price applied to subprime is a single factor in what likely was a multi factor model (think no income and no doc mortgages) even if the housing bubble had enough volume for price discovery to offset the high transaction costs that usually kill these models. I would love to see the 2nd derivative idea backtested against other housing bubbles while statistically adjusting for other explanatory factors. And be sure to strictly apply the null hypothesis or whatever outcome you are testing for (is it price increases or rising deliquencies etc). Another problem is that AI DC and GPU backed loans are not subprime resi housing. Commercial RE lands closer. Infra loans is even better. Fracking (bc of the similar declines) loans for GPUs. LNG, midstream, and other LT loans for the DCs. The biggest problem is that AI allows a writer to weave a lot of these misapplied and misaligned ideas and forces it into a coherent sounding framework. On deeper thought (or running it through another AI to specifically look for problems) and you will find those errors. The biggest tell of this is a all encompassing framework. Those just do not really exist prior to the age of AI, simply because human writers have to take the time to error check, especially in publication in academic magazines with peer review.
All encompassing frameworks didnt exist prior to AI? Wut? Tell that to BCG or Deloitte or Bearing Point who have spent decades "over fitting" unvalidated frameworks and models to sector after sector.
As for what is the best analogy, subprime resi or commercial RE or infra loans I'd argue they all largely flawed, particularly when it comes to the GPU part of the build out. A GPU that's sat unboxed in a warehouse for 2-3 years waiting for a DC to be built to house it is in reality worth $0, even if hyperscalers are fiddling with amortisation schedules to stretch the out to 5yrs.
On the plus side, we'll probably enter a second Golden Age of animation and CGI as they'll probably make great render farms if you can pick them up for pennies on the dollar.
Lol. I like the sarcasm. I mean robust and statistically tested encompassing frameworks. Not for profit ones from consultants... look, now they can spin out even more of those with AI. I think that it is in aggregate inaccurate to say GPUs sit in warehouses for 2-3 years. That would say that H100 or A100s are sitting undeployed which is not true at this point. There is a lag certainly esp if you consider time to actually producing tokens but on average, it is likely sub 1 year.
Ed Zitron's research would beg to differ re Nvidia's claimed GPU sales and the GW of data centre capacity needed to run them. To whit, a massive mismatch, so a LOT of those GPUs are in fact sat in warehouses. And the DC build out is equally "behind schedule" at best or "complete fantasy" at worst. It's my understanding that there just is not a viable DC build pathway to get a lot of those units deployed and commissioned until *well* into their real-world amortisation, particularly given that Rubin is in theory due in 2H2026.
Hence why the hyperscalers have quietly stretched out the amortisation schedules on the GPU to 5yrs, which is frankly ridiculous accounting BS.
Feynman, the generation AFTER Rubin is due in 2028, yet Blackwell will still be on the books at 60% of face value in 2028, assuming straight line depreciation over 5yrs starting in 2025. It's just silly.
I would recommend not taking Ed Zitron's “research” as anywhere close to gospel. I find his stuff unsubstantiated to misleading at best. His goal is to build his AI bubble brand. He benefits just like (way smaller magnitude) Jensen benefits from hyping AI. I do find Jensen way more informed.
Whilst I agree Zitron's got a brand to build I don't think it's quite so easy to dismiss his core argument. He's not claiming any inside sources or special knowledge. Most of what's he's saying is drawn from public filings, press releases, analyst calls, annual reports etc.
The facts on the ground re the rate of data centre build out, power supply and commissioning are empiricallly verifiable. Ditto the accountancy games with capex amortisation schedules is there in black and white in the annual reports.
Whether you like his style of hype journalism or feel his case is over-blown is neither here nor there, really.
His core thesis when you distill it down is
AI needs GPUs
GPUs need Data Centres
Data Centres need land, water and power.
Currently there isn't enough of all 3, and the rate at which they are being made available is probably at least 1 order of magnitude too slow to match the financial engineering house of cards built on top of all the circular promises between the AI companies, hyperscalers, chip mfrs, and finance bros.
