The AI Boom Is Moving Risk Off the Balance Sheet—Not Making It Disappear

Big Tech has found a powerful way to finance the AI infrastructure buildout: move the borrowing away from the corporate balance sheet while retaining enough economic support to make the financing work.

That distinction matters. A liability can move. Risk is harder to relocate.

New Financial Times reporting estimates that major technology companies have provided as much as $300 billion of residual-value guarantees tied to AI data centers and computing infrastructure over roughly the past year. The structures can support debt raised by special-purpose vehicles rather than by the technology companies themselves, leaving much of the associated borrowing outside conventional corporate debt totals.

The financing mechanism matters more than the label

A residual-value guarantee is straightforward in economic terms. Investors finance an asset—such as servers or a data center—based partly on what that asset should still be worth later. A financially strong technology company agrees to absorb some of the shortfall if the asset ultimately sells for less than an agreed value.

That support can reduce the lender’s downside and lower the project’s financing cost. But it also means the guarantor retains exposure to precisely the risk the financing structure is designed to address: what happens if today’s enormously expensive AI infrastructure becomes less valuable than expected?

Broadcom’s own regulatory filing makes the economics unusually visible. In its Form 10-Q filed September 10, the company disclosed a backstop associated with AI racks and customer leases. Broadcom said its maximum potential liability after deployment of all the racks would be approximately $29 billion on an undiscounted basis, while the fair value of the backstop was not material at quarter-end. The filing also says no amounts had been paid under the backstop.

Those statements are not contradictory. They illustrate the difference between maximum contractual exposure, expected economic loss, and recognized accounting liability.

Three balance sheets, not one

For CFOs and investors, the useful analytical framework is to stop thinking about a company as having only one balance sheet.

• The accounting balance sheet records assets and liabilities under the applicable recognition rules.
• The contractual balance sheet adds leases, guarantees, purchase commitments, backstops and other obligations that may sit outside headline debt.
• The economic balance sheet asks a harder question: under a stressed but plausible scenario, which cash flows and asset-value risks ultimately come back to the company?

That third view is where judgment becomes important.

The Bank for International Settlements highlighted this issue earlier this year, describing AI infrastructure arrangements involving dedicated vehicles, operating leases, capacity commitments and guarantees as a form of “shadow borrowing.” Economically, these structures can substitute large upfront capital expenditures with multi-year commitments while directing debt toward project vehicles and private-credit investors.

This does not make the structures improper. Project finance exists for good reasons. Assets, funding and risks can often be allocated more efficiently when each sits with the party best equipped to manage it.

But balance-sheet efficiency and risk elimination are not the same thing.

AI creates an unusual residual-value problem

The most interesting risk is not simply whether AI demand grows. It is whether the assets being financed retain enough value while the technology changes at extraordinary speed.

A warehouse can become obsolete, but usually slowly. A semiconductor architecture can become economically inferior much faster. A data center optimized around today’s power density, cooling requirements and hardware can remain useful for decades—or require substantial reinvestment much sooner than expected.

That uncertainty makes residual value unusually important. It also explains why guarantees from highly capitalized technology companies can unlock financing that lenders might otherwise price much more conservatively.

Nvidia’s recent Form 10-Q provides another useful window into the ecosystem. The company says it has undertaken initiatives that include securing and providing guarantees around land, power, data-center shells and capacity where doing so enables customer deployment of its products. Nvidia also notes that its exposure declines as OpenAI fulfills lease payments and that the guarantees cover defined portions of obligations rather than the full cost of a site.

The CFO lesson: reconstruct the economic leverage

Traditional debt-to-EBITDA or debt-to-capital measures remain useful. They are increasingly incomplete for businesses making enormous infrastructure commitments through multiple financing channels.

A better review starts with reported debt and then separately maps:

• uncommenced and long-duration leases;
• purchase and capacity commitments;
• financial and residual-value guarantees;
• vendor-financing exposure;
• special-purpose vehicles dependent on the company’s contracts;
• minimum payments or take-or-pay arrangements; and
• assets whose stressed resale value determines whether a guarantee becomes economically meaningful.

The objective is not to mechanically add every commitment to debt. A purchase agreement is not identical to a bond, and a contingent guarantee is not identical to cash already borrowed.

The objective is to understand how the obligations behave under stress.

Capital structure is becoming part of the AI competitive advantage

There is another implication. The AI race is increasingly not just a contest in models, chips or software. It is a contest in financial architecture.

Companies able to use investment-grade credit strength, long-term contracts and carefully structured guarantees can mobilize outside capital without funding every dollar of infrastructure directly. That can preserve liquidity and expand faster than a conventional balance-sheet-funded model would permit.

That is genuine financial innovation. It can also make conventional comparisons of capital intensity and leverage less informative.

The right conclusion is therefore not that off-balance-sheet financing is inherently dangerous. It is that investors and finance leaders need to follow the economics through every entity in the structure.

If the project succeeds, the structure may prove remarkably efficient. If utilization disappoints, technology depreciates faster than expected, or refinancing becomes difficult, contracts determine where the losses ultimately land.

Financial engineering can redistribute AI risk. It cannot make the risk disappear.

---

Sources

• Financial Times, “Big Tech uses guarantees to keep $300bn of AI exposure off balance sheets,” September 20, 2026.
• Broadcom Inc., Form 10-Q for the quarter ended August 2, 2026, filed September 10, 2026.
• NVIDIA Corporation, Form 10-Q for the quarter ended July 26, 2026.
• Bank for International Settlements, “Financing the AI infrastructure boom: on- and off-balance sheet borrowing,” March 2026.