The AI Trade Just Became a Credit Trade

The Setup

There is a sentence that credit investors repeated to each other for the better part of two years, and it went something like this: do not worry about the AI capital spending, because it is funded by cash flow, so it is an equity risk, not a credit risk. The logic was sound at the time. Companies like Microsoft, Google, Amazon, and Meta generate cash on a scale that dwarfs most national economies, and they were pouring it into data centers without borrowing a dollar. If the AI bet disappointed, their stock prices would fall, but no lender was exposed. It was somebody else's problem.

That sentence has quietly stopped being true, and the implications have not caught up to the conversation. The spending has grown so fast that it now exceeds even these companies' prodigious cash generation, and the gap is being filled with borrowed money. The most careful way to see it is in a single ratio: capital expenditure as a share of operating cash flow. Historically, the largest technology companies spent something like 40% of their operating cash flow on capex. In 2026, that figure is pushing toward 100%. When a company spends essentially all the cash it generates on building one thing, anything else it wants to do, including continuing to build, has to be financed.

So this piece is about a migration. The AI trade has been described endlessly as a story about chips, models, and demand. Underneath it, a different story has been forming, about bonds, leases, and credit spreads. The risk that used to sit with equity holders, who signed up for volatility, is moving onto the balance sheets of bondholders, who did not. And because the sums involved are large enough to reshape the corporate bond market itself, this is no longer only a question about a handful of technology stocks. It is a question about credit.

The Contract That Broke

The scale of the shift is easiest to grasp through the bond deals themselves, because they arrived fast and large. In 2025, the five biggest AI infrastructure spenders issued roughly $121 billion in bonds, more than four times their combined annual average over the prior five years. The back half of the year produced a run of record-setting transactions: Oracle sold $18 billion in September, Meta followed in October with a $30 billion offering that was the largest investment-grade corporate bond sale ever completed outside of a merger, and Alphabet and Amazon added $17.5 billion and $15 billion deals in November. Four of the five biggest high-grade bond deals in the United States that year came from this group.

That was the opening act. Analysts at Morgan Stanley and JP Morgan estimate the technology sector will need to issue roughly $1.5 trillion in new debt over the next few years to fund the AI buildout. Barclays projects total United States corporate bond issuance will reach about $2.46 trillion in 2026, and identifies AI capital spending as the single largest force pushing that number higher. To put the concentration in perspective, the group of hyperscalers is on track to rival the largest banks as a source of investment-grade bond supply. A sector that was famous for sitting on mountains of net cash is turning into one of the biggest borrowers in the market.

One credit manager captured the change well: for years investors were told this spending would be funded by internal cash flow, that it was speculative equity risk and nothing for a lender to worry about. That understanding, he said, has been broken. The spending did not slow to fit the cash flow. The financing changed to fit the spending. And a promise about who bears the risk, made implicitly to every bondholder who owns investment-grade technology paper, was quietly rewritten.

Where the Stretch Shows

Here is the crucial distinction, and it is the reason this is not a simple bubble call. The largest hyperscalers can carry this debt. Microsoft, Google, Amazon, and Meta have real cash flows, top credit ratings, and balance sheets strong enough that even large bond programs sit comfortably against their earnings. If the story ended with them, there would be little to say. It does not end with them. The stretch shows up at the edges, and the edges are where credit accidents usually begin.

Oracle is the clearest pressure point. It is spending like a hyperscaler without the same fortress underneath it. Its capital spending reached roughly $56 billion in fiscal 2026 and could climb toward $95 billion in fiscal 2027, and it has told investors it plans to raise close to $40 billion more through a mix of debt and equity. Layered on top are more than $248 billion in data center lease commitments that have not yet even commenced, against roughly $124 billion of existing borrowings. Some of its notes stretch to maturities in 2065. Analysts at Barclays have reportedly warned the company could face a genuine cash squeeze, and the credit default swaps on its bonds, the instruments that pay off if it cannot repay, have had sharp bouts of volatility. This is what a company financing an ambition larger than its balance sheet looks like.

Below the named giants sits a more fragile layer still. A class of specialized providers, the so-called neoclouds, has been buying enormous quantities of chips using debt collateralized by the chips themselves. One of the largest funds GPU purchases through facilities that carry interest rates near 14%, the kind of rate that signals the market is charging real risk. And a growing share of the entire buildout is being financed off balance sheet entirely, through leases and private infrastructure partnerships that do not show up cleanly in the headline debt figures. On-balance-sheet hyperscaler debt is around $420 billion, but as one chief investment officer put it, the larger commitments live in the leases. The visible debt is not the whole iceberg.

The Money That Goes in a Circle

Now the part that should make any careful investor pause. A large and growing share of the money financing the AI boom is moving in a circle, and the circle has a way of making demand look more organic than it is.

The center of the circle is Nvidia. Its vendor financing, the practice of investing in the very customers who then buy its chips, now totals around $110 billion in commitments, plus billions more in debt backed by GPUs. The marquee deal is a commitment of up to $100 billion to OpenAI, structured in tranches tied to how much Nvidia hardware gets deployed. OpenAI's own chief financial officer described the arrangement plainly, acknowledging that most of the money will flow back to Nvidia. The chipmaker holds similar positions across the ecosystem, including a stake in the neocloud CoreWeave, an investment in Elon Musk's xAI, and deals with a string of smaller cloud providers that rent Nvidia chips they bought with borrowed money.