And to your very last point, citizens are beginning to resist against the building of new data centers. The state of New York just put a temporary halt on all large data center projects. Any assumption that the rate of building data centers will remain static is flawed. The public--as the author of this point notes--is the most fickle investor of them all and the public sentiment is turning against data centers due to their massive suck of public resources.
This is by far the best analysis I have read and great work on trying to discern the similarities and differences between AI capital cycles and past cycles (housing and dotcom). Uncle Sam seems to be coming much earlier here to rescue than it did with housing cycle because - hey AI is about national security. That said, not everyone will be bailed out / supported - the neoclouds are swimming naked. Their model collapses with second derivative as you point out.
1) The build rate of data centres is facing a growing resistance with delays and cancellations resulting. Also, something as simple as power transformers now have a lead time of four years. How will this impact the revenue forecasts of the whole chain?
2) The Iran war gives a real risk of stagflation. AI utilisation, or the appearance of utilisation, will be difficult to predict. Can the CAPEX justification survive even a small assumed contraction in the usage/market demand?
3) OpenAI is very vulnerable to a Chinese onslaught aimed at the volume market users. What is known about Chinese buildup of capability?
4) And let's add a black swan: a breakthrough in biological computers.
Interesting analysis. I’d say the AI bubble has elements of both the Dotcom bubble and the 2008 mortgage crisis. The analogy to the dotcom bubble is that the buildout of data centers is likely erroneously front running associated tech that would be needed to make AI provide ROI in the right time frame. That happened with the fiber buildout. There were not sufficient use cases yet developed and ancillary hardware for using the fiber capacity had yet to be developed or was too expensive. There were companies in that space (eg JDSU) that were in that space that were vaporized, at least in part because they pursued tech that turned out to be impractical. It’s quite possible that AI won’t be fully actualizable in a commercial sense using the current mega data center compute architecture. It’s may be too complex, expensive, energy intensive and depreciating to be a worthwhile solution for most applications. It may end up just an overgrown lab prototype. You’re already starting to see competition from open weight models running on local devices that may herald a paradigm shift in how AI is deployed for most applications. One difference from the dotcom bubble is that unlike fiber all those GPUs are fast depreciating special purpose assets. However some of the data center infrastructure may have discounted secondary uses.
Really good piece, and it would seem that OpenAI delaying their IPO, $META switching to leasing out compute capacity, like $SPCX did too, and $CRWV, and $ORCL retracing their gains from Mid-April, with $ORCL down 9 straight days, and $ORCL down 18 of the last 22 days, all indicative of an environment and backdrop that is changing, hidden in plain sight.
Thanks. Most people don’t know what a second derivative is. Think will end badly. A trillion a year on capx for AI 🤖 based on LLMs. Like Thelma and Louise driving off a cliff
I recall in 2008 that the biggest contributors to the problem were CDO bonds that were rated AAA which allowed institutional investors to invest in them and caused insurers to cover these bad positions. I’m not sure the AI credit is at this level of systemic risk.
When the Great Recession happened in 2008, to most people it came out of the blue. "No one could have seen this coming," said the cable TV channels. In reality, there were plenty of people calling it out in 2006-2008.
This is the best piece that I have seen explaining what to watch for to see it all happen again with AI.
I love Substack because incredible people take the time to share their insights, and there is an audience of people who read it. This is one of those pieces that is a great investment of your time and effort to read and to understand. Because, sooner or later, this will be a key factor in an economic recession.
Groundbreaker relies on first-semester calculus here.
The basic principle is this. Think of using your car. The distance you travel, your speed, and how fast you are breaking (or stomping on the gas) are three different things, right? But they are related. Speed measures how far you can go in an increment of time. High speed means that you are covering ground really fast; but it doesn't mean that you necessarily are covering a lot of ground (because that depends on how long you keep it up).
