Borrowed Time

Robert Cohen runs credit at DoubleLine, the bond shop founded by Jeffrey Gundlach, and he is not a man given to theatrics. So when he was asked recently to put a probability on whether the wave of debt now financing the artificial-intelligence build-out would end in a bubble, his answer was striking less for its drama than for its precision. He did not say "likely." He did not hedge with "if conditions persist." He said the probability was 100%.

Not because the technology is fake, or the companies are frauds, or the demand is imaginary. Cohen's reasoning was older and colder than that. Every great wave of infrastructure investment in financial history — the railroads, the telegraph, the fiber-optic boom of the late 1990s — has been financed, in its final innings, by debt raised on the assumption that the future would arrive on schedule and on budget. And every one of them, without exception, eventually issued more bonds than the future could service. The bubble, in Cohen's telling, is not a risk. It is a stage. The only question is when.

This is the part of the AI story that the equity markets, transfixed by the soaring share prices of Nvidia and its customers, have been slow to look at directly: the boom is not being paid for out of profits. It is being paid for with borrowed money, on a scale and with a set of structures that should be familiar to anyone who lived through 2008 — and the people whose job is to price risk for a living have begun, quietly, to back away.

The number, and the number behind the number

Start with the visible debt, because it is alarming enough on its own.

The big US technology companies — the "hyperscalers," the Microsofts and Metas and Oracles racing to build the data centers that train and run AI models — have together sold more than $155 billion of unsecured bonds globally, a figure already up more than 45% from their entire issuance the year before. Wall Street expects the hyperscalers alone to issue $250 billion to $300 billion of bonds in 2026. And that is just the investment-grade, on-the-books portion — the debt these companies are willing to show you.

Behind it sits a number that is harder to see and more worrying. According to estimates of the capital required, the data-center build-out will consume something on the order of $3 trillion through 2028 by Morgan Stanley's reckoning, and more than $5 trillion by JPMorgan's once you include the power plants and grid connections to feed it. Cohen's own framing is the one that should give a bond investor vertigo: if the hyperscalers fund merely half of their projected $4.1 trillion of capital spending between 2026 and 2030 with debt, they would account for roughly 15% of all non-financial corporate bond issuance on earth. A single industry, betting on a single technology, would become a sixth of the global corporate bond market.

When one theme grows to that weight inside the thing that prices the world's debt, it stops being a sector and becomes a systemic exposure. And the most aggressive part of the borrowing is the part designed not to appear on any balance sheet at all.

Off the books

Here is the mechanism that should make the hair stand up.

Oracle, Meta, xAI and CoreWeave — among the most voracious builders in the field — have together moved an estimated $120 billion of AI-infrastructure debt off their own balance sheets, into special-purpose vehicles: legally separate shell entities that own the data centers and borrow against them, so that the parent company can enjoy the compute without carrying the loan where its shareholders and creditors can plainly see it. The capital for these vehicles has come from the giants of private finance — JPMorgan, Pimco, BlackRock, Blue Owl — the same private-credit machine that has swelled, largely in the dark, into a $1.7 trillion market over the past decade.

The template is Meta's Hyperion campus in Richland Parish, Louisiana. To build it, Meta and Blue Owl assembled a roughly $30 billion package — about $27.3 billion of fully-amortizing senior secured notes, arranged by Morgan Stanley through a purpose-built vehicle — that keeps the borrowing off Meta's own books even as Meta uses the data center. The notes are due in 2049. Hold that date.

If the phrase "off-balance-sheet special-purpose vehicle funded by the shadow-banking system" rings a faint and unpleasant bell, it should. It is the precise architecture that turned a manageable problem in American mortgages into a global catastrophe in 2008. The point is not that history will repeat in detail — the assets here are real data centers, not liar loans — but that the structure is the same: leverage moved out of sight, risk repackaged and sold on to investors a step removed from the underlying, and a collective comfort that because everyone is doing it, it must be safe. The banks themselves have begun quietly offloading their AI-data-center exposure to private-equity buyers through "significant risk transfer" trades — which is to say the institutions closest to these loans are paying to get the risk off their own books, even as they originate more of it.

The maturity mismatch

Now the detail that turns a debt story into something stranger, and which Cohen flagged almost in passing.

Many of the bonds being sold to finance this build-out will not mature for decades — Meta's Hyperion notes run to 2049. They are being used to buy, and house, graphics processors with an economically useful life, as the bears of CoreWeave and the warnings of Michael Burry have made vivid, of perhaps three to five years. Read that mismatch slowly. Investors are lending money until 2049 against machines that will be electronic scrap by the early 2030s — and will have been replaced four or five times over before the note they financed even comes due. The bond will still be paying coupons long after the asset it financed has been hauled to a recycler and replaced two or three generations over.

That is only sustainable if the cash flows — the rent paid by AI companies to run their models — keep coming reliably enough, for long enough, to service debt that outlives the hardware many times over. It is a bet not on any individual data center but on the permanence of the entire AI economy: that demand will not merely survive but compound, uninterrupted, across decades, through every recession and technology shift between now and the 2050s. Bond investors are, in effect, being asked to underwrite the proposition that the AI boom is not a boom at all but a permanent condition. History's verdict on permanent conditions is not encouraging.

The smart money is already leaving

The tell, as ever, is in what the most informed participants do rather than what they say.

DoubleLine and Oaktree — two of the most respected distressed- and credit-investing shops in the world, the latter built by Howard Marks on a career of buying what others were forced to sell — are not waiting for the bubble Cohen says is certain. They are positioning for it now: buying the safer, shorter, more defensible pieces of the AI-credit complex and bracing for the pain in the rest. A Bank of America survey of credit investors recently found, for the first time ever, that an "AI bubble" had become their single greatest concern — displacing inflation, recession and war at the top of the worry list. Bond desks have reportedly begun circulating internal "bubble survival guides."

None of this means the reckoning is imminent. Bubbles inflated on real technology and real demand can run for years past the point of reason — the fiber boom of the late 1990s laid genuinely transformative infrastructure on its way to vaporizing hundreds of billions of dollars of bondholder capital. The data centers being built today are real, the demand for AI compute is real, and a great deal of this debt will be repaid in full. But "a great deal" is not "all," and the structures now proliferating — the off-balance-sheet vehicles, the decades-long bonds against years-long assets, the risk laundered through private credit and shunted off bank books — are precisely the structures that turn an ordinary cyclical downturn into a credit event.

Cohen did not say the AI companies would fail. He said the debt would reach bubble levels with certainty, and that the only open question was timing. The equity market is still pricing the dream. The credit market — quieter, older, and far better at counting — has started pricing the bill. When those two prices disagree this sharply about the same companies, it is usually the bond market that turns out to have been reading the documents.

The build-out is running on borrowed money. It is also, the people who lend for a living have begun to suspect, running on borrowed time.