Big Tech's $1.65 Trillion Off-Balance-Sheet AI

Here's a number worth sitting with: $1.65 trillion. That's not what Alphabet, Microsoft, Amazon, Meta, and Oracle owe in the traditional sense. It's what they owe in the sense that accounting lets you not mention on the first page.

A study by Nikkei Asia, analyzing recent financial statements and related documents from all five companies, found that their AI-related contractual obligations total approximately $1.65 trillion—and that this figure exceeds the companies' actual reported debt. The Seoul Economic Daily puts the Korean-won equivalent at roughly 2,455 trillion won, which doesn't make it feel smaller. Per Phemex News, this represents an eightfold increase over roughly four years.

Eightfold. In four years. That's not a trend line—that's a cliff face.

What "Hidden" Actually Means Here

Let's be precise, because "hidden debt" sounds like something from a financial crime drama, and the reality is both less dramatic and more interesting.

Off-balance-sheet obligations are, broadly, real financial commitments that don't appear as liabilities on a company's balance sheet under standard accounting rules. Think long-term purchase agreements, operating leases before updated GAAP rules pulled them onto balance sheets, and crucially: multi-year supply and capacity contracts. When a hyperscaler locks in years of GPU supply from Nvidia, or commits to decades of power purchase agreements to feed a data center, those forward commitments may not show up as "debt" in the headline number—but they're money owed, full stop.

The Hacker News thread on the Nikkei piece surfaces an illuminating framing: someone invokes "off-balance-sheet financing 101" from the 1980s. That's not a throwaway snark. The 1980s were the golden era of creative corporate financing structures—special purpose vehicles, synthetic leases, partnerships designed to keep obligations off the books. Enron perfected this into an art form. The parallel isn't that these five companies are doing anything fraudulent—there's no evidence of that—but the mechanics rhyme. Large obligations, real economic exposure, structured to minimize what appears on the face of the balance sheet.

The difference is that today's rules require extensive footnote disclosure of these commitments. They're not secret in the legal sense. They're just buried deep enough that most retail investors and tech watchers skip past them. AI Weekly frames it well: "The single number in Nikkei's new study is small enough to read past, and big enough to matter."

That's the crux. Technically visible. Practically invisible.

Why the AI Buildout Generates This Specific Kind of Debt

To understand why AI spending produces off-balance-sheet exposure at this scale, you have to understand what AI infrastructure actually costs—and how you pay for it.

Training frontier models requires enormous compute clusters: thousands of high-end GPUs running for months. But training is a one-time cost per model generation. What makes the economics stranger is inference—actually running the models at scale for users. Inference is ongoing, and its cost scales with usage in ways that are hard to forecast. To lock in supply and pricing, companies sign multi-year agreements with chip manufacturers, cloud infrastructure providers, and power suppliers.

Add to that the data center construction pipeline. Microsoft, Google, Amazon, and Meta have all announced data center investment plans totaling hundreds of billions over multiple years. Oracle, somewhat surprisingly, has emerged as a major AI infrastructure player—its cloud pivot under Larry Ellison has been aggressive enough to land it in this particular five-company club. These aren't expenses; they're commitments. Long-dated, binding, and—until they hit the income statement—living in footnotes.

The Next Web notes that the off-balance-sheet figure now exceeds the companies' actual on-balance-sheet debt. That's the structural oddity worth flagging: the obligations these companies chose not to call debt are now bigger than the obligations they did. At some point the distinction between "debt" and "contractual obligation that is economically indistinguishable from debt" stops being meaningful.

Is This Actually Dangerous?

Here's where I want to resist the easy narrative in both directions.

The bearish read: $1.65 trillion in obligations tied to a technology buildout that hasn't yet proven it can generate returns commensurate with its cost is a real risk. AI revenue exists—Microsoft Copilot subscriptions, Google's AI search monetization, AWS and Azure AI services—but the aggregate revenue from AI products is nowhere near $1.65 trillion. The gap between commitment and return is being funded by cash flows from legacy businesses: search advertising, cloud compute, e-commerce, social advertising. If any of those core businesses softens while the AI obligations remain fixed, the math gets uncomfortable fast.

The bullish counterpoint isn't crazy either: these are five of the most cash-generative companies in the history of capitalism. Alphabet and Microsoft have historically carried manageable debt-to-cash ratios. Amazon generates cash at industrial scale. The argument that they can't service $1.65 trillion in contractual obligations over multiple years is not obviously correct.

What's harder to dismiss is the opacity argument. When off-balance-sheet obligations exceed on-balance-sheet debt, markets are pricing these companies with incomplete information—not because anything is being hidden illegally, but because the standard headline metrics don't capture the full picture. Analysts who dig into footnotes know this. Most retail investors don't.

The dot-com parallel the brief gestures at is real but needs calibration. The dot-com bubble involved companies with no revenue and no path to it. These five companies have enormous, profitable core businesses. The AI spending isn't a lottery ticket; it's an infrastructure arms race where falling behind may genuinely cost market position. The question isn't whether AI is worth building—it's whether the specific bets being made, at this specific scale, at these specific price points, will generate returns that justify the commitment structure.

That question doesn't have an answer yet. Anyone who tells you it does is selling something.

The Transparency Problem Is the Actual Story

What Nikkei's study surfaces, underneath the headline number, is a disclosure architecture that lags the economic reality of how AI is being financed.

Current accounting standards weren't designed for a world where a company's most consequential financial exposure takes the form of GPU purchase agreements and power contracts with 15-year terms. The rules allow—sometimes require—these to live outside the headline balance sheet. That's not a conspiracy; it's a mismatch between accounting frameworks built for an industrial economy and a capital deployment pattern that looks more like a military procurement cycle than a traditional tech R&D budget.

Regulators and standard-setters will eventually catch up. The 2016 update to lease accounting (ASC 842 in the US) that pulled operating leases onto balance sheets was exactly this kind of catch-up. A similar reckoning for long-term AI infrastructure commitments seems likely, eventually.

Until then, the most honest thing you can say is this: the balance sheets of the five most important AI companies in the world currently understate their economic exposure by at least $1.65 trillion, and the gap has grown eightfold in four years. Whether that exposure is wise, reckless, or somewhere in between depends entirely on whether the AI economy develops at the pace these companies are betting it will.

If it does, these commitments look like prescient infrastructure investment—the equivalent of buying fiber in 2001 before broadband demand caught up. If it doesn't, the footnotes are going to get a lot more attention than they're getting now.

Tyler Nakamura is a Consumer Tech & Gadgets Correspondent for BuzzRAG.