Five Tech Giants Are Hiding $1.65 Trillion in AI Debt Off Their Balance Sheets. That's the Ground Your Token Prices Are Standing On.

A Nikkei investigation published this week put a number on something a lot of people had been feeling in their gut: Alphabet, Amazon, Meta, Microsoft and Oracle are carrying an estimated $1.65 trillion in off-balance-sheet obligations tied to AI, on top of the roughly $1.35 trillion of debt they actually report. The liabilities investors can't see in a normal debt-to-equity ratio are now larger than the ones they can. Meta alone accounts for around $420 billion of the hidden pile: nearly three times its visible debt.

You don't own any of that debt. But your token prices are sitting on top of it, and it's worth understanding why.

What "off-balance-sheet" actually means here

The trick isn't fraud, and calling it Enron (which several outlets have) is doing a lot of work. Here's the mechanic. Instead of borrowing a pile of money to build a data center and putting that loan on the balance sheet where everyone can see it, the hyperscalers increasingly sign long-term leases on the data centers, take-or-pay commitments on GPU supply, and joint ventures with private-credit funds. Each of those locks in years of future cash outflows. None of them shows up as a formal debt line today.

So when you read that Microsoft or Alphabet has a manageable debt load, that's true in the narrow accounting sense and misleading in the way that matters. Nikkei's figure (up roughly eightfold in four years) says the real forward obligation is close to double what standard credit models capture. The combined visible-plus-hidden number lands near $3 trillion.

I'm not writing this to short anyone's stock. I'm writing it because the entities in that paragraph are the ones you rent inference from, and their financing structure is the actual foundation your unit economics rest on.

Why this reaches a solo operator at all

Trace your dependency chain honestly. Your product calls a model API. That API runs on a lab's infrastructure. That infrastructure runs, more and more, on data-center capacity these five companies are financing with commitments that don't appear as debt. Every token you resell inside your app is priced today at a level those companies are subsidizing to win market share: a level that assumes the capital keeps flowing on the current terms.

The thing about future cash commitments is that they're rigid. A lease is due whether or not the AI revenue showed up on schedule. A GPU take-or-pay contract gets paid whether or not you filled the capacity. If the AI revenue curve flattens even for a few quarters, the obligations don't flatten with it, and the first lever a CFO under that kind of pressure reaches for is the price of the thing customers are currently getting cheap. That thing is your inference.

This is the same lesson every operator eventually learns about any subsidized input. Cheap is a strategy someone else is funding, and strategies end. You don't have to believe in an imminent crash to take the point seriously. You just have to notice that your margins quietly assume a price that a lot of leveraged balance sheets are working hard to hold in place.

The honest counter-take

Now the other side, because a post that only rings the alarm is a bad post.

Off-balance-sheet leasing is not inherently dishonest, and these are not Enron. Enron hid obligations to fake profits it didn't have. Alphabet, Microsoft and Amazon have enormous, real, cash-generating businesses underneath the AI spend: search, cloud, ads, retail. The leases fund assets that mostly get used. A structure that spreads a data-center payment over fifteen years instead of parking a bond on the balance sheet is normal corporate finance, not a smoking gun.

And these companies can service a lot. If the AI bet pays even part of what they're projecting, $1.65 trillion of forward commitments against several trillion in market cap and hundreds of billions in annual cash flow is aggressive, not reckless. The genuinely bad scenario isn't "they go bankrupt." It's milder and more likely: capex discipline arrives, subsidies on inference get trimmed to protect margins, and the price you pay per million tokens stops falling (or ticks up) right when you'd budgeted for it to keep dropping.

That's not a doomsday event. It's a Tuesday, and it's the version worth planning for.

What I'd actually do

Nothing dramatic, which is the point. Panic-selling your AI product because five companies use aggressive lease accounting would be silly. But there's one honest exercise this story earns.

Open your cost model and find the line where you assumed inference keeps getting cheaper. A lot of solo AI products are quietly banking on the last two years' price trend continuing: margins that only work because the token cost in eighteen months is projected well below today's. Ask what your business looks like if that curve simply flattens at current prices. If the answer is "still fine," great, you've got a real margin. If the answer is "underwater," you don't have a business yet, you have a bet on someone else's capex generosity.

Then do the boring hedge, the same one that covers half a dozen other AI-stack risks: keep your model calls behind a single swappable interface so provider and price are a config change, not a rewrite. If one vendor gets financially disciplined at your expense, you move to whoever's still competing for share.

What I wouldn't do is treat a Nikkei headline as a reason to rearchitect anything today. The hidden debt is real, the risk is slow, and the correct response is to stop assuming free money you never had a contract for, not to sprint away from a fire that hasn't started.

Author

Lukas

@lukcombinator

Sources

• Five US tech giants' hidden debts soar to $1.65tn on opaque AI funding (Nikkei Asia)
• Concerns grow over AI giants' hidden debts (Semafor)
• Five Tech Giants Are Using Enron's Accounting Strategy to Conceal $1.65 Trillion in AI Debt (Yahoo Finance)
• Big Tech is hiding $1.65tn in off-balance-sheet AI debt (The Next Web)