You can be going fast, but suddenly a stoplight changes color and you stomp on the brake. You were going fast, but only covered a short distance, right? So even though your speed was high, the distance was short. That is because your acceleration was great. What?? How can your acceleration be great when you slowed down? Because it was "negative" acceleration.
I had a typo as I wrote that paragraph to say that "suddenly a stoplight changes color and you stomp on the gas." In that case, the acceleration is significant and in the "positive" direction. You have positive acceleration. This increases your speed. And you keep going, so you increase your distance as well.
The article's point is to understand what this means for the AI industry. In the case of AI, distance correlates to the total number of data centers and compute. Speed corresponds to how fast data centers are being added. And acceleration corresponds to changes in how fast data centers are being added. When data centers are not being added as fast, that will upset the apple cart of how their financing works. And this is inevitable.
Groundbreaking work!
F**k f**king substack’s ios app. I merely glance at my unposted comment the wrong way and it slides off the page into oblivion.
Sigh…typing for the third time - a friendly counterpoint: I think ‘the Big Short’ has skewed memories of the Great Recession.
There was nothing sudden about it at all by my recollection. Subprime delinquencies dominated headlines starting in 2006. Lenders and hedge funds failed throughout 2007. Bear Stearns was on its death bed for a solid year mid 07 to mid 08z Anyone surprised by Lehman and Merrill bankruptcies in Sept 2008 was seriously not paying attention…
Most people don't pay attention, then or now. That's what scares me about AI, actually. People who don't think won't understand that it doesn't think, either. AI is Sean Hannity with Tucker Carlson's demeanor. (Ha. I like that one.)
Re no one could see this coming. That is ChatGPT’s version of the housing crash. Which just goes to show it is a ginormous internet word salad regurgitation machine.
And right on cue, OpenAI proposes a 5% stake by the federal government.
IIRC this was a free 5% stake offered to the government. Effectively a proposed bribe in return for a bailout.
Trump calls everybody "communists", but under Trump, the government is Intel's largest shareholder (and he strong-arms other companies to buy Intel products). The government is also a major shareholder of MP Materials and of US Steel and gets a cut of Nvidia's sales.
And they call the left “socialists”. Government bailouts for investors only, not the elderly, disabled, the ill, children or students.
Scary thought when Gov steps in...
Corporate Socialism in action.
Very insightful article. I especially like the second-derivative framework — the idea that markets often break not when growth stops, but when the acceleration of growth disappoints.
My only challenge with applying that framework too aggressively to AI capex is that it risks treating a multi-dimensional phenomenon as a single-variable equation.
Yes, AI infrastructure spending growth will eventually decelerate. Mathematically, it has to. No exponential curve keeps accelerating forever. But a slowdown in the rate of investment growth does not necessarily imply a slowdown in the surface area of impact.
A useful comparison is exactly the one you used - the subprime mortgage crisis. That was largely a one-dimensional leverage cycle: more credit → more housing demand → higher home prices → more collateral value → more credit. Once the underlying assumption (housing prices only go up) broke, the feedback loop reversed. The second derivative told you almost everything because the system depended heavily on one dominant variable: the continuous expansion of mortgage credit.
AI capex feels fundamentally different.
The current investment wave is not only chasing one asset class or one revenue stream; it is building a new computational layer. The second-order and third-order effects are spread across productivity, software, robotics, science, healthcare, energy, education, manufacturing, defense, and entirely new categories we probably have not imagined yet.
Ten years from now, there is a reasonable chance we look back at today’s “AI use cases” the same way we look at early internet websites — useful demonstrations, but nowhere close to the final economic architecture.
The risk is real: overinvestment, misallocated capital, and unrealistic timelines can absolutely create painful corrections. But the harder question is not simply “will AI capex acceleration slow?”
It is: “how much economic surface area gets created before and after that slowdown happens?”
The second derivative may capture the investment cycle.
