The following piece represents my opinion only. Please read my full disclaimer below.
It’s tough timing bubbles bursting. Just ask any short seller or Austrian economist, including myself, who is constantly ridiculed for being a “broken clock” over and over until, of course, their thesis plays out…just never on the timetable that made it comfortable to hold. Today I’ll do what I rarely do, which is make a measured guess about the AI bubble and timing, hereinafter referred to as hanging my balls out there.
After years of euphoria, soaring valuations and trillions in increasingly circular investment, I think the end of 2026 and beginning of 2027 might finally mark the moment the AI boom begins to live up to next season’s name: the fall. I have five key reasons I believe this, which I will discuss below.
But first, let’s set the table. The most dangerous sentence in markets today is that artificial intelligence is real. It is real, of course, and it will almost certainly reshape enormous parts of the economy, but that fact has become the de facto answer by every lobotomized automaton that appears on CNBC daily to a completely different question: whether the trillions of dollars now being spent, borrowed, committed and capitalized around AI can possibly earn an adequate return.
It’s a total non-sequitur. Most people know the basics of the “overshooting the mark on the buildout” thesis. I highlighted both Michael Burry and Jim Chanos’ excellent thoughts on this. Railroads changed America and destroyed fortunes. Fiber optic cable became the backbone of the internet after helping bankrupt the companies that laid too much of it. The internet itself exceeded almost every grand prediction made about its importance in 1999, while the Nasdaq still managed to fall roughly 80 percent.
So someone inform the next brilliant Series 63 holder on CNBC that transformative technology and catastrophic overinvestment have always been perfectly capable of occupying the same room.
That is increasingly the right framework for understanding AI right now. The technology is extraordinary, but the capital cycle built around it has gone much further than the economics can support, in my opinion. Investors have taken perhaps the most consequential technological development since the internet and done what investors reliably do with consequential technological developments: extrapolated the eventual destination backward into today’s valuation, financed the intervening decade (or even more) in advance, and then congratulated themselves on the resulting growth.
In other words…we are no longer merely betting that AI becomes enormous. We are betting that it becomes enormous quickly enough to justify an infrastructure buildout measured in trillions of dollars, while maintaining attractive returns on capital despite rapidly falling inference costs, intense competition and technological obsolescence that can make expensive hardware look old remarkably quickly.
The timing could hardly be worse. The AI capital boom is reaching its most aggressive phase after years of positive real interest rates have worked their way through the rest of the economy, as I have consistently written about.
Not only that, but bonds both in the U.S. and Japan are *screaming *that risk is underpriced. Bond yields are scorching higher on both sides of the Pacific. In the last 24 hours, Japan’s 10-year JGB touched 2.93%, its highest since 1996, while the U.S. 30-year Treasury recently hit 5.27%, a level not seen since 2007.
The simultaneous surge in long-term borrowing costs is striking given Japan’s weak growth backdrop, underscoring how inflation, currencies and fiscal concerns are increasingly driving global bond markets.
Monetary tightening and higher yields do not hit everything at once. Households spend savings before they miss payments, companies refinance gradually, commercial property owners extend loans before recognizing losses, and private equity firms possess the wonderfully convenient ability to discover that assets without daily market prices are much less volatile than assets with them.
The damage accumulates quietly…before appearing suddenly.
American households are now carrying roughly $18.8 trillion of debt as of the end of June 2026, up about $400 billion, or 2.2%, from $18.39 trillion a year earlier and at or near an all-time high.
Credit card balances stood at $1.26 trillion as of the end of June 2026, up roughly $50 billion, or 4%, year over year and near record highs, while auto loan balances reached roughly $1.71 trillion as of the end of June 2026, up about $55 billion, or 3%, from a year earlier and at an all-time high.
Delinquencies remain elevated. As of Q2 2026, 4.7% of all outstanding household debt was in some stage of delinquency. More importantly, credit-card and auto-loan serious delinquency rates remain elevated relative to their pre-pandemic norms.
