AI Companies Are Trying to Hide a Staggering Amount of Debt

A short Futurism item by Victor Tangermann, published 22 July 2026. It is aggregation rather than reporting, and the reporting it aggregates is worth separating from the frame Futurism puts around it.

What Nikkei found

The underlying work is a Nikkei Asia investigation into how the data centre buildout is being financed. Its findings, as relayed:

Alphabet, Microsoft, Amazon, Meta, and Oracle together carry an estimated $1.65 trillion in debt that does not appear on their balance sheets. That is larger than the $1.35 trillion the same five reported in their most recent quarter. Meta accounts for roughly $420 billion of the off-balance-sheet total on its own. The structures involved are special purpose vehicles and similar arrangements, including legally distinct subsidiaries. The same companies are also issuing new shares to raise funds, which carries dilution risk if investor confidence slips. Four of the five were scheduled to report second-quarter earnings within weeks of publication.

The one quote from outside the Nikkei piece is from technical accounting consultant Tom Selling, speaking to Bloomberg Tax: "The accounting treatment itself is in fashion. But what if one of these companies was a house of cards and was propping itself up with this accounting treatment? To me, that's the risk."

What Futurism adds

The headline says the companies are "trying to hide" the debt, and the body says they are "hiding" it. Nothing in the reported material establishes intent, and the vehicles described are a standard financing structure whose existence is disclosed — the reason Nikkei could estimate the figure at all is that the information is available to someone willing to add it up. The gap between "not consolidated onto the balance sheet" and "concealed" is where most of the article's rhetorical work happens.

The Enron comparison is Futurism's, sourced to a Yahoo Finance piece rather than to Selling, and Selling's own quote is conditional: he is describing a risk that would exist if one of these companies were a house of cards, not asserting that any is. The line calling off-balance-sheet arrangements "often a glaring sign that something is deeply amiss behind the scenes" is unattributed. "Experts continue to warn of an AI bubble" is the vague-attribution move, and its link goes to Futurism's own earlier article.

The piece also does not say what the vehicles are actually for. Joint-venture data centre financing, where a chip or infrastructure partner takes part of the capital risk, is the ordinary use of these structures in this industry, and knowing which share of $1.65 trillion is that versus something less benign would change the story considerably. No comment from any of the five companies appears, and Nikkei's methodology for arriving at the estimate is not described.

What survives

The magnitude and the concentration are the parts worth carrying forward: off-balance-sheet obligations exceeding reported debt, across five companies, tied to a single capital programme whose demand assumptions are unproven. Whether that is fraud, prudent risk-sharing, or an accounting fashion that will look reckless in hindsight is not something this article can settle.

It slots into an existing thread. subprime-ai-crisis is Zitron's case that AI demand is subsidy-driven, with a predicted collapse sequence; ai-too-expensive is his follow-up with the break-even arithmetic. ai-subsidy-economics traces the subsidies link by link. ai-bubble-pale-horses is the vault's running checklist of deflation warning signs, and a debt structure larger than reported debt is the sort of item it exists to track. speculative-growth-ai-bubble is the argument that cuts the other way: Caballero's model in which an unsustainable valuation still leaves a permanent real legacy, because the capital gets installed before the correction lands. On that reading the financing structure matters less than whether the concrete and the transformers end up built, and hold-on-to-your-hardware documents how much of the world's memory and storage supply the buildout is already consuming.