It may not fully capture the innovation curve…
Are you saying "this time is different"? The point of the author's analysis is that this is, for all practical purposes, a credit bubble a la 2008. The homes built at that time eventually found long term owners. It was the transition from pre-2008 to post 2008 that was painful. AI use cases may indeed explode just as internet use cases did, but the precondition to that massive adoption maybe a dramatic drop in token pricing precipitated by the bubble bursting. It's the mechanics of a potential burst we are pondering here--and the entire focus of the analysis.
Yes, that could absolutely happen — and I agree the transition mechanics matter. A capex cycle can be directionally right in the long run and still create enormous pain if capital allocation gets ahead of near-term monetization.
My point is slightly different: the analysis feels incomplete if it only models the liability side of the equation and not the asset being created.
In 2008, the world discovered that a significant portion of the “asset” backing the credit expansion was mispriced housing demand. The houses remained useful, yes, but the marginal economic productivity unlocked by an extra subdivision was relatively limited.
AI infrastructure is a different question. The debate is whether we are overbuilding empty capacity or whether we are building a new productivity layer. If the second is true, then some portion of this capex has to eventually appear on the world’s balance sheet as productivity expansion, lower costs, faster innovation cycles, new businesses, and new economic output.
Almost nobody knows the magnitude or timing of that yet — and that uncertainty is my point.
If AI capex were purely a 2008-style credit bubble with no offsetting productivity asset, you could argue it should have broken already because the gap between investment and immediate cash flow is historically extreme. The fact that it has persisted this long is itself information: markets may be pricing something beyond the traditional bubble template.
The question is not “is this time different?”
The better question is: “what variable exists in this cycle that was missing in previous cycles?”
For AI, the answer may be the scale and speed of the productivity feedback loop. Whether that is large enough to justify the current capex spend is the trillion-dollar question…
thank you for the response. Appreciate the discussion.
Pleasure is all mine Krishna. I also do enjoy these debates…
MLPortfolios you have the right question. I want to add a third position to you and Krishna's debate, if I may
The bear watches the credit structure and says it breaks on deceleration. The bull watches the productivity feedback loop and says the missing variable justifies the capex. Both are right on their own terms.
But both assume a knowable revenue ceiling. The bear needs it to price credit risk. The bull needs it to justify the multiple. Neither asks whether the ceiling itself is the assumption that breaks.
AlphaFold had no knowable revenue ceiling before it ran. The discoveries it made possible were not in any model because they had not been made yet. If AGI-adjacent systems do that across multiple domains simultaneously, the feedback loop you describe cannot be sized using current instruments.
Krishna is right that the precondition for massive adoption may be the bubble bursting. The dotcom infrastructure survived. The financing around it did not. Survivors bought stranded assets for pennies.
Both can be true simultaneously. Credit structure breaks on deceleration while long-run productive value remains unbounded. That is the position the retail investor inside this through an index fund has no framework to navigate.
Thanks Sovereign. I really like this framing and find it quite illuminating as It deftly reconciles what initially looked like opposing views.
Where I’d extend your idea is this: perhaps the real bottleneck isn’t estimating the revenue ceiling, but measuring the rate at which the ceiling itself expands?
Historically, we’ve valued businesses on relatively stable addressable markets. AI may be changing the market while simultaneously participating in it. Every productivity gain increases humanity’s capacity to create the next productivity gain. That’s a recursive system, not a linear one.
I also agree that a bubble and a productivity revolution are not mutually exclusive. The dot-com analogy is instructive: the financing layer can implode while the infrastructure layer compounds for decades. Credit can be catastrophically wrong while technology is directionally right.
What fascinates me is the missing state variable.
Markets are exceptionally good at pricing cash flows. They’re much less equipped to price new possibility space (I am one of those who strongly believe that new technology should be priced more like options rather than using using traditional tools such as DCF) —economic activity that doesn’t yet exist because the tools to create it don’t yet exist.
If that’s the regime we’re entering, then both the bull and the bear are working with incomplete state vectors. The debate shifts from “Who is right?” to “What variables are our models still missing?”
