The Line That Could Pop the AI Bubble
Ray Dalio said something in Singapore last week that stopped me cold.
“Everybody says ‘I’m worth a billion dollars’ but OK, try to spend that,” the Bridgewater founder told the Forbes Global CEO Conference. “In order to spend that you have to sell wealth in order to get money, and so the bubble usually pricks at that”.
He wasn’t talking about some abstract economic theory. He was describing the exact moment a bubble stops being a story about paper gains and becomes a story about cash. And right now, the AI boom is sprinting toward that moment with a borrowed credit card in one hand and a spreadsheet in the other.
Here’s the thing nobody wants to say out loud: the line that pops this bubble isn’t a technology failure. It’s a financing failure. A simple, boring, brutal math problem. And the numbers are getting ugly.
When capex eats your entire paycheck
Let me put this in terms you can feel in your gut.
UBS crunched the numbers and found that hyperscaler capital expenditures in 2026 are on pace to consume close to 100% of operating cash flows. The ten-year average? Forty percent.
Read that again. These companies are spending every dollar they generate just to keep building. Not most of it. Not eighty percent. All of it. And then some, because the actual capex figures are climbing faster than the cash can keep up.
Fitch projects that capital expenditure by Alphabet, Amazon, Meta and Microsoft alone will jump more than 75% this year to $700 billion. JPMorgan’s Jamie Dimon has put the broader hyperscaler ecosystem spending at around $700 billion for this year, with whispers of $1 trillion in 2026.
Where does the extra money come from when you’ve already spent everything you earn?
You borrow it.
And that’s where the line starts to reveal itself.
The $200 billion question
In the first six months of 2026, five companies, Amazon, Microsoft, Alphabet, Meta and Oracle, sold approximately $200 billion worth of investment-grade bonds. That’s nearly double what those same five companies issued across all of 2025.
Global AI-linked debt issuance is on track to hit $570 billion this year. Morgan Stanley says that’s more than double last year’s total.
The Bank for International Settlements put it plainly in a July report: the AI build-out’s “scale, reliance on debt and circular equity ties raise questions about the boom’s sustainability and financial stability”. The BIS economist who wrote the study, Phurichai Rungcharoenkitkul, added a line that should be printed on every investor’s forehead: “The larger the boom, the deeper the eventual bust. The race to commit early through debt and circular financing also makes a bust more likely”.
I keep coming back to that phrase: the race to commit early.
Nobody wants to be the company that underinvested. So everybody borrows. And when everybody borrows at the same time, the cost of that borrowing starts to matter in ways that spreadsheet models from eighteen months ago never anticipated.
Interest rates: the quiet assassin
Ray Dalio didn’t mince words about the mechanism. “There’s a lot of pressure from all the supply for much more rises in interest rates, and then you have the dynamic of when it starts to have an effect, and that is the point at which the bubble begins to burst,” he said. “We’re in that part of the cycle that is before that, but approaching that, I think we’re close to that”.
Here’s the arithmetic that keeps me up at night. If you’re a hyperscaler that just issued $50 billion in bonds at 4.5%, and rates climb to 6%, your interest expense jumps by roughly $750 million a year. Not catastrophic on its own. But multiply that across the entire AI supply chain, data center operators, chip manufacturers, cloud providers, the private credit funds that financed half of this buildout, and the cascade becomes real.
The BIS warned in June that a “risk reassessment triggered by rising bond yields or an AI bubble collapse could cause a corporate credit freeze with broader implications for total investment”. That’s central-bank language for the money stops flowing.
And when the money stops flowing, projects get shelved. When projects get shelved, revenue forecasts get slashed. When revenue forecasts get slashed, equity prices fall. When equity prices fall, the collateral backing all that debt weakens. And when collateral weakens, lenders pull back.
That’s the chain reaction. It doesn’t start with a bad product. It starts with a line on a bond trader’s screen.
