Central bankers are careful with language. They don’t reach for dramatic phrases, because when they do, markets move. So when the Bank for International Settlements, the central bank for central banks, used the phrase “shadow borrowing” in a report this March1, it was worth stopping to read what they actually meant.
What they meant is this: a large and growing share of the money funding the AI boom doesn’t appear where you’d expect to find debt. It isn’t on the balance sheets of the companies that owe it. It isn’t in the regulated banking system. It isn’t marked at a price you can look up. It has been moved, deliberately, into the parts of the financial system that nobody watches closely.
And here’s the part that should hold your attention: the same move showed up in three different markets in the same few months. Not because they’re connected by accident, but because they share a single mechanic. Once you see it, you can’t unsee it.
This isn’t a doom prediction. It’s a description of where the risk went.
Shadow one: the debt that isn’t on the books
Start with the companies building the AI infrastructure — the “hyperscalers.” Amazon, Alphabet, Microsoft, Meta, Oracle. The handful of firms spending staggering sums on data centers, chips, and power.
For years, these were the most self-funded companies on earth. They threw off so much cash that they didn’t need to borrow. That changed. The buildout got so big that cash flow stopped covering it, and they started raising debt… more than $100 billion in bonds in 2025 alone, most of it locked in for over five years.
That part is visible. You can see the bonds. You can watch the warning sign that comes with them: the cost of insuring that debt against default has been climbing, especially for the lower-rated names. When the price of insurance goes up, the market is telling you it’s less sure it gets paid back.
But the bonds are only the part you can see. Alongside them, the hyperscalers built a quieter structure. Here’s how it works, in plain terms:
A separate shell company is created to own the data center. Outside investors put up the equity. The shell borrows the rest… privately, not in public markets. The tech giant takes only a small stake, then signs a long-term contract promising to rent the capacity for years, and sometimes quietly guarantees the debt if things go wrong.
Read that again, because it’s the whole trick. The tech company gets the data center it needs. The debt that paid for it sits on someone else’s books. On its own financial statements, what was a massive loan becomes a tidy line item: rent. A capital expense becomes an operating expense. The obligation is just as real, the company is still on the hook, but it has been moved somewhere it doesn’t show up as debt.
That’s what the BIS means by shadow borrowing: money that is economically debt, parked outside the borrower’s balance sheet. And who’s holding that debt? Increasingly, private credit funds and insurers… the same institutions we’re about to meet again.
Shadow two: lending to the companies AI might kill
Private credit is money lent by funds rather than banks. Over the last decade it exploded into a market north of $2 trillion, and a large slice of it went to a single kind of borrower: software companies. The subscription-software businesses, the ones that sell you a seat, a login, a monthly plan.
The numbers are stark. Loans from these funds to software firms went from about $8 billion in 2015 to over $500 billion by the end of 2025… roughly a fifth of all the direct lending these funds do. A third of private credit funds now have money out to the software sector.
Then AI arrived as a threat to exactly those businesses. If an AI tool can do what a software subscription used to do — cheaper, faster, without the seat — then the company that borrowed against those subscriptions has a problem. And so does the fund that lent to it.
You can watch this happening in the corner of the market that’s forced to report. Publicly traded versions of these lenders fell about 10% as software stocks collapsed nearly 30% between October and February. The ones with the most software exposure fell hardest. That tracks.
But here’s the deeper issue, and it’s the one the whole industry was built to avoid talking about.
Private credit's entire selling point was that it doesn't trade. That feature is now the danger.
The loans sit in a fund, valued at what the fund says they’re worth, “marked to model,” in the jargon, not marked to the market. No daily price means no scary headlines, no panic. That was the feature. It made private credit feel calmer and safer than the stock market.
That feature is now the danger. When the underlying borrowers are under threat but the loans are still valued at full price on paper, the gap between the official number and the real number is invisible… until something forces it into the open. And the industry has spent the last few years selling these funds to ordinary investors, who, unlike the big institutions, tend to ask for their money back when they get nervous. Some funds have already limited how much you can withdraw. When a fund puts up a gate, it’s telling you the calm was always partly an illusion.
Shadow three: the gold and silver blowup that wasn’t about gold and silver
Now the third market, which looks unrelated and isn’t.
Gold and silver ran hot through 2025 and into January 2026. Silver doubled over the year, then jumped another 50% in a single month. Then, in late January, it fell about 30% in one day… its worst day since the 1980s. Gold did a milder version of the same thing.
Nothing fundamental happened to silver in those 24 hours. No mine flooded. No factory stopped buying. So what broke?
Two things, and you should recognize the shape of both.
- First, who was buying. The money pouring in came mainly from ordinary retail investors, largely through exchange-traded funds. The big institutions weren’t chasing it… many were quietly trimming. When the buyers are mostly small and mostly late, the market gets fragile.
- Second, leverage. A lot of that retail money went into leveraged ETFs: funds that promise to amplify each day’s move. To keep that promise, they have to trade in the same direction as the price every single day. Prices up, they buy more. Prices down, they sell. That’s not investing; it’s a machine that pushes whatever’s already happening harder. On the way up it inflates the rally. On the way down it accelerates the crash.
