Working Paper 24
Hidden Leverage in the AI Infrastructure Boom: A Forensic Investigation and Historical Comparison
October 2026
Abstract
This paper conducts a forensic investigation of how the global AI infrastructure buildout is financed, and whether its financing structures resemble six historical episodes of hidden leverage: Enron, WorldCom, Tyco, the 2007-2009 global financial crisis, Lehman Brothers' Repo 105, and Greensill Capital. We reconstruct obligation-level exposure maps for Microsoft, Amazon, Alphabet, Meta, Oracle, OpenAI, Anthropic, CoreWeave, Nvidia, Broadcom, and xAI from a five-vintage dataset of 116 obligations across 84 entities with primary committed exposure of $4,315.4 billion; catalog seven financing mechanisms (SPVs, sale-leasebacks, residual-value guarantees, operating and finance leases, purchase commitments, supplier-arranged customer financing) with their treatment under US GAAP and IFRS; and run nine-scenario coverage stress tests over 2027-2030 with a collateral-depreciation model for GPU-secured debt. We find no evidence of fraud: the structures are disclosed in filings, and the post-2002 accounting regime (ASC 810, ASC 842) blocks the specific tricks used in the historical cases. The genuine parallels are structural, not criminal. The funding architecture of the buildout resembles pre-crisis shadow banking (private-credit intermediation outside the bank perimeter, SPV layering, maturity transformation, originate-to-distribute incentives), and Greensill is the closest single-firm analogue (private-credit intermediation of future economic claims, guarantee dependence, arranger opacity). The solvency core holds: Microsoft, Amazon, and Alphabet never breach 1.0x fixed-obligation coverage in any scenario, including the severe case, and Meta breaches only in 2027 of the severe case (0.71x) against a $90.3 billion cash buffer. The periphery is fragile: CoreWeave breaches from 2027 in every case, Oracle compresses to 1.33x base coverage by FY2029 and breaches in the severe case, Meta faces a one-year 2027 squeeze, and GPU-secured facilities go underwater on hardware value within two years under every depreciation assumption. The verdict is a solvent core funding a fragile periphery through increasingly circular structures: a credit cycle with genuine systemic-fragility elements in its transmission architecture, not a fraud story and not a clean bill of health.