The $300 Billion Question: Who Owns the Downside on Big Tech

Roughly $300 billion of off-balance-sheet AI debt has built up behind Big Tech's data center and chip construction over the past year. The financing runs through guarantees that never show up as borrowings in company accounts. The pattern repeats across issuers: a technology company takes a minority equity stake in a facility, commits to a lease of 15 to 20 years, and guarantees the asset's residual value if the project misses expectations. Accounting rules treat that guarantee as a contingent liability judged out-of-the-money and not yet probable, so reported leverage and credit ratings do not change.

Over the past twelve months, technology companies have pledged as much as $300 billion against AI data centers and chips. Most of that exposure sits in footnotes rather than in reported debt. Meta, Nvidia and Broadcom are among the issuers. Morgan Stanley, using a wider definition that folds in lease commitments and credit support, counted more than $3.1 trillion of obligations across seven hyperscalers and AI chipmakers.

How the Guarantees Work

The mechanics follow a repeatable template. The technology company supplies a minority of the equity in a special-purpose vehicle that owns the facility. That vehicle borrows the remainder, and the technology company signs a long lease whose payments are the cash flow lenders rely on. The residual-value guarantee is the clause that changes the credit calculus, because it moves part of the asset's terminal value risk back to the guarantor.

For lenders, the guarantee means underwriting partly as if the borrower were Nvidia, Meta or Broadcom rather than a startup or a single-purpose data center company. That lowers the cost of funding for projects that would otherwise be priced as speculative infrastructure. For the guarantor, the accounting classification keeps the obligation in the footnotes while the finished capacity is counted as productive assets.

The classification itself is a judgment made at origination. A guarantee is deemed out-of-the-money and not yet probable only as long as a project's projected cash flows support the residual value assumption. If those projections weaken, the accounting question returns and the liability has to be reassessed.

The structure also has a role in equipment supply. The same guarantee format backs hardware delivered to Anthropic, where a chip supplier's credit strength supports financing for compute the model developer needs but cannot fund on its own balance sheet.

The size of each guarantee depends on the residual value written into it, and that figure rests on how slowly the underlying hardware is expected to depreciate. Brookings identifies rapid technological change as one of the risks the whole arrangement depends on, which puts the residual assumption at the center of the credit analysis rather than in the appendix.

The Numbers Behind the Off-Balance-Sheet AI Debt

Alphabet's disclosed data center guarantees moved from $16.9 billion to $43.8 billion within six months, with less than 2% of that total reaching the company's actual balance sheet. The guarantee book grew by more than 150% across two quarters while reported debt barely moved, the clearest available measure of how fast the structure scales.

Item	Disclosed scale	Balance-sheet treatment
Residual-value guarantees, sector-wide	Up to $300 billion over 12 months	Contingent liability in footnotes
Alphabet data center guarantees	$16.9 billion to $43.8 billion in six months	Under 2% recorded on balance sheet
Morgan Stanley tally, seven hyperscalers and chipmakers	More than $3.1 trillion	Off balance sheet
Lease terms on guaranteed facilities	15 to 20 years	Lease commitment

The gap between the $300 billion headline and Morgan Stanley's $3.1 trillion figure is a definitional one. The smaller number covers residual-value guarantees alone. The larger number folds in lease commitments and other credit support, which carry similar demand risk but sit outside the guarantee line. Both matter for anyone sizing the exposure, because the lease is what makes the guarantee callable in practice: an operator that walks away from a 15-year commitment triggers the residual-value clause.

The Enron Parallel, and Where It Stops

Investor Steve Eisman has described the financing methods used by Oracle and Meta as echoing structures from the Enron era. He did not accuse Meta of fraud, and the distinction is material. Enron's collapse involved concealed losses, while residual-value guarantees are disclosed in filings and permitted under current accounting standards.

The more concrete warning comes from the Brookings Institution, which argued in a paper published on September 25 that AI investment risk is shifting from transparent on-balance-sheet financing at large corporations toward opaque arrangements including joint ventures, private credit, securitization, special-purpose vehicles, lease commitments and loan guarantees. Those arrangements depend on the cash flows and collateral values of AI companies, which are exposed to uncertain demand, rapid technology turnover and timing risk.

Who Owns the Downside

The debt is disclosed, aggregated in footnotes rather than itemized as borrowings. The open question is who holds the downside if AI demand undershoots the assumptions built into these leases and residual values. A residual-value guarantee turns a demand shock into a balance-sheet event for the guarantor at the moment an asset is written down or a lease is terminated early.

Concentration is the transmission channel. Where several hyperscalers guarantee capacity serving overlapping pools of AI customers, one demand correction reaches multiple guarantors at once, and the collateral values lenders relied on fall together rather than independently. Morgan Stanley's tally indicates that credit support extends well beyond the residual-value guarantees that have drawn public attention.

The Trade-Offs

The case for the structure is straightforward. It lets operators add capacity at a lower cost of capital without a ratings downgrade, and it preserves cash for chip purchases and model development. The cost is visibility into correlation. Investors see each guarantee on its own, but rarely see the combined exposure across a portfolio of facilities that all depend on the same demand curve.

For the operators, the arrangement also shifts the timing of recognition. Capacity arrives and revenue starts, while the guarantee obligation surfaces only if and when the residual test fails. That mismatch explains why the exposure tends to grow quietly in the years when AI demand is strong and appears abruptly in the year it is not.

Two failure modes deserve attention. The first is a demand plateau rather than a collapse, where facilities keep running at utilization below the levels their lease assumptions require, producing repeated write-downs across several years instead of one clean loss. The second is a technology shift that shortens the useful life of current-generation accelerators well before the 15-to-20-year lease terms expire, leaving residual values anchored to hardware that no longer holds a market.

Neither scenario requires misconduct, and neither would automatically breach accounting rules. That is what makes the exposure difficult to price. The contingent liability stays contingent until it does not, and the trigger is a demand forecast rather than an event a lender can watch in real time.

For decision-makers, the practical consequence is that AI capacity and AI credit exposure are now separate figures. A company can contract for substantial new compute while reported debt stays flat, so diligence has to move into footnote disclosures, lease schedules and residual-value assumptions. Comparing hyperscalers on leverage ratios alone now compares numbers that no longer measure the same risk.

Why this matters

The AI buildout is being financed on collateral values that rest on demand assumptions no one has tested through a downturn. Because the guarantees stay contingent, the market prices AI capacity on credit strength that looks intact precisely because the risk has been moved to footnotes. When those assumptions meet a real demand test, the write-downs will land on the guarantors named in the filings, and the ratings that stayed unchanged will be the first thing to move.

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.