The BIS found the market impact of this daily rebalancing roughly doubled over the course of 2025… the machine got bigger right before it mattered.
Add margin calls, forced selling when leveraged bets move against you, made worse when exchanges raise the required collateral mid-panic and you get a self-reinforcing spiral. Lower prices trigger selling, selling lowers prices, repeat.
There’s even an echo of the private-credit problem hiding in here. For a while, these ETFs traded above the value of the metal they held, because demand was so one-sided. When the mood flipped, that premium didn’t just vanish, for silver it inverted into a discount. The official value and the trading price came apart, fast, in the direction nobody wanted.
But first, someone has to build it
Every structure you just read about rests on one quiet assumption: that the data centers actually get built.
They might not, at least not all of them, and not on schedule. Across the country, towns and states have started saying no. Maine has frozen new large data centers until late 20272. New York’s legislature just passed a one-year pause3. Roughly a dozen states have similar bills in motion, and two members of Congress have proposed a national freeze on any facility that draws as much power as a small city4. In Seattle, developers walked away from projects after the local utility said it couldn’t power them without forcing big users to bring their own electricity.5
And you don’t need an outright ban for the trouble to start. A long enough delay does the same job. Here’s the mechanism, in plain terms.
Go back to the shell company from the first shadow… the one that owns the data center and borrowed the money to build it. That loan gets paid back out of rent: the long-term payments the tech company makes once the building is up and running.
The detail that matters is when those payments start. They start when the capacity is delivered. So if the building never gets delivered — blocked by a town, stalled in permitting, stuck for years in the line just to connect to the power grid — the rent never starts flowing. But the interest on the loan keeps running the whole time. A project that was supposed to generate cash quietly becomes one that only burns it.
So who eats the loss? It splits two ways, and both lead somewhere you’ve already been.
- If the tech giant promised to cover that loan no matter what, the debt it so carefully kept off its books comes flooding back onto it, at the worst possible moment, with the asset behind it frozen. The whole clever trick runs in reverse.
- If the tech giant’s promise came with fine print, and limiting how much they’re really on the hook for is usually the entire point, then the loss stays stranded in the shell company. Which means it lands on the private credit funds and insurers holding that debt.
And now you are standing in the second shadow, surrounded by funds that price their own loans and small investors lining up to pull their money out. Notice the pattern repeating… the risk doesn’t disappear. It slides to whoever was least able to say no.
But the real danger isn’t any single frozen project. It’s what happens to all of them at once. The moment “a town might block us” stops being a freak event and becomes an expectation, lenders stop judging each project on its own merits and start charging more for every data center loan everywhere… including the ones with no problem at all. A single county board voting for a six-month pause doesn’t just stop one building. It quietly raises the price of the entire buildout.
And this is where it loops back to the machines themselves. The grand bet of the AI era — the borrowing on one side, the AI revenue on the other — was always one bet wearing two hats. You borrow to build the data centers. The data centers are how you serve the AI demand that’s supposed to pay back the borrowing. Block the building and you sever both halves with a single cut: the loan goes bad and the revenue that was meant to justify it never arrives.
Power and permits turn out to be the one bolt holding the whole machine together.
Which exposes the single risk none of these structures was built to survive. They were designed by people who are very good at pricing interest rates, defaults, and demand. They were not designed for a county commissioner, a state senator, or a neighbor who’s tired of watching their electricity bill climb to cool someone else’s servers. You cannot buy insurance against a town saying no. You cannot spread that risk across a hundred projects when the opposition is turning up in red states and blue states alike.
The most sophisticated financial engineering of the decade has a blind spot shaped exactly like a zoning board.
One pattern, three masks
Step back and the three stories collapse into one.
In each case, risk that used to live somewhere with guardrails — a regulated bank, a company’s own balance sheet, a price that updates in daylight — was relocated to somewhere without them. An off-balance-sheet shell. A fund that marks its own homework. A product that amplifies instead of cushions.
And in each case, the money that ended up holding the risk at the bottom was increasingly ordinary money. Retail investors in the private credit funds. Retail investors in the leveraged ETFs. And through your index fund, retail exposure to the handful of AI giants that now make up about a third of the entire S&P 500.
Call it the last holder problem. Every one of these structures works beautifully until it doesn’t, and when it doesn’t, the question is simply: who’s holding the asset when the music stops? The structures of the last few years have quietly arranged for that person to be the smallest, slowest, least-informed participant in the chain. Not by accident.
The opacity is the product.
The whole appeal of shadow borrowing, model-based marks, and amplified ETFs is that they feel calmer than the real thing… right up until the calm is the first thing to break.
The BIS isn’t predicting a crash. Neither am I. What they did, and it’s rarer than it sounds, was point at three separate fires and say, quietly, notice that these are all the same fire. Risk didn’t shrink over the past few years. It moved. It moved to where the rules don’t reach and the prices don’t update and the headlines don’t form.
So what tips