Big Tech Keeps $300 Billion Out of Frame

The Financial Times reports that technology companies are using guarantees and separate financing structures to support roughly $300 billion of AI infrastructure exposure without placing it directly on their balance sheets. Wall Street can lend against the credit strength of major technology groups while a project vehicle owns the data center. The source material supplied here does not identify the individual companies, lenders, campuses, guarantee terms, or accounting judgments, so the aggregate should not be mistaken for a verified list of hidden corporate debts.

The mechanism is straightforward. A special-purpose company raises money, builds or buys the server campus, and expects lease payments, capacity purchases, or other contracted revenue to repay lenders. The technology customer may promise minimum payments, reserve years of computing capacity, or guarantee part of the financing. Legal ownership stays outside. Economic dependence may not.

That distinction matters because off-balance-sheet does not mean obligation-free. Accounting treatment turns on the exact contract: who controls the asset, whether the arrangement functions as a lease, what a guarantee covers, and how probable a payment has become. A purchase commitment is not automatically debt. It can still behave like debt when the buyer has little practical ability to leave.

The structure preserves the familiar image of the asset-light technology company. Concrete, transformers, cooling equipment, and power contracts sit in another entity. Yet the revival of heavy industrial construction has not abolished industrial risk. It has distributed that risk among project lenders, infrastructure investors, utilities, equipment suppliers, and the future customers expected to pay for AI services.

If demand meets the forecasts, the arrangement can work well. Dedicated vehicles let specialist investors finance long-lived infrastructure, and a strong customer commitment lowers borrowing costs. The technology company gains capacity without funding every wall and cable itself. Lenders receive contracted revenue rather than a wager on an unproven tenant.

The stress case is less tidy. Weak AI demand could leave a vehicle with excess servers and fixed power bills. Rising electricity costs could erode project cash flow. Fast equipment obsolescence could reduce collateral value before the loan matures. If a tenant withdraws, the decisive question is whether lenders absorb the loss or invoke a guarantee, minimum-payment clause, termination fee, or parent-company support agreement.

Investors and regulators therefore need a contract map, not another adjusted-debt slogan. Companies should disclose project names, counterparties, maximum guaranteed amounts, minimum purchases, termination payments, repayment sources, equipment lives, consolidation judgments, and stress scenarios. Only then can formal debt be compared with obligations that apply the same pressure through a different pipe.