Zoom out and the loops multiply. OpenAI has reportedly committed on the order of $1 trillion across a handful of suppliers, Broadcom, Oracle, Microsoft, Nvidia, AMD, Amazon, and CoreWeave among them, while its own projections show it burning tens of billions of dollars for years before any profit. By 2026, analysts have identified more than $800 billion of these interlocking arrangements, where the same dollar can be counted as one company's investment, another's funding round, and a third's revenue. Observers have taken to calling it an infinite money glitch, and the joke carries a real warning. When every major participant is simultaneously the others' investor, supplier, and customer, revenue can look robust and growth can look organic even when much of it is the same money traveling around the ring.

Is This 2001 Again?

Anyone who watched the last great infrastructure boom will feel a shiver of recognition, and the comparison deserves to be made honestly, in both directions.

The bearish case writes itself from history. During the telecom and dot-com buildout of the late 1990s, equipment makers like Lucent, Nortel, and Cisco lent money to the customers who bought their gear. Near the peak, the financing those suppliers had extended exceeded a tenth of their annual revenue, and when demand failed to arrive on schedule, the loans went bad in enormous numbers. Nortel went from a peak value of roughly $398 billion to essentially zero. The fiber that got laid sat mostly unused for years. The technology was not wrong, it was just early, and being early is indistinguishable from being wrong if the debt comes due first. There is a specific version of that risk here: lenders are extending credit for five and eight year terms against data centers full of chips that could be made obsolete in three, whether by a more efficient competitor or by the hyperscalers' own custom silicon, which would gut the value of the collateral underneath the neoclouds.

The bullish case is genuinely stronger than the doom crowd admits, and honesty requires giving it full weight. The 2001 comparison breaks down at the balance sheet. The companies at the center of this cycle are not speculative carriers with no revenue. Nvidia lends against more than $50 billion in annual operating cash flow, a fortress net cash position, a top-tier credit rating, and audited books with no hint of the accounting games that later sent telecom executives to court. More important, the end demand appears real and is growing fast. One leading AI lab saw its annualized revenue jump from $1 billion to $5 billion in under six months, and a major cloud provider reported that demand for its AI services nearly tripled in half a year. Vendor financing is not automatically a scam. It helped build the railroads and the telephone network. It magnifies losses only if the underlying demand fails to show up.

So the honest answer is that both things are true at once, and the whole question reduces to a single variable. The circular money can flatter revenue for a while, but the one number it cannot fake is durable demand from customers outside the loop, real businesses and consumers paying real money for AI they actually use. Today that gap is wide: the industry is spending on the order of $400 billion a year building this out against perhaps $100 billion in realized AI revenue, and only durable growth from real customers can close it. If that external revenue keeps compounding, this is aggressive but rational growth financing. If it stalls, the loop unwinds and the losses cascade exactly the way they did a generation ago. Everything hangs on which of those it turns out to be.

What Could Go Wrong

Grant the bull case its due and the picture is still riskier than it was a year ago, because the shift to debt changes how a disappointment travels through the system.

A stumble is now a credit event, not just a stock move. When the buildout was funded by cash, a slowdown in AI demand meant lower earnings and a falling share price, painful but contained. Now the same slowdown threatens the ability to service and refinance debt, which is a different and more contagious kind of trouble. Credit problems do not stay politely inside one company. They widen spreads across the sector, raise the cost of the next deal, and can freeze the financing that the entire buildout now depends on. The AI trade and the corporate bond market have been wired together, and a jolt in one is now a jolt in the other.

The refinancing runs straight into a hostile rate backdrop. That $1.5 trillion of borrowing does not get issued once and forgotten. It has to be rolled over, and it is being issued into a world where the Federal Reserve under Kevin Warsh has signaled higher rates for longer, a shift we examined in our piece on the end of the Fed's roadmap. Debt taken on cheaply gets refinanced expensively if rates stay elevated, and a buildout planned around one cost of capital can look very different when the bill to roll the debt arrives at another.

The interconnection is the amplifier. The same circularity that makes the boom look self-reinforcing on the way up makes it self-reinforcing on the way down. When each participant is a creditor or counterparty to the others, one failure does not stay isolated. If a major AI customer cannot pay its cloud provider, the cloud provider struggles to pay the chipmaker, and the stress travels around the ring rather than stopping at one balance sheet. Concentration that looks like strength in a boom is correlation that looks like fragility in a bust.

None of this is a prediction of collapse. It is a description of how the failure mode changed. A year ago, the worst case for AI was a correction in some expensive stocks. Today, the worst case runs through the credit market, and that is a larger and less forgiving place for a mistake to surface.

The Bottom Line

The AI debate has been conducted almost entirely in the language of technology: how good the models are, how fast demand is growing, whether the productivity is real. Those questions matter. But while everyone argued about them, the financing underneath the boom changed character, and that change may end up mattering more than any of them. A buildout that was paid for with cash is now increasingly paid for with debt, and debt does not care how impressive the technology is. It cares whether it gets repaid.

For an investor, this reframes what to watch. The signal is no longer only in the earnings calls and the demand forecasts. It is in the credit market: in how wide hyperscaler bond spreads trade, in the volatility of the credit default swaps on the stretched names like Oracle, in whether the next jumbo deal gets absorbed smoothly or demands a fatter premium. Those are the readings that will move first if confidence cracks, and we have written before about how much the bond market tends to know before the stock market admits it, in The Bond Market Is Shouting and Credit Cracks.

At Wealth Engine Pro, we follow the numbers, not the narrative, and the number that matters most here is the one the circle cannot manufacture: durable revenue from customers outside the loop. As long as that keeps growing, the debt is a rational bet on a real market. If it falters, the same debt becomes the mechanism that turns an AI disappointment into a credit event. The technology may well deliver everything its champions promise. The open question, the one the bond market is quietly starting to price, is whether it delivers before the financing has to be repaid.