That, to me, is the more interesting research problem.
MLPortfolios your options point is right. DCF requires a terminal state. Options pricing requires only a distribution of possible states and a timeframe. If the ceiling expands recursively, the terminal state is undefined by construction and DCF cannot reach it.
The missing state variable point is where I would go further. Markets are not just less equipped to price new possibility space. They are structurally incapable using current instruments, because those instruments require knowable cash flows in knowable timeframes. That is not a calibration problem. It is a category error.
Which brings it to the practical question for the investor sitting inside this right now. If both bull and bear are working with incomplete state vectors, and the market is pricing the asset using instruments that cannot reach the relevant variable, then conviction framework matters more than the valuation model.
You cannot know the ceiling. You can know what you believe about the recursive system, document it, and define the conditions under which that belief would be falsified. That is a different kind of analysis from DCF. It is the only one that works when the state vector is incomplete.
Alas, we converge. This has been my point all along.
If the state vector is incomplete, then your positioning in the AI capex trade inevitably becomes conviction-based rather than valuation-based.
But here’s the question almost no one asks: where does conviction come from?
My observation is that conviction is rarely created by reading. It is created by experience.
If you’re only consuming articles about AI, your conviction is largely borrowed from other people’s models. It will naturally be fragile because every new headline can change your mind.
If, however, you’re a daily power user building with AI, watching it fail, improve, surprise you, and reshape your own productivity, your conviction becomes grounded in first-hand evidence. It doesn’t make you right—it simply means your priors come from direct interaction rather than second-hand narratives.
Ironically, this is probably why the AI debate is so polarized. Many of the strongest bulls are heavy users. Many of the strongest bears evaluate AI primarily through financial statements, capex charts, and valuation multiples. Both are observing something real—but through different lenses.
The truth, as usual, is probably somewhere in between.
My advice to anyone navigating this trade is simple: first identify your bias. Bullish, bearish, or indifferent. Then actively seek evidence that could falsify your position. If you’re bullish, spend time with the best bearish arguments. If you’re bearish, spend time building real things with AI.
When we look back a decade from now, I suspect both bulls and bears will have been right—and wrong—in different ways.
The biggest winners may not be those who picked a side. They may be those who understood the asymmetry, managed their risk accordingly, and positioned themselves to benefit regardless of which part of the thesis unfolded first…
I see the innovation curve as a echo of the high-speed internet build out in the early 2000s - fiber optic cable was the game changer for high speed internet but the economics played out different.
I actually like this fibre analogy as it is probably the closest historical precedent.
The difference, though, I think, is that fiber primarily reduced the cost of moving information. AI has the potential to reduce the cost of producing intelligence itself and the distinction between both matters more than we think.
When bandwidth became abundant, we didn’t just browse websites faster, rather we got streaming, cloud computing, social media, mobile apps, and entire industries that barely existed before.
If intelligence follows a similar cost curve, today’s use cases may end up looking like email did in the early internet era i.e. important, but we maybe looking at only a tiny glimpse of what the infrastructure may ultimately enable.
Ironically, that doesn’t invalidate the economics either. The fiber build out was painful for many investors, but transformative for society. AI may follow a similar path where the infrastructure outlasts the financing cycle.
The hardest part, as investors, is separating the timing of returns from the permanence of the technology. History suggests those two are often very different.
Overbuild the houses and no one will use them, their marginal value drops to 0. Overbuild the AI compute layer and the marginal value of the chips, while still drops, doesn't go to 0. People will always find a way to use it. So I think while there's merits to compare with 2008, it's not complete analysis to not look at the demand part of the equation. Real compute demand, as you say, in different industries. Right now it's mostly consumers and coding, but there would be more overtime. So only looking at second derivatives is really over-simplifying.
Exactly. I think you’ve put your finger on the variable that’s easiest to overlook (and explained it so simply): demand isn’t static.
An oversupplied housing market and an oversupplied compute market don’t necessarily behave the same way.
