Big Tech Hides $300B in AI Liabilities Through Off-Balance-Sheet Guarantees

The accounting machinery that underpins investor confidence in Big Tech is quietly concealing roughly $300 billion in artificial intelligence liabilities through a mechanism most shareholders never examine: financial guarantees that sit nowhere on a balance sheet.

These guarantees function as a kind of financial sleight of hand. When a major technology company commits to backing a loan or obligation—say, a guarantee on debt issued by a subsidiary or partner—that commitment doesn't appear as a liability on the parent company's financial statements. It lives in footnotes, in disclosures buried deep in SEC filings, in the fine print that most investors never reach. Yet the obligation is real. If the underlying loan defaults, the guarantor must pay. The risk is genuine; the visibility is not.

The scale of what's being hidden has begun to alarm financial analysts. Michael Burry, the investor known for his prescient warnings before the 2008 financial crisis, has flagged that hyperscalers—the massive cloud and AI infrastructure companies—may be concealing as much as $3 trillion in total AI-related liabilities through these off-balance-sheet arrangements. The $300 billion figure represents what has been documented so far, but the true exposure could be substantially larger. Wall Street, Burry and others argue, is not adequately interrogating these arrangements. The financial system is pricing in the risks it can see; it may be blind to the ones it cannot.

The concern is not merely academic. These guarantees represent real money that could be demanded from these companies if underlying obligations fail. In a sector where capital expenditures are already staggering—hyperscalers are spending tens of billions annually on AI infrastructure—hidden liabilities of this magnitude could materially alter how investors should value these firms. A company that appears to have $50 billion in debt might actually have exposure to $80 billion or more once you account for what's been tucked away in guarantees.

The market is beginning to sense something is amiss. Credit default swap spreads—the cost of insuring against default on corporate debt—have been widening for hyperscalers, suggesting that traders and institutional investors are pricing in risks that traditional balance sheet analysis does not capture. When CDS spreads widen, it signals that sophisticated market participants believe the actual risk of default is higher than what the published financials suggest. It is a canary in the coal mine, a warning that the gap between what companies are disclosing and what they actually owe may be larger than anyone wants to admit.

The structural problem is that accounting rules, written before the AI era transformed corporate finance, allow companies significant discretion in how they treat guarantees. A guarantee that is deemed "remote" in likelihood of being called can be disclosed minimally. A guarantee that is deemed more probable might require a reserve, but even then, the accounting treatment can be conservative or aggressive depending on how a company's auditors interpret the rules. In an industry moving as fast as AI infrastructure, where demand forecasts shift monthly and partnerships are constantly being renegotiated, the probability of a guarantee being called is genuinely hard to assess. Companies have incentive to assess it conservatively. Investors have no way to independently verify.

What happens next depends partly on whether regulators and auditors begin demanding greater transparency, and partly on whether the market's growing skepticism forces companies to disclose more. For now, the $300 billion sits in the shadows, a liability that is real but not quite visible, a risk that is priced into some trades but not into the valuations most retail investors rely on. The music, as Burry suggested, is still playing. But the question of what happens when it stops—when these guarantees are called, when the true scale of obligation becomes undeniable—remains unanswered.