Skip to content
Komposite News
Markets · Technology · Capital
Thursday, September 10, 2026Global EditionGot a tip?
AI Infrastructure

Ninety cents on the dollar

A Brookings paper published on Wednesday puts the American AI build-out at $10.3tn by 2032 — larger relative to output than the railroads, electrification or the interstates — and argues its risk has migrated to structures that do not report it. On Thursday, Oracle sent a force majeure notice on a campus whose loans were already trading below par.

Priya Raghavan · Markets Reporter
September 24, 20264 min read
A half-built concrete and steel structure on a muddy construction site beneath heavy grey cloud, a tower crane above itJan Huber via Unsplash
A construction site under heavy cloud. The paper's benchmark is that a 200-megawatt AI campus costs about $8.2bn, roughly two-thirds of it equipment whose useful life is shorter than the debt raised against the building.Photograph via Unsplash

Stijn Van Nieuwerburgh of Columbia Business School has produced the first serious accounting of how the American AI build-out is being paid for. His conference draft for the Brookings Papers on Economic Activity, summarised by Brookings on Wednesday, estimates $10.3tn of investment over 2025 to 2032, averaging 3.63 per cent of GDP a year. His own comparison table is the context: the railroads of 1870–90 ran at 2.24 per cent of GDP, the interstate highways at 1.13, the telecoms and fibre boom of 1996–2003 at 1.10. Nothing in American history has been this large relative to output. Against 57 gigawatts of operating US capacity he counts an announced pipeline of 509, of which he assumes 183 gigawatts arrive by 2032 and 227 never.

The financing argument is where the paper earns its keep. Combined capital expenditure at Oracle, Microsoft, Amazon, Meta and Alphabet rose from about $97bn in 2020 to more than $400bn in 2025 and is projected above $800bn this year, surpassing their combined operating cash flow for the first time. Beyond that point the money comes from outside. Morgan Stanley, whom the paper cites, expects more than half of the $2.9tn needed through 2028 to be external capital, with private credit supplying about $800bn of it.

Meta's Hyperion campus is the worked example, and worth setting out precisely because it is the template — about two gigawatts and $30bn. Meta sold 80 per cent of the equity to Blue Owl for roughly $2.5bn; the joint venture then raised $27bn of debt in October 2025, rated A+ — one notch below Meta's own rating, and the largest single investment-grade corporate issue in American history. That is a debt-to-asset ratio of about 90 per cent at project level and a coverage ratio of 1.12 times, against Meta's own book leverage of roughly 25 per cent. The bonds priced at 6.58 per cent, at least a percentage point above what Meta would have paid unsecured — more than $5bn of extra interest over the life of the deal, reaching bondholders through the lease. Meta is not buying cheaper money. It is buying a thinner balance sheet.

The credit rests on five four-year leases running to the 2049 maturity, each with a termination right and a residual-value guarantee covering any shortfall on sale. Neither reaches Meta's balance sheet until it takes effect, though one must. Moody's puts hyperscaler lease commitments at roughly $970bn, about $660bn of it unrecognised.

Then the circularity. In August a private-credit consortium stood up a $500bn facility to finance purchases of Nvidia chips, against which Nvidia wrote a 25 per cent residual-value guarantee — a backstop of up to $125bn from the supplier of the collateral. Exposures that look independent deal by deal rest on the same tenants, chips and forecast; the beta of listed data-centre REITs against the S&P 500 has duly risen from about 0.5 to about 1.

The underwriting question reduces to one figure: the build-out needs about $3.7tn of annual revenue by 2032 to earn a 10 per cent unlevered return, or roughly $5.50 per installed GPU-hour. That sits inside today's range of frontier rental prices, and has to stay there after 180 gigawatts of new supply arrives. "This is freaking complicated," the author told Reuters, adding that "this opacity of all these special purpose vehicles is somewhat reminiscent of what happened in the subprime mortgage crisis". The paper is more careful: it would be "premature to conclude that AI infrastructure already poses systemic risk comparable to earlier credit booms".

Thursday supplied a data point. Oracle sent a force majeure notice to a Blue Owl unit over Project Jupiter, the 2.45-gigawatt Stargate campus in New Mexico, as TechCrunch reported. Oracle is not walking away as anchor tenant; it is preserving its position on payments should the site miss 2028. The causes are what the paper files under execution risk: an Energy Transfer pipeline delayed six months after repeated permit refusals, and an air-quality permit for the campus fuel cells still pending. "Project Jupiter remains on our planned schedule," Oracle said, and Blue Owl said the notice changed nothing about its commitments. The shares fell about 4 per cent.

One thing about that campus is worth getting right, because most accounts do not. The $165bn of industrial revenue bonds approved by Doña Ana County is not borrowing: in New Mexico an IRB is a property-tax abatement under which the county takes nominal title for up to thirty years while the developer arranges its own financing. The actual financing is roughly $18bn of syndicated leveraged loans, which Bloomberg reports were quoted between 89 and 91 cents on the dollar before the notice went out. This desk could not read Bloomberg at source — its 23 September newsletter on data-centre debt returned an error to us, and nothing here rests on it — but Oracle's own paper points the same way: S&P cut the company to BBB− in July, and its ten-year bonds now yield around 6.5 per cent.

A project loan at ninety cents is not a crisis. It is a syndicate asking for more cushion on a build that slipped two quarters because a pipeline did not arrive. It is also the first public price on the risk the paper describes: long-dated debt written against one tenant, one site and one demand forecast, in a structure whose ultimate holders are not disclosed.

MORE ON THIS STORY