Fitch’s warning shot
Fitch Ratings doesn’t usually do drama. So when the agency published its third-quarter Global Risk Outlook and named “mounting vulnerability to an AI-related market correction” as one of the two dominant short-term credit risks facing global markets, I paid attention.
The report noted that the S&P 500’s cyclically adjusted price-to-earnings ratio has climbed to levels close to those seen during the late-1990s dotcom boom. U.S. corporate bond issuance surged 26% in the first half of 2026, driven largely by AI-related fundraising.
Fitch stopped short of predicting a crash. But it flagged something more interesting: the structural interdependence. “The extent to which capital markets and economies have become intertwined with AI have created a vulnerability for credit,” the agency wrote.
Translation: a tech correction won’t stay in tech. It will bleed into credit markets, corporate lending, consumer spending, and eventually the real economy. Because AI has become too big to fail before it even finished being built.
The circular financing problem nobody wants to discuss
Here’s a detail that doesn’t get enough attention.
A significant chunk of the AI investment boom is being financed through circular structures. Company A invests in Company B. Company B buys services from Company A. Company A books the revenue and uses it to justify more investment. Company B does the same. The circle spins, and on paper, everything looks fantastic.
Franklin Templeton CEO Jenny Johnson, speaking on the same panel as Dalio, put it bluntly: “These capital-light businesses are now capital heavy; they’re now using creative financing and structuring and doing off-balance sheet (financing)… You really have to build out the web of liabilities to understand it”.
That web of liabilities is the real story. It’s not just the bonds you can see on the balance sheet. It’s the special purpose vehicles, the joint ventures, the vendor financing arrangements, the private credit funds that lent against future revenue streams that haven’t materialized yet.
The BIS warned that “increasingly complex off-balance sheet vehicles and circular investment structures” likely make current investment figures understated and more fragile than they appear.
Understated. Let that sink in. The $850 billion in global AI investment that Macquarie tracked in 2026 might actually be higher. And the fragility might be worse.
The revenue side of the equation
I’m not saying AI is a mirage. Far from it.
Macquarie’s analyst Viktor Shvets noted that annualized AI revenues are already estimated at close to $175 billion, enough to cover current operating expenses and depreciation, and growing roughly three times faster than previous IT waves. Microsoft’s AI revenue for fiscal year 2026 hit $34.33 billion.
These are real numbers. Real businesses. Real customers paying real money for real services.
But here’s the gap that matters: $175 billion in revenue against $700 billion in annual capex. The revenue is growing fast. The spending is growing faster.
The question isn’t whether AI will eventually generate enough revenue to justify the investment. The question is whether the financing structure can survive long enough to get there. And that’s a liquidity question, not a technology question.
Chris Wood at Jefferies put it starkly: “AI capex will destroy capital, could trigger major credit event”. He’s not saying AI is useless. He’s saying the financing math doesn’t close at current spending levels.
What the “rolling bubble” theory misses
Macquarie has a thesis that the AI boom won’t pop all at once. Instead, it will deflate through a series of “rolling bubbles”, different segments of the ecosystem heating up and cooling down at different times.
It’s a comforting thought. Gradual adjustments. Sector-specific corrections. No single catastrophic event.
But here’s my problem with that theory: debt doesn’t roll. Debt has maturity dates. Interest payments. Covenants. When a private credit fund that lent to a data center developer needs its money back, it doesn’t care that the chip sector is still doing fine. It wants cash. And if the data center developer doesn’t have it, the whole chain feels the squeeze.
The BIS study compared the current AI boom to canal mania in the 1830s, railway building in England in the 1840s, and the dotcom boom of the 1990s. All of those “ended in sharp corrections” that had far-reaching economic implications. The BIS added that the scale and speed of AI investment suggests the impact this time could be larger.
Larger.
The specific line
So what is the line? The specific number, rate, or event that triggers the cascade?
It’s not one thing. It’s a combination, and that’s what makes it so dangerous.
