In partnership with

Years ago I spent a night on call watching a green dashboard while the product was down. Every panel healthy, no alerts firing. The dependency that broke had never been wired into the monitoring, so it emitted nothing and nothing looked wrong. It existed, at debug level, on a box nobody was watching.

I got a version of that feeling last week reading a Nikkei study that puts the off-the-books obligations of five US tech giants at $1.65 trillion. Alphabet, Microsoft, Amazon, Meta, Oracle. That's more than the $1.35 trillion they report as debt, and it has grown eightfold in four years. Meta carries about $420 billion of it, close to triple its reported debt.

They filed all of it.

Every one of these obligations is disclosed, down in the footnotes, filed with regulators, exactly where the rules say to put it. Data center leases that haven't started yet, chip supply agreements, server purchase commitments. Hidden implies somebody buried it. You can look all of this up. In March the Bank for International Settlements, the bank that central banks use, gave the category a better name: shadow borrowing, "obligations that are economically akin to debt but largely reside outside corporate balance sheets."

The rules written after Enron to drag exactly this into daylight are being followed, carefully. The obligations sit in the one place a stock screener and a debt ratio never reach. The error was in the logs the whole time. Nobody had an alert on it.

A four-year lease against a twenty-four-year bond.

In October, Meta and a group of funds run by Blue Owl set up a joint venture to build Hyperion, a 2,250-acre campus in Louisiana. Blue Owl put in 80%, Meta 20%. The venture borrowed $27.294 billion through a separate company, Beignet Investor LLC, built so its debts cannot reach Meta if it fails. Those bonds pay 6.581% and get paid down through 2049.

Meta leases the campus back from the venture it owns a fifth of, on an initial four-year term with options to renew. It also promised to cover the lenders for sixteen years if the campus ends up worth less than expected.

So the debt runs 24 years. The guarantee runs 16. The lease Meta records runs 4.

The lease accounting rule allows this on purpose: a renewal option only counts toward what you record if using it is "reasonably certain," which auditors read as 70% or better. Four years is what's certain. Everything past it is real money at a number nobody writes down. Across the big cloud builders, Moody's counts $662 billion of leases signed but not yet started, more than all the debt they do report.

One smaller detail bothers me more. Two weeks after those bonds priced, Meta sold $30 billion of its own ordinary bonds, nothing pledged against them, and drew $125 billion of orders. It could borrow whatever it wanted at a price that made bond desks emotional. It built the 80/20 structure anyway, at 6.581%, rated a notch lower.

Access to capital was never the constraint. So what was the wrapper for? Nobody adds a layer of indirection for free. When you find one in a codebase you don't admire it, you ask what it was optimizing for, and the answer usually lives in whatever it keeps from the caller.

If you own an index fund, you own this three ways.

Concentration first. The top ten companies are north of 40% of the S&P 500, so a broad market fund is a bet on four or five building programs. The fact sheet gives you a ticker and a weight.

Then somebody had to buy all that debt. Data center loans are now about 61% of a $79 billion market that bundles them into bonds and sells them on, and insurers are the heavy buyers, since the payments stretch as far as the policies they owe. The largest life insurers already hold close to a trillion dollars of private loans. Ed Zitron put it more bluntly than I would: "if these data centers do not pay out, we have retirement funds and insurance premiums that cannot get paid."

Banks lend to these ventures too, which that same central bank report warns opens fresh routes for a shock to travel.

Here's what bothers me about my own side. Public markets were supposed to be the transparent ones and private markets the murky ones. I work in the murky ones. When I look at a seed company I get the data room, the raw Stripe exports, and a founder who answers the same question twice if I didn't follow it the first time. A shareholder in the largest companies on earth gets a footnote and the hope that somebody read it.

Try the AI that knows your customers. No commitment.

Most platform evaluations start with a demo request and end three weeks later in a conference room. This one takes 15 minutes and puts you directly inside Gladly's interface — navigating it on your own terms.

See how AI surfaces real-time customer context before a conversation starts. Watch how a single conversation thread pulls in purchase history, channel history, and account details without a handoff.

No installation. No commitment. Start the interactive demo and see the platform for yourself.

Debt decides who pays when the building stops.

So does any of this make AI a bubble?

That's the wrong question to ask a debt structure. Overbuilding is a demand question, and this is all liability plumbing. Amazon, Microsoft, and Alphabet are sitting on something like $1.45 trillion of cloud contracts already signed.

What borrowed money changes is the blast radius. Dot-com was financed with stock, so when it broke shareholders watched the Nasdaq fall 78%. Painful, and exactly what stock is for. The 2001 telecom bust is the other template, debt raised against fiber that took a decade to light, and it traveled out through the lenders into the banking system.

Stock just sags. Debt has a due date, and the interest is owed whether or not anybody rented those chips that quarter.

Oracle is the live experiment. S&P cut its credit rating this month to one notch above junk, and the cost of insuring its debt against default just hit an eighteen-year high. It burned $23.7 billion of cash last fiscal year. Oracle's stock story and its credit story describe two different companies, and when those disagree this loudly I side with the desk that reads footnotes for a living.

The best pushback came from a rating agency. Moody's analyst David Gonzales says it is "not as if [the big cloud companies] have avoided a liability through structuring," and that the obligations are "yet to be on the balance sheet" rather than "missing." Fair, and the leases do land as they start. The narrower worry, from accounting consultant Tom Selling, is a company propping itself up with the accounting. Four of these five generate enough cash to make that a strange question. One doesn't.

I don't have a position to defend. My fund owns none of these five, and I've spent this year arguing with founders over what it costs to land one more customer.

My retirement account is a target-date fund, the set-it-and-forget-it default. I had never opened the holdings.

I opened them Thursday. All five are in there, and between the stock side and the bond side I own the shares and the debt. Then I read Meta's lease footnote for the first time. Eleven minutes. I have never given eleven minutes to anything I actually own.

Twelve years of being paid to be careful with other people's numbers. My biggest holding, I took on faith.

— SWEdonym

Reply and tell me: have you ever opened the holdings of your own retirement account?

Know a founder who'd want this? Refer a friend - rewards at 1, 3, 5, and 10.

New here? Grab the free guide: Top 10 Things SWEs Get Wrong About VC.

Reply

Avatar

or to participate

Keep Reading