None of this means the consumer has already collapsed, but households have nevertheless spent several years absorbing substantially higher financing costs while carrying record or near-record nominal debt loads. At the same time, rising stocks, home values and retirement accounts have helped support household balance sheets and consumption. That combination becomes considerably less comfortable if expensive credit persists while asset prices begin falling.
And as I’ve constantly reminded…the bigger the bubble, the more leverage people take on, the smaller the move has to be in the markets for everyone’s “cushion” (i.e. illusion of wealth) to simply go “POOF” into thin air.
Like I’ve reminded constantly, behind the consumer sits an even larger collection of assets whose prices have not really been tested. God only knows what is going on over the crypto/stablecoin world (my latest thoughts on that opacity here). Meanwhile, private equity marks are not necessarily fraudulent, nor is every commercial property secretly worthless, but neither market has experienced the sort of continuous clearing mechanism imposed on public equities.
A private company can remain valued at the price of its last financing round long after comparable public companies have fallen. A commercial building can remain marked to an appraisal until refinancing or a forced sale introduces the owner to someone willing to write an actual check. Private credit can appear remarkably stable when lenders amend, extend and avoid crystallizing losses. There is nothing mysterious about this. In fact we’re watching sociopaths literally repeat 2008 history while trying to figure out a way to move dogshit private credit assets that currently have near zero liquidity for a damn good reason. Assets without frequent transactions naturally adjust more slowly. The problem comes when investors mistake slow (or no) price discovery for economic stability.
This is why the experience of Michael Burry before the financial crisis is more relevant than the caricature of it. Burry did not discover the housing problem in October 2008 and short it the week before Lehman collapsed. He spent an extended period watching the underlying mortgage data deteriorate while the securities supposedly representing those mortgages refused to reflect reality.
His investors became furious, counterparties continued marking instruments at prices he believed were absurd, and for a meaningful period the market’s verdict was that Burry was wrong. Eventually prices caught up and Burry was proven right. That does not make Burry automatically correct about AI today, but it should permanently retire the argument that a bearish thesis is disproven because the market has continued rising for another 6 or 12 months.
The more troubling feature of the present boom is that AI has developed a financing structure in which the success of one participant increasingly helps finance the success of another. AI companies raise capital at enormous valuations and use that capital to buy compute. Compute purchases become revenue for semiconductor companies and commitments for data center operators. Those revenues and commitments support higher valuations, which make financing easier, which provides more capital to customers, which produces more orders, which supports still higher valuations.
Hyperscalers sign enormous long term capacity agreements because they expect AI demand to compound, infrastructure companies finance construction against those commitments, and investors treat the resulting construction boom as further evidence that AI demand must be enormous.
Think of a group of people sitting around a poker table using the same stack of chips to prove that everyone at the table is rich. As long as the chips keep circulating, everyone looks wealthy and every transaction seems to prove the chips are valuable. Nobody needs to be lying or doing anything irrational. The same money just keeps moving around the table and supporting higher bets. *Everyone wants these chips, and there’s plenty to go around. *
But eventually the rake takes some chips, a couple big players walk away and the money stops circulating. Suddenly everyone realizes there were never as many chips at the table as they thought. Those left without chips who were hoping to “win” their chips back are shit out of luck, as my mother would say.
This is where companies such as Nvidia become more complicated than a simple discussion of earnings multiples.
Nvidia may be the most important cog in AI’s circular financing machine because it increasingly sits on both sides of the transaction. It intends to invest up to $100 billion in OpenAI as OpenAI builds at least 10 gigawatts of Nvidia-powered infrastructure; it invested $2 billion in CoreWeave, one of its major GPU customers, while also agreeing to buy up to $6.3 billion of CoreWeave capacity that goes unsold.
And Nvidia is now working with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on platforms targeting more than $500 billion of AI-infrastructure financing, with Nvidia retaining the option to backstop up to $125 billion, according to Reuters.