An empty house creates very little new demand simply because it exists. Compute is different. As its marginal cost falls, entrepreneurs, researchers and businesses find entirely new ways to consume it. History is full of examples where cheaper computation didn’t just satisfy existing demand—it created demand that didn’t previously exist.
That doesn’t mean AI capex can’t overshoot. It absolutely can. Some projects will destroy capital, and some companies will almost certainly overbuild.
But I struggle with analyses that treat compute as if it’s just another cyclical asset. Compute is a general-purpose input into innovation itself. That changes the demand equation in ways our traditional bubble frameworks don’t fully capture.
To me, that’s the missing state variable. The question isn’t only “How much compute have we built?” It’s also “What new economic activity does cheaper, more abundant compute unlock?”
That’s a much harder variable to model—and probably the one that matters most…
That's very eloquently put. I think that's where bulls and bears differ. Bulls assume new high value use cases will spring up right on time to absorb extra compute supply. Bears treat total demand as deterministic, and once supply is overbuilt, the whole thing collapses. I think both are wrong, and we will be experiencing both compute under-supply (as of now), and compute oversupply (maybe in 2027), but then demand will catch up again because who doesn't need more intelligence? Right now it doesn't seem anyone has an answer yet.
Couldn’t agree more Hoang.
What strikes me is that everyone wants a static answer to what is fundamentally a dynamic problem.
Compute can be undersupplied in one regime, oversupplied in the next, and then undersupplied again as new applications emerge. Those states aren’t contradictory—they’re part of the same feedback loop.
Perhaps the real mistake is asking, “Will there be oversupply?” There almost certainly will.
The better question is, “How quickly does intelligence create the next wave of demand for intelligence?”
That’s the variable I don’t think we’ve measured well yet. And until we do, both the bull and bear cases will remain incomplete—which is exactly what makes this one of the most fascinating investment problems of our generation…
Great piece! Hopefully I can fully grasp this after rereading it 3 times. Maybe I’ll just ask AI to explain it to me. lol.
I had to open a lot of tabs to translate the jargon, but yeah--it was worth the effort.
It would help if Substack would offer a define option the way e-book readers do.
Did the same thing as you lol. Great piece
Chock full o' LLM gobbledygook. I couldn't make it past the 50% mark, all the "it's not this... it's that" phrases and techno-econ jargon were making me nauseous. Which is too bad because I thought it had some good ideas and views on how this AI financing vicious cycle is playing out. The article just needs some serious human touch and de-Dickens'ing.
I was looking for this comment. Agree!
Yeah glad someone else noticed. May well contain some truth but also surely contains bs. Ai slop, not written by a human author who understands things.
I just kept re-reading each paragraph until I could kinda make sense enough to go to the next paragraph, lol.
glad I’m not the only one 😅
In 2006, the only growing sector of the economy was homebuilding. There is no inherent value creation; we use already-created value to invest in a home. If no value is being created, all you can do is to borrow from the future. Entropy says that this can't extend forever.
There is some value being created by AI. What that value is will probably play an important role in how long this can be sustained.
Your analysis is spot-on.
Building a home that someone lives in creates value.
Relatively low value compounding, though. Building housing doesn't raise productivity much (which is ultimately what drives growth) and is the difference between housing and AI (or the 2000 boom).
Most "value creation" is terminal consumer goods rather than investment - about 3x from GDP figures.
Still, it's not exactly the case that housing is a pure consumer good. More housing in more productive cities in particular can create compounding returns as it allows workers to produce more, just like regular old technology or capital investments.
My point was that you cannot buy it unless you have created value elsewhere. It's from the excess of that value that you can buy it.
I think you are quite correct. About the vulnerability, the mechanism and the trigger (OpenAI). I do not know how to answer the timing question. And I need conviction on the timing to make this a tradable premise.
So I am watching carefully. Sentiment is shifting, has shifted. OpenAI is the canary.
