The $1.65 trillion that isn't on the balance sheet

A Nikkei Asia study published on 21 July found that five US tech giants, Alphabet, Microsoft, Amazon, Meta and Oracle, collectively carry an estimated $1.65 trillion in liabilities tied to AI data-centre construction that do not appear as debt on their balance sheets.¹ That figure exceeds the roughly $1.35 trillion in debt those same companies report on their books. The instruments are legal. The disclosures exist, in footnotes. The question I want to ask is a different one: whose money is on the other side.

Because that is what an off-balance-sheet liability is. It is debt that has been moved — into a legally distinct vehicle, funded by someone else, and taken out of the ratio that headline analyst reports use to gauge how leveraged a company is. The debt still exists. The risk still exists. Someone still holds the paper.

What the structure actually does. A special purpose vehicle, or SPV, is a separate legal entity a parent company sets up to take on debt for a specific asset — in this case, data centres and long-term GPU supply contracts. Under current US and international accounting rules, if the parent does not exercise sufficient control over the SPV, the SPV's liabilities stay off the parent's consolidated balance sheet. This is standard project finance. It is how airports, toll roads and power plants have been built for decades. The mechanism is not new and it is not, by itself, a scandal.

What is new is the scale, the speed, and the asset class. The Nikkei study reports the aggregate has risen roughly eightfold in about four years across the same five companies.¹ Meta alone accounts for approximately $420 billion of the total, which Futurism, reporting the Nikkei figures, describes as nearly triple its transparent debt.²

Who is on the other side. The reporting mostly does not answer this, and it is the question that matters most for a distributional read. When a hyperscaler's SPV borrows to build a data centre, the lender is typically some combination of private credit funds, insurance companies, infrastructure funds, and increasingly, structured vehicles that repackage the exposure further downstream. Those funds have beneficiaries: retirees drawing pensions, policyholders paying premiums, savers whose money managers have reached for yield in a market where yield has been scarce.

The Bank for International Settlements flagged part of this pattern in a bulletin dated approximately June 2026, according to the explainx.ai breakdown of the Nikkei story: AI capex was outrunning hyperscaler internal cash flow, and private credit was extending to AI-related borrowers at spreads that looked too tight for the risk, with an aggregate reported at roughly $200 billion.³ Treat that number as second-hand until the BIS bulletin is read directly. The direction of travel is what matters: credit is flowing into AI infrastructure faster than the priced risk suggests it should, and the Nikkei figure now puts a much larger denominator behind the same warning.

What is different from Enron, and what is not. The comparison is the one everyone reaches for, and it does real rhetorical work that the facts do not fully support. Enron's offence was fraud: hiding losses from auditors and inventing earnings. The current SPV structures are disclosed in regulatory filings. They passed the post-2001 consolidation reforms that were designed specifically to stop the Enron pattern. That is a genuine difference and it should not be waved away.

The disanalogy sharpens somewhere else. What Enron hid was the truth about a specific company. What the current structures obscure, at scale, is risk concentration across an entire industry whose central financial assumption — that AI revenue will grow fast enough and durably enough to service the debt behind the data centres — is unproven. Tom Selling, described as a technical accounting consultant, told Bloomberg, in a quote reproduced by Futurism, that the treatment itself was in fashion but that the risk was what happens if one of these companies turns out to have been propping itself up with the accounting.² Selling is talking about a single house of cards. The structural version of his question is whether the industry is holding up the industry, and where the exposure lands when a data-centre cash-flow assumption underperforms.

The information asymmetry is not random. An institutional investor with an analyst team can model SPV exposure from the footnotes of quarterly filings. It is tedious, but it is possible. A retail investor cannot. A pension beneficiary certainly cannot. The people best equipped to price the risk are the people least likely to be holding it at the tail; the people least equipped to price it are, disproportionately, the people whose retirement or insurance depends on funds that have reached into private credit for yield.

The pattern keeps repeating in every recent financial episode where opacity does distributive work. The 2008 mortgage collapse worked the same way: risk migrated to counterparties whose ultimate exposure ran to households that had no line of sight on what their money markets or pension funds held. The current structures are not equivalent to subprime CDOs. But the mechanism by which sophistication and information map onto the wealth distribution is the same mechanism, and it produces the same asymmetry in who eats losses when losses arrive.

What I do not know. I do not know Nikkei's methodology for the $1.65 trillion — which leases, which GPU contracts, which discount rates.¹ I do not know the individual breakdown across the four companies that are not Meta. I do not know, and the reporting does not tell me, which specific private credit funds and insurance vehicles hold the largest concentrations of this paper, and I do not know what disclosure their beneficiaries receive. Those are the things I would want to see before saying with any confidence how badly a repricing of AI infrastructure risk would land, and on whom.

Four of the five companies were scheduled to report Q2 earnings in the days and weeks following the Nikkei publication, per Futurism.² The specific figure to watch is whether any of them voluntarily consolidates SPV liabilities into their headline debt disclosures, or expands the footnote detail — and if not, whether analysts on the earnings calls press for it.

Footnotes