The circularity is simple: Nvidia helps finance the companies buying Nvidia-powered compute, those purchases become Nvidia revenue, and that revenue gives Nvidia more firepower to finance the next round. Nothing about that is inherently improper, but it makes Nvidia not just AI’s dominant chip supplier, but one of the financial engines keeping the entire boom moving. **If Nvidia were to…say…****ever ****miss earnings, it’s not hyperbolic to say it could put a dent in the entire global economy overnight. US markets would get hit, as would overseas markets. Sentiment across millions of trading desks would change instantly. **
That reflexivity works wonderfully on the way up and becomes vicious on the way down. If AI customers disappoint, their valuations fall and their cost of capital rises. If their cost of capital rises, they buy less compute. If compute purchases slow, expected semiconductor earnings decline at exactly the moment investors are compressing the multiple they are willing to pay for those earnings. Falling semiconductor valuations weaken confidence throughout the ecosystem, financing terms tighten for data center developers, weaker developers cut orders, lenders become more selective, private capital marks fall and suddenly the demand forecasts that justified the original infrastructure buildout are being revised because the financing that created part of that demand has disappeared.
The size of the commitments makes this especially dangerous. Goldman Sachs has predicted that hyperscaler AI capex could reach as much as $1.4 trillion in 2027, while Morgan Stanley estimates the same companies and Nvidia already have roughly $1.8 trillion of off balance sheet commitments, including about $982 billion in purchase obligations and $822 billion in leases that have not yet commenced, creating a massive pool of future cash obligations that could become a serious problem if AI revenues fail to materialize fast enough.
Not all of this is debt in the accounting sense, and calling every dollar of it hidden borrowing would be sloppy. Economically, however, these are claims on future cash flows, and the distinction between a contractual commitment, a lease obligation and funded debt becomes rather less comforting when the revenues expected to service all three are being revised downward.
During the boom, investors look at these obligations and see visibility. A ten year data center commitment proves future demand, a massive equipment order demonstrates confidence, and a long term lease signals strategic necessity. During a bust, precisely the same spreadsheet receives different labels. Visibility becomes rigidity, commitments become liabilities, capacity becomes overcapacity and strategic investment becomes sunk cost. Nothing about the underlying contract needs to change. Only the denominator in the expected return calculation does.
The larger danger is that an AI correction would not occur inside a sealed technology sector. The financial system is currently carrying trillions of dollars of assets whose apparent values depend to some degree on abundant liquidity and the absence of forced selling.
Private equity portfolios contain companies marked from financing rounds and valuation models rather than continuous transactions. Commercial real estate still contains properties financed under a radically different interest rate regime. Venture portfolios contain companies whose most recent marks assume future exits into receptive public markets. Crypto contains enormous quantities of nominal wealth that can disappear without anybody technically defaulting on anything. Public equities themselves embed assumptions about margins, growth and discount rates that have been extraordinarily generous to the winners of the AI trade.
The first meaningful break in AI could therefore trigger something much larger than a semiconductor selloff. A meaningful decline in the AI complex would reduce household and institutional wealth, create margin pressure and force investors to raise cash. The assets they can sell first are public equities, so those fall further while private marks initially remain unchanged. Institutions then discover that their allocation to private markets has mechanically increased because the public side of the portfolio declined, forcing secondary sales and creating pressure on private valuations. Lenders tighten standards, credit spreads widen, refinancing becomes more difficult and commercial property owners approaching maturities discover that an appraisal produced during better conditions is not the same thing as a bid.
Crypto would probably contribute its traditional public service of discovering the clearing price at three in the morning on a Sunday. Tokens that appeared to represent hundreds of billions of dollars of wealth would reveal that market capitalization is calculated using the price of the marginal transaction, not the price at which every holder can exit simultaneously. Private equity would experience the same lesson more slowly and with better PowerPoint presentations. Commercial real estate would experience it when loans mature. Venture capital would experience it…