The first line is the bond market. If 10-year Treasury yields climb meaningfully above current levels and stay there, the cost of refinancing all that AI debt becomes prohibitive. Bank of America has reportedly told clients that bond yields would need to climb considerably higher before posing a genuine threat to the AI trade. That’s a reassuring sentence until you realize that “considerably higher” isn’t that far from where we are.
The second line is the cash conversion problem. Dalio’s point about paper wealth is critical. Right now, the AI boom is supported by equity valuations that exist largely on screens. Nvidia, Apple, and Microsoft alone make up over 21% of the entire S&P 500. If institutional investors decide they want to rotate out of AI and into something else, anything else, the selling pressure alone could trigger a cascade.
The third line is the credit market’s tolerance for opacity. The circular financing structures, the off-balance sheet vehicles, the private credit exposure, none of this is fully transparent. When lenders start asking questions they can’t easily answer, they stop lending. And when they stop lending, the whole thing seizes up.
What happens if the line gets crossed
I want to be clear about something: I don’t think this ends with AI disappearing. The technology is too useful, too embedded, too transformative for that.
But I do think it could end with a brutal repricing of every AI-adjacent asset on the planet.
Fitch warned that a correction “could trigger a potentially significant and prolonged market correction, with widespread macroeconomic implications”. The BIS warned of a “corporate credit freeze.” Dalio warned that “we’re close to that” moment.
When I talk to people in finance about this, the conversation usually splits into two camps. The first camp says the AI boom is different because the revenue is real. The second camp says it doesn’t matter how real the revenue is if the financing structure collapses before the revenue scales.
Both camps are right. And that’s the problem.
The uncomfortable math of scaling
Let me try to make this concrete.
Suppose you’re a hyperscaler that’s spending $100 billion a year on AI infrastructure. Your operating cash flow is $90 billion. You’re borrowing $10 billion a year to cover the gap. Fine. Manageable.
Now suppose your capex needs to climb to $150 billion because your competitors are spending more and you can’t afford to fall behind. Your operating cash flow grows to $100 billion. Now you’re borrowing $50 billion a year. The debt accumulates. The interest payments compound.
Multiply this across five or six major players, plus dozens of smaller companies in the supply chain, plus the private credit funds that financed the data centers, plus the utilities building power plants to feed the compute demand. The aggregate borrowing requirement balloons.
And here’s the kicker: the revenue doesn’t scale linearly with the spending. It scales with adoption. With customer willingness to pay. With regulatory frameworks that don’t exist yet. With labor market disruptions that are already showing signs of strain.
Macquarie noted that the real pressure points are emerging in labor markets, where lower hiring rates and declining education premia point to early evidence of AI-related disruption that isn’t fully captured in official statistics.
That disruption affects the demand side of the equation. If people lose jobs, they spend less. If they spend less, the companies that buy AI services earn less. If those companies earn less, they cut back on AI spending. And the circle closes.
What I keep coming back to
Ray Dalio’s point about money versus wealth.
Money is a claim on goods and services. Wealth is the capacity to produce. The AI boom has created enormous money wealth on paper. Trillions of dollars in market capitalization. But the real wealth, the productive capacity, the actual goods and services being produced, is still catching up.
When the gap between money wealth and real wealth gets too wide, the correction is inevitable. The only question is whether it’s orderly or chaotic.
Fitch said the exposure of the economy and capital markets to a correction is “significant”. The BIS said the potential impact could be larger than previous technology busts. Dalio said we’re approaching the point where it starts to have an effect.
I don’t know exactly where the line is. Nobody does. But I know it’s closer than the stock market’s record highs would suggest. And I know that when it gets crossed, the popping sound won’t come from a research lab or a product launch.
It’ll come from a bond desk.
The line between a boom and a bust isn’t drawn on a chart. It’s drawn in the willingness of lenders to keep lending. And that willingness has a limit. The AI industry is about to find out where that limit is.