Thanks for the essay.
Yeah. That is the big question. Are we in 2003 or 2005 or 2007?
OR, are we in 1990 or 1980 (or 1970!) of the Japan bubble, when P/E ratios topped out at around 99(!)
This article is why a reminder of everything I hate about AI and how users use it to write. There are a good number of points that stand strong in isolation in the article. However, they are misapplied without sufficient specific context. Overfitting is the term of use. The second derivative idea is very useful to spot blow off tops in markets but the limit is that it generally requires rapid feeback in the form of price discovery. The second derivative price applied to subprime is a single factor in what likely was a multi factor model (think no income and no doc mortgages) even if the housing bubble had enough volume for price discovery to offset the high transaction costs that usually kill these models. I would love to see the 2nd derivative idea backtested against other housing bubbles while statistically adjusting for other explanatory factors. And be sure to strictly apply the null hypothesis or whatever outcome you are testing for (is it price increases or rising deliquencies etc). Another problem is that AI DC and GPU backed loans are not subprime resi housing. Commercial RE lands closer. Infra loans is even better. Fracking (bc of the similar declines) loans for GPUs. LNG, midstream, and other LT loans for the DCs. The biggest problem is that AI allows a writer to weave a lot of these misapplied and misaligned ideas and forces it into a coherent sounding framework. On deeper thought (or running it through another AI to specifically look for problems) and you will find those errors. The biggest tell of this is a all encompassing framework. Those just do not really exist prior to the age of AI, simply because human writers have to take the time to error check, especially in publication in academic magazines with peer review.
All encompassing frameworks didnt exist prior to AI? Wut? Tell that to BCG or Deloitte or Bearing Point who have spent decades "over fitting" unvalidated frameworks and models to sector after sector.
As for what is the best analogy, subprime resi or commercial RE or infra loans I'd argue they all largely flawed, particularly when it comes to the GPU part of the build out. A GPU that's sat unboxed in a warehouse for 2-3 years waiting for a DC to be built to house it is in reality worth $0, even if hyperscalers are fiddling with amortisation schedules to stretch the out to 5yrs.
On the plus side, we'll probably enter a second Golden Age of animation and CGI as they'll probably make great render farms if you can pick them up for pennies on the dollar.
Lol. I like the sarcasm. I mean robust and statistically tested encompassing frameworks. Not for profit ones from consultants... look, now they can spin out even more of those with AI. I think that it is in aggregate inaccurate to say GPUs sit in warehouses for 2-3 years. That would say that H100 or A100s are sitting undeployed which is not true at this point. There is a lag certainly esp if you consider time to actually producing tokens but on average, it is likely sub 1 year.
Ed Zitron's research would beg to differ re Nvidia's claimed GPU sales and the GW of data centre capacity needed to run them. To whit, a massive mismatch, so a LOT of those GPUs are in fact sat in warehouses. And the DC build out is equally "behind schedule" at best or "complete fantasy" at worst. It's my understanding that there just is not a viable DC build pathway to get a lot of those units deployed and commissioned until *well* into their real-world amortisation, particularly given that Rubin is in theory due in 2H2026.
Hence why the hyperscalers have quietly stretched out the amortisation schedules on the GPU to 5yrs, which is frankly ridiculous accounting BS.
Feynman, the generation AFTER Rubin is due in 2028, yet Blackwell will still be on the books at 60% of face value in 2028, assuming straight line depreciation over 5yrs starting in 2025. It's just silly.
I would recommend not taking Ed Zitron's “research” as anywhere close to gospel. I find his stuff unsubstantiated to misleading at best. His goal is to build his AI bubble brand. He benefits just like (way smaller magnitude) Jensen benefits from hyping AI. I do find Jensen way more informed.
Whilst I agree Zitron's got a brand to build I don't think it's quite so easy to dismiss his core argument. He's not claiming any inside sources or special knowledge. Most of what's he's saying is drawn from public filings, press releases, analyst calls, annual reports etc.
