The AI buildout began as a cash story and turned into a credit story. For its first years the spending was funded out of the deepest corporate cash flows in the world, and the question was only how much of it there would be. That changed once the numbers outgrew the cash. CoreWeave showed the template early, converting from crypto mining to AI cloud on two loans from Magnetar in 2021 and 2022 and spending nearly all of it on Nvidia chips — debt secured against GPUs. By October 2025 the pattern had reached the largest balance sheets, when Meta sold $30 billion of investment-grade bonds in one six-part deal to term out data-centre spending. This page follows what happened after that: how much of the buildout is now borrowed, through which instruments, and who is left holding the exposure when it goes wrong.
The borrowing did not arrive as one thing. It arrived as bonds, as syndicated bank loans, as private-credit facilities, as data-centre securitisations, and — increasingly — as structures designed so the debt does not appear on the borrower's balance sheet at all. Moody's counted roughly $460 billion of direct debt across six hyperscalers alongside $1.2 trillion of lease commitments, more than $820 billion of it on leases that have not yet begun because the buildings are unfinished. Goldman Sachs, measuring the five biggest spenders, found $108 billion of investment-grade issuance against 2025 capital spending and $194 billion in the first half of 2026 alone, and expects debt to fund more than a third of the buildout by 2027. Alongside those totals runs a parallel channel that the totals miss: joint ventures in which an asset manager owns most of a campus and raises the debt, as BlackRock did for Meta's El Paso site and Blue Owl for Hyperion; private-credit vehicles anchored by insurance money, as Apollo and Blackstone built with Broadcom to fund Anthropic's compute; and bank syndicates lending against a chip supplier's or a cloud provider's guarantee rather than the tenant's own credit. The collateral itself has now begun to qualify as well: a broadly syndicated loan secured on GPU servers carries an investment-grade rating, which brings the hardware-backed borrowing CoreWeave pioneered into the same rated markets that price ordinary corporate bonds.
The consequence is that the risk keeps moving away from the people who chose it. Banks arranging a $15 billion loan for a Google-guaranteed, Anthropic-leased campus in Texas moved within days to refinance it into the bond market, leaving investors — not lenders — carrying the construction-phase risk that the guarantee does not yet cover. American regulators then determined that a large class of data-centre securitisations are not asset-backed securities at all, removing the post-2008 requirement that a deal's sponsor keep a slice on its own books. And the exposure has now reached the most passive holders of all: hyperscalers are approaching 5% of the dollar investment-grade index and eight of the ten largest bond deals of 2026, so index-tracking funds buy this credit as a function of its size rather than a judgement about it. Prices have begun to register the shift — Oracle was downgraded to one notch above junk, the cost of insuring hyperscaler debt against default is at record levels, and the El Paso bonds priced wide of a comparable deal a year earlier.
What is not yet established is whether any of this is distress or simply scale. No hyperscaler has defaulted; leverage at most of them remains inside investment-grade thresholds, with Oracle the outlier; the swaps market pricing the alarm is small enough that a few dozen trades a day set its level, and the analysts nearest it read the widening as hedging rather than expectation. The measurements that would settle the argument are already named in the record — whether debt-funded capital spending passes a third of the total, whether free cash flow deteriorates far enough to move leverage ratios, whether securitisation issuance climbs once sponsors need retain less of it. Each is recorded here with a date attached and will be checked against filings when it comes due.