The $1.65 Trillion Hidden Inside America's AI Buildout

On 20 July 2026, Nikkei Asia published a number that should embarrass every credit analyst who spent the last four years calling artificial intelligence a once-in-a-generation infrastructure cycle. Hidden, off-balance-sheet debt at five US technology giants has climbed to an estimated $1.65 trillion, an eightfold increase in four years, as the companies financing the AI buildout push more of the obligation into vehicles the headline numbers do not capture.

The thesis here is uncomfortable but worth stating plainly. The AI capex boom is not being financed the way the equity market has been told it is being financed. Earnings calls celebrate operating cash flow; balance sheets quietly accumulate guarantees, lease commitments, and structured-finance vehicles tied to data centres, GPUs, and the electricity contracts that feed them. The harder question is whether this opacity is a feature of the cycle or a fault line running through it.

What "hidden" actually means

The phrase "hidden debt" is doing a lot of work in the Nikkei reporting, and it deserves unpacking. Standard accounting rules let companies keep certain financing arrangements off the consolidated balance sheet when the entity doing the borrowing is technically separate, even if the parent has issued a guarantee, holds a put option, or has contracted to absorb the asset at the end of the structure's life. Special-purpose vehicles, synthetic leases, and supply-chain financing programmes are the usual culprits.

In the AI context, the vehicle of choice is the data-centre joint venture. A hyperscaler commits to a multi-year lease on a building it does not own, owned by a partnership it does not consolidate. The partnership borrows against the lease; the hyperscaler guarantees the cash flows. The obligation is disclosed, technically, in the footnotes. But it does not appear in the headline leverage ratio that drives ratings and equity multiples.

The eightfold jump Nikkei documents therefore reflects two things happening at once: the absolute scale of AI infrastructure spending has exploded, and the legal architecture used to finance that spending has grown more creative. The companies have not necessarily become more leveraged in a way rating agencies would call reckless. They have become more leveraged in a way the summary financials cannot see.

The insider signal

A second data point landed the same day, and it sits awkwardly next to the first. Unusual Whales, a US market-data service that tracks corporate insider transactions, published a comparison of current insider selling with prior cycles. The headline finding: the last time insider selling reached similar levels was during the dot-com bubble, which was followed by a significant market correction.

The caveat is structural. Insider selling is a noisy signal. Executives sell for a thousand tax, liquidity, and diversification reasons that have nothing to do with their private view of the share price. A wave of selling around the all-time highs of 2026 could reflect the same lock-up expirations and diversification patterns that produced selling at every prior market peak, without any insider actually believing the AI trade is finished.

The honest read is that the signal is consistent with a late-cycle distribution, not a verdict. The combination, though, is what matters. The people closest to the capex decisions are, in aggregate, monetising their equity exposure at the same moment their companies are quietly parking ever-larger financing obligations in structures that minimise their visible footprint. Whether by coincidence or by coordination, the flow tells a story the press release does not.

Why the structure matters now

A $1.65 trillion obligation is not, on its own, alarming. US non-financial corporate debt is well above $10 trillion. What makes the AI tranche unusual is its velocity. Eightfold growth in four years implies a doubling roughly every eighteen months, a pace that has to either decelerate sharply, get absorbed into visible balance sheets, or continue expanding the off-balance-sheet footprint.

Each of those paths carries different risks. A deceleration would imply the hyperscalers have decided the demand curve for AI compute is flattening faster than their contracted commitments assumed, which puts downward pressure on the chipmakers and power utilities that have been planning around continued acceleration. An absorption into visible balance sheets would, for a quarter or two, look like a sharp increase in reported leverage and could trigger ratings actions that raise the cost of the next round of capex. A continuation of the off-balance-sheet path leaves the system exposed to a funding shock if the structured vehicles lose access to wholesale credit at the moment they need to roll.

The structural frame here is older than the AI trade. The same playbook, securitised mortgages in 2007, supply-chain finance at Greensill in 2021, private-credit NAV loans at present, runs through every credit cycle of the last twenty years: an asset class grows rapidly, the financing gets more sophisticated, and the headline leverage ratios stay calm until the day they do not. The defence offered by the AI trade is that the underlying assets, data centres full of GPUs with multi-year contracted tenants, are cash-flowing in a way subprime mortgages never were. That defence is plausible, but it is also exactly what the holders of every prior late-cycle structure said before the structure broke.

What the critics on both sides say

The bear case, articulated by credit analysts who have been raising the alarm for two quarters, is straightforward: the AI buildout is financing itself with duration-matched, non-recourse vehicles that depend on continued low rates and continued hyperscaler appetite. If either assumption cracks, the structures have to be refinanced on worse terms or absorbed by parents that will then face the leverage their vehicles were designed to hide. The first signs would show up not in equity prices but in the spreads on private credit and in the cost of data-centre construction financing.

The bull case, articulated by sell-side equity analysts and the companies themselves, is that this is a transitional accounting choice, not a structural risk. The hyperscalers are cash-flowing at historic levels; the debt is being incurred against assets that are contracted for a decade; the vehicles are a way of matching long-dated infrastructure cash flows with long-dated financing without polluting the parent balance sheet with mark-to-market noise on short-term rate moves. The eightfold growth reflects the rapid scaling of a fundamentally cash-generative business, not the accumulation of hidden risk.

Both cases can be partially right. The honest read is that the visible obligations are manageable, the off-balance-sheet obligations are real, and the question is whether the cash flows from the underlying AI compute demand will be sufficient, predictable, and contractually durable enough to service both.

The geopolitics underneath the capex

The financial architecture is not the whole story. The AI buildout is also a sovereign-industrial project, which is why the financing question matters beyond the credit desks.

The five companies at the centre of Nikkei's reporting are not just private actors chasing returns. They are the operational backbone of the US position in a technology contest with China that has been openly framed, in successive US National Security Strategies and in the export-control architecture of the last three administrations, as a defining national priority. Every dollar of off-balance-sheet financing is, functionally, a dollar of US strategic capacity to train and serve frontier models. The fact that the financing is structured to look smaller than it is, on the relevant financial statements, is a feature inside the corporate finance department and a vulnerability inside the national-security planning process.

The Chinese counter-frame deserves the same weight. Beijing's industrial-policy apparatus has spent the last decade building parallel capacity in compute hardware, model training, and energy supply for data centres, financed through state-directed bank credit that appears on balance sheets and does not pretend to be anything other than strategic. The Western press tends to describe this as opaque; it is, by the standards of US corporate disclosure, less opaque, because the obligations are sovereign, not synthetic, and the strategic intent is declared rather than inferred. The trade-off is real: the US system is faster, more innovative, and produces frontier models at a pace the Chinese system has not matched; the Chinese system is more legible to its planners and less exposed to a credit-cycle shock in the middle of a strategic contest.

Neither model is winning cleanly. Both are running into constraints. The US model is running into the credit architecture described above. The Chinese model is running into the export controls that limit its access to leading-edge compute and into a domestic economy whose consumer demand has not recovered at the pace the industrial buildout assumed.

What to watch

Three near-term indicators will determine whether the $1.65 trillion becomes a footnote or the opening chapter of a credit event. First, the rating-agency posture. If Moody's or S&P starts formally incorporating the off-balance-sheet obligations into their leverage calculations for any of the five, the cost of the next tranche of capex rises immediately. Second, the spread on private credit and on data-centre construction financing. A widening of 50 to 100 basis points, even without a ratings action, would meaningfully change the internal-rate-of-return maths on the next round of capacity. Third, the insider-selling flow over the next two earnings cycles. If the current pace continues or accelerates, the late-cycle distribution case hardens; if it moderates, the noise-floor explanation gains weight.

The question for readers outside the credit complex is the same one that has surfaced in every prior cycle: how much of the AI trade is a productive deployment of capital into infrastructure the economy will use for decades, and how much is a financial structure whose returns depend on assumptions that stop being true if growth disappoints even modestly. The answer, for now, is that both descriptions are partially true, and the gap between them is where the next twelve months of market narrative will be written.

This publication framed the Nikkei disclosure through the credit-cycle lens rather than the standard "AI boom is unstoppable" treatment, because the off-balance-sheet growth rate is the more actionable fact in the public record.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

• https://t.me/NikkeiAsia
• https://t.me/nikkeiasia
• https://t.me/TSN_ua
• https://t.me/TSN_ua