The facts on the ground re the rate of data centre build out, power supply and commissioning are empiricallly verifiable. Ditto the accountancy games with capex amortisation schedules is there in black and white in the annual reports.
Whether you like his style of hype journalism or feel his case is over-blown is neither here nor there, really.
His core thesis when you distill it down is
AI needs GPUs
GPUs need Data Centres
Data Centres need land, water and power.
Currently there isn't enough of all 3, and the rate at which they are being made available is probably at least 1 order of magnitude too slow to match the financial engineering house of cards built on top of all the circular promises between the AI companies, hyperscalers, chip mfrs, and finance bros.
And to your very last point, citizens are beginning to resist against the building of new data centers. The state of New York just put a temporary halt on all large data center projects. Any assumption that the rate of building data centers will remain static is flawed. The public--as the author of this point notes--is the most fickle investor of them all and the public sentiment is turning against data centers due to their massive suck of public resources.
This is by far the best analysis I have read and great work on trying to discern the similarities and differences between AI capital cycles and past cycles (housing and dotcom). Uncle Sam seems to be coming much earlier here to rescue than it did with housing cycle because - hey AI is about national security. That said, not everyone will be bailed out / supported - the neoclouds are swimming naked. Their model collapses with second derivative as you point out.
Another brilliant one. I thought "Peak Cheap" was real good and you just raised the bar higher. Thank you, I am learning so much 🙏
Three small questions:
1) The build rate of data centres is facing a growing resistance with delays and cancellations resulting. Also, something as simple as power transformers now have a lead time of four years. How will this impact the revenue forecasts of the whole chain?
2) The Iran war gives a real risk of stagflation. AI utilisation, or the appearance of utilisation, will be difficult to predict. Can the CAPEX justification survive even a small assumed contraction in the usage/market demand?
3) OpenAI is very vulnerable to a Chinese onslaught aimed at the volume market users. What is known about Chinese buildup of capability?
4) And let's add a black swan: a breakthrough in biological computers.
Interesting analysis. I’d say the AI bubble has elements of both the Dotcom bubble and the 2008 mortgage crisis. The analogy to the dotcom bubble is that the buildout of data centers is likely erroneously front running associated tech that would be needed to make AI provide ROI in the right time frame. That happened with the fiber buildout. There were not sufficient use cases yet developed and ancillary hardware for using the fiber capacity had yet to be developed or was too expensive. There were companies in that space (eg JDSU) that were in that space that were vaporized, at least in part because they pursued tech that turned out to be impractical. It’s quite possible that AI won’t be fully actualizable in a commercial sense using the current mega data center compute architecture. It’s may be too complex, expensive, energy intensive and depreciating to be a worthwhile solution for most applications. It may end up just an overgrown lab prototype. You’re already starting to see competition from open weight models running on local devices that may herald a paradigm shift in how AI is deployed for most applications. One difference from the dotcom bubble is that unlike fiber all those GPUs are fast depreciating special purpose assets. However some of the data center infrastructure may have discounted secondary uses.
The second derivative of velocity is called "jerk"... And a lot of AI investors will be learn that soon enough I suspect!
Great analysis, thank you.
Really good piece, and it would seem that OpenAI delaying their IPO, $META switching to leasing out compute capacity, like $SPCX did too, and $CRWV, and $ORCL retracing their gains from Mid-April, with $ORCL down 9 straight days, and $ORCL down 18 of the last 22 days, all indicative of an environment and backdrop that is changing, hidden in plain sight.
Thanks. Most people don’t know what a second derivative is. Think will end badly. A trillion a year on capx for AI 🤖 based on LLMs. Like Thelma and Louise driving off a cliff
Great essay!
I recall in 2008 that the biggest contributors to the problem were CDO bonds that were rated AAA which allowed institutional investors to invest in them and caused insurers to cover these bad positions. I’m not sure the AI credit is at this level of systemic risk.
But people have been treating equity like it’s credit since 2020