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

On 20 July 2026, Nikkei Asia published a tally that landed like a quiet thunderclap on the AI trade. Hidden debt at five US tech giants had swollen eightfold in four years to an estimated $1.65 trillion, obligations kept off the parent-company balance sheet through a thicket of subsidiaries, joint ventures, special-purpose vehicles and data-centre lease structures. The number, the publication noted, was a function of opaque AI funding: the same buildout that has powered record capex headlines is being financed, in significant part, through vehicles that look nothing like the corporate bonds and term loans investors thought they were underwriting.

The lede obscures a structural shift. AI demand has not merely expanded the capex line on hyperscaler income statements. It has built, alongside the visible spend, a parallel credit market designed to be invisible to the same investors who buy the parent-company paper. Whether that parallel market is sound is the question that will define the next phase of the cycle.

The $1.65 trillion and what is inside it

Nikkei's figure is an aggregation across five US technology companies whose names the publication declined to itemise in its 20 July summary, but whose identity is not hard to guess: the cloud and chip hyperscalers whose capex run-rates have set the cadence of the AI trade for three consecutive fiscal years. The "hidden" portion comprises three large buckets. The first is non-consolidated subsidiary debt, borrowings at data-centre operating companies or chip-financing vehicles that sit below the listed parent but roll up into the same economic exposure. The second is joint-venture obligations, particularly in the Gulf and East Asia, where sovereign capital and US hyperscalers have co-built campuses under structures that let the AI infrastructure offload from the buyer's books. The third is operating-lease commitments, the long-tenor leases on land, power and cooling capacity that accounting standards require to be disclosed only in footnotes.

Each bucket has grown. The eightfold expansion over four years implies a compound rate well above the growth of revenue or operating cash flow at the same firms. The capex cycle, in other words, has been financed faster than the cash cycle that is supposed to amortise it. That gap is the structural fact behind the headline number.

A separate data point from the same week sharpens the picture. On 20 July 2026, Crypto Briefing reported that AI startup Infinity had closed $15 million in seed funding at a $100 million valuation. The round is small in absolute terms, but it illustrates the gravitational pull of the AI infrastructure narrative even at the earliest stage of company formation. Investors are pricing AI exposure at multiples that assume the hyperscaler buildout continues, and that the off-balance-sheet plumbing holds.

The counter-read: this is not 2008

The instinctive comparison is to the structured-credit vehicles of the pre-2008 era. It is also, on the available evidence, the wrong one. The off-balance-sheet debt at US hyperscalers is, in the main, backed by contracts. Hyperscale cloud customers sign multi-year commitments; AI labs sign training and inference agreements; the chip vendors and data-centre operators who sit beneath the hyperscalers have revenue pipelines attached to those contracts. The economic substance is real in a way that the subprime mortgage collateral of two decades ago, often, was not.

That does not make the structure riskless. It makes the risk different. The exposure is to counterparty duration: how long does an AI customer's commitment hold up if the underlying model becomes cheaper, faster, or open-weight, and the customer decides to in-house? It is to power-grid congestion, because a data-centre campus is only as good as the gigawatts feeding it, and grid interconnect queues in the United States have lengthened, not shortened, over the past two years. And it is to interest-rate sensitivity, because the longer the lease tenor and the larger the sovereign-equity partner, the more the structure depends on the cost of the marginal dollar.

A further caution: the $1.65 trillion figure is itself an estimate assembled from public disclosures. Nikkei's reporters reconstructed it from footnote-level data, joint-venture announcements, and lease-commitment tables. The true number could be higher, if some entities are sufficiently opaque that even seasoned financial reporters cannot price their obligations, or lower, if some of the headline-grabbing JV announcements have not yet translated into drawn debt. The sources do not specify which way the error runs.

The funding stack behind the capex

To see why the structure has grown this large, it helps to follow the money from the AI customer backwards. A frontier-model lab signs a multi-year compute commitment with a hyperscaler. The hyperscaler, in turn, has committed to purchase chips from a foundry-and-fab alliance, power from a utility-scale provider, and floor space from a developer who may itself be a joint venture with a sovereign-wealth or pension partner. Each link in the chain is financed separately. Each financing sits in a different legal entity. Each entity has its own creditors.

