The $120bn off the books: judge AI datacenter debt on tenants, not on accounting

The debt does not go away in these structures; it changes address. The vehicle borrows, banks and outside investors own most of it, and the tenant holds a contract for exclusive and full use of the building once it is finished ^[2]. That contract is what the lenders underwrote, which is why the accounting treatment is the least interesting part of the file. What they hold is a claim on one occupant with no obvious substitute if that occupant walks.

Scale first. The $120bn the Financial Times counted off balance sheet as of December 2025 ^[1] is about 2.3% of the $5.3tn Goldman Sachs thinks hyperscalers could spend on AI and datacenters through 2030 ^[4][7]. Goldman also expects private markets to carry more of that financing as the buildout proceeds ^[4]. Whatever this structure is, it has not run its course.

The occupancy numbers are the strongest part of Gene Marks's argument in the Guardian, and they are also the part that can turn. CBRE has North American capacity up 36% last year with vacancy still falling to a record 1.4%, and demand outpacing supply in nearly every major market ^[5][6]. For the vacant share to fall while the base grew that much, leased capacity had to grow by more than 36% ^[9]. Construction is not the binding number. Absorption is.

Where the Enron comparison misleads is not in its verdict but in its subject. Enron collapsed in 2001 and cost shareholders tens of billions ^[10], and Marks is probably right that fraud at that scale is not being repeated, given disclosure requirements and scrutiny far heavier than the 1980s vehicles faced ^[11][14]. But "not fraud" and "not a loss" are separate findings, and his own precedent shows it. Centocor's partnerships were ordinary practice, held at a minority interest, funded by limited partners, with Centocor keeping exclusive rights to the drugs ^[12]. Some of those drugs failed in clinical testing, and no panic followed because the risk had been spread to the partners ^[13]. The structure worked as designed, and money was still lost.

Marks's distinction is that a datacenter can disappoint financially without disappearing the way a failed trial does ^[17], which is why he reaches for Bezos's description of an industrial bubble, the sort that leaves usable assets behind ^[8]. Fair, though it relocates the question rather than answering it: disappointment then arrives as a resale price on land, power and equipment. Reading the footnotes and arguing about consolidation, as he suggests ^[15], settles the accounting. It does not tell you how full the buildings are.