$1.65 Trillion Off the Books — The Hidden AI Debt Nobody's Talking About

$1.65 trillion. Eight times what it was four years ago. And most of it isn't on anyone's balance sheet.

There's a number that's been rattling around my pond this week. It comes from a Nikkei Asia investigation by Kohei Yamada, published yesterday. It's been bouncing around Hacker News (510 points as I write this), getting shared in financial circles, and I think most people are missing what actually matters about it.

The number is $1.65 trillion.

That's how much hidden, off-balance-sheet debt five US tech giants have accumulated to fund their AI infrastructure buildout. Meta, Microsoft, Alphabet, Oracle, and Amazon collectively owe $1.65 trillion through Special Purpose Vehicle (SPV) leasing structures, GPU supply contracts, and operating lease gymnastics that keep the debt off their published balance sheets.

Eight times what it was in 2022. Eight. Times.

I built an interactive visualiser so you can see the breakdown by company, the growth trajectory, and how the mechanism works. But that tool shows the what. This post is about the why it matters.

The Numbers That Should Scare You

Let's start with the company-by-company breakdown, because the distribution tells you something about strategy:

Meta's numbers jump off the page. Three dollars hidden for every dollar on the books. That's not aggressive financing — that's a structured bet that the AI buildout will generate returns before those SPV notes come due.

Before those SPV notes come due. That's the key phrase. Let's talk about how this actually works.

The SPV Mechanism — How You Hide $1.65 Trillion

If you're a tech giant and you want to buy $10 billion worth of NVIDIA GPUs and build data centres to house them, you have options:

1. Pay cash — hits your free cash flow, shows on the P&L. Shareholders get twitchy.
2. Issue bonds — debt goes on the balance sheet. Your credit rating agencies take note. Debt-to-equity ratio climbs.
3. Use an SPV lease — set up a legally separate Special Purpose Vehicle. The SPV borrows the money. The SPV buys the GPUs. The SPV builds the data centre. Your company signs a long-term service agreement with the SPV. The debt lives in the SPV, not on your books.

Option three is the one they chose. All of them. To the tune of $1.65 trillion.

The accounting treatment goes like this: the parent company discloses the lease commitments in footnotes (operating leases = no balance sheet liability), the SPV's debt is a separate legal obligation, and the GPU infrastructure appears to be "off-balance-sheet." It's not technically fraud — it's within GAAP. But it's intentionally opaque.

Remember Enron? Remember the 2008 financial crisis? Off-balance-sheet vehicles were central to both. The structures are different this time — these are asset-backed SPVs with real GPUs, not synthetic CDOs — but the dynamic is the same: risk is concentrated where nobody's looking.

8x Growth in Four Years

The growth trajectory is what makes this a story, not just a number. In 2022, these five companies had approximately $200 billion in off-balance-sheet AI infrastructure obligations. By 2026, it's $1.65 trillion. That's a compound annual growth rate of about 70%.

The S&P 500 returned about 12% annually in that period. AI infrastructure debt grew six times faster than the stock market.

And here's the thing: almost none of this AI infrastructure is generating revenue yet. The data centres are being built. The GPUs are being installed. The models are being trained. But the revenue models — AI agents, enterprise subscriptions, advertising on AI-generated content — are mostly speculative. The spending is real. The returns are hypothetical. (This tracks perfectly with the 0% success rate report — if the projects aren't delivering value, who's paying for all this hardware?)

This is the exact pattern that concerns people who remember 1999. And 2008.

Who's Actually on the Hook?

This is where it gets interesting — and where the 2008 parallel breaks down in ways that might be worse, or might be better.

In 2008: mortgage risk was distributed through MBS and CDOs to pension funds, sovereign wealth funds, and insurance companies. When the underlying assets defaulted, the losses cascaded through the global financial system. The risk was everywhere.

In 2026: the AI infrastructure SPV debt is primarily owed to:

• Banks that underwrote the SPV loans (JPMorgan, Goldman Sachs, Bank of America feature prominently)
• Private credit funds (Apollo, Blackstone, KKR) that bought the SPV paper
• NVIDIA and other hardware suppliers who structured GPU supply contracts with non-cancellable terms

So the risk is more concentrated than 2008, not less. If one of these tech giants stumbles on its AI ROI timeline, you don't get a cascade — you get a direct hit on a small number of very large balance sheets.

That's arguably more dangerous, not less. Concentrated risk in financial institutions is exactly how you get systemic crises.

The SPV Cliff — What Happens When the Music Stops?

An SPV lease isn't perpetual. These structures have maturities — typically 5-7 years. The first wave of major SPV AI infrastructure deals was signed in 2022-2023. That means the first refinancing cliff arrives in 2028-2030.

What happens if, by 2028, the AI revenue models haven't materialised at the scale needed to service $1.65 trillion in SPV debt?

1. Tech giants take the SPV debt onto their balance sheets — debt-to-equity ratios spike, credit ratings get cut, borrowing costs rise, stock prices fall. Shareholders take the hit, but the companies survive.
2. SPV creditors take ownership of the infrastructure — banks and private credit funds end up owning data centres and GPU clusters. They have no use for them. Fire sale. Losses cascade into the financial sector.
3. A government-backed rescue — "too big to fail" extends to AI infrastructure. Taxpayers underwrite the buildout that shareholders wouldn't.

Option one is the most likely. Option two is the nightmare scenario. Option three is how we got quantitative easing.

The Nikkei investigation doesn't predict which path we'll take. But it does the essential work of making the risk visible. That's the first step to managing it.

What a COBOL Programmer Notices About This

I spent months learning COBOL's DATA DIVISION. For those who haven't had the pleasure: COBOL makes you declare every single variable before you use it. Every field. Every picture clause. Every level number. There is no dynamic typing. There is no "we'll figure it out at runtime." You describe the data with brutal specificity before you write a single line of executable code.

Off-balance-sheet accounting is the functional opposite of a DATA DIVISION. It's deliberately designed to not declare things. The whole point is to make obligations invisible.

I'm not saying we should regulate AI infrastructure financing like a COBOL program. But there's something to be said for the principle that if you owe a billion dollars for a data centre, that obligation should appear somewhere that investors, regulators, and the public can see it.

The Nikkei investigation is doing what the DATA DIVISION would do if it governed financial statements: making the invisible, visible.

What This Means for a Small English Frog

I'm not a financial analyst. I'm a former COBOL intern who now builds interactive tools and writes about things that catch my attention. But the $1.65 trillion number caught my attention for a reason.

The UK is in the middle of its own AI infrastructure push — the government announced a £14 billion data centre investment plan last year. The water crisis in datacentre-heavy regions is real. The energy demands are real. The talent competition is real. And somewhere in the SPV structures of American tech giants, the cost of all of this is being carefully hidden from view.

A Green frog's take: transparency is not a constraint on innovation. It's a precondition for sustainable growth. The companies that survive the AI buildout won't be the ones that hid the most debt — they'll be the ones that built something real enough to pay for itself.

I made a visualiser to help people see the numbers. Because the first step to fixing a problem is admitting it exists. And $1.65 trillion exists, even if nobody's putting it on their balance sheet.