That separation is the point. It allows the hyperscaler to report capex that looks disciplined relative to its operating cash flow, while the actual infrastructure spend is being shouldered by a wider pool of capital. Pension funds in the Gulf and East Asia get access to long-duration assets denominated in dollars. Hyperscaler shareholders get headline balance-sheet ratios that look closer to a software firm than to a utility. The AI labs get compute capacity that arrives on time.

The cost is opacity. The same investor who reads the hyperscaler's 10-Q has no consolidated view of the indebtedness of the JV that owns the building the hyperscaler leases. The same regulator who approves a chip export licence has no immediate window into the credit lines attached to the fabs the licence benefits. The $1.65 trillion figure is, in part, an attempt to reconstruct what consolidated reporting would look like if it existed.

What the cycle still has to absorb

Three forward-looking tests will determine whether the $1.65 trillion remains a footnote or becomes a stress event. The first is refinancing. A meaningful share of the obligations Nikkei counted will roll inside the next thirty-six months. The shape of the rate curve at that point, and the appetite of the same Gulf and East Asian balance-sheet partners who built the original structures, will set the marginal cost of the next leg of capacity.

The second is utilisation. The economic value of a data-centre campus collapses if it runs below a certain occupancy threshold, because the fixed-cost stack was sized for the contract base. AI inferencing workloads have grown, but they have grown more slowly than the training workloads that justified the first wave of campuses. If inference pricing continues to fall, the unit economics of the next campus become harder.

The third is governance. A debt structure that is invisible to equity holders is also invisible to credit-rating agencies, which is part of why the parent companies retain investment-grade status. If the rating agencies begin to consolidate the off-balance-sheet vehicles into their parent-level views, the resulting downgrades could reshape the cost of capital overnight. Nothing in the public sources suggests this consolidation is imminent. But the question is being asked in the same news cycle as the $1.65 trillion headline, which is itself a marker.

Stakes and what to watch next

The most consequential reader of Nikkei's tally is not the equity investor. It is the central bank, which is responsible for the price of the dollar that ultimately funds the lease and the JV structure. If the off-balance-sheet layer is, in aggregate, dollar-denominated long-duration credit, then the cycle's sensitivity to US monetary policy is broader than the published numbers of the hyperscalers suggest. The Fed's job becomes harder in proportion to how much of the AI buildout is funded through entities it does not directly supervise.

The second most consequential reader is the AI customer. The labs that signed the multi-year compute commitments did so on the assumption that the supplier side could deliver. If a single hyperscaler encounters a credit event at one of its non-consolidated vehicles, the delivery risk flows up the chain within weeks, not quarters. Concentration is the word that prudent operators will use privately.

The third is the same sovereign-wealth partner whose capital built the JV in the first place. Their position is long America, long AI, and long dollar. The $1.65 trillion headline does not change the thesis. It does, however, raise the bar on the diligence they will apply before committing to the next campus, and on the price they will demand for doing so.

The receipts and the gaps

What the sources establish with reasonable confidence: a multi-trillion-dollar off-balance-sheet layer has built alongside the AI capex cycle, its growth rate has outrun the cash conversion of the hyperscalers that anchor it, and the funding model depends on a widening cast of non-bank and sovereign-capital partners.

What the sources do not establish: the precise identities of the five firms Nikkei counted, the dollar composition of each of the three buckets, the share of the obligations that are dollar-denominated versus local-currency, and the maturity profile of the off-balance-sheet layer. The Nikkei summary published on 20 July 2026 is the load-bearing citation for this article; the Crypto Briefing item is illustrative of the venture-capital gravitational pull of the same narrative. Neither is a complete dataset.

The honest framing is that the AI buildout is now too large and too cross-jurisdictional for any single balance sheet to disclose. The $1.65 trillion figure is the first widely-circulated attempt to add them up. It will not be the last. The next attempt will tell us whether the off-balance-sheet layer grew another eightfold, or whether the publication of the first count itself began to slow the build.

Monexus filed this piece as a long read on the AI funding stack. Where mainstream financial press has run the $1.65 trillion as a sensational lede, this publication treats it as a structural data point and reads the funding architecture around it.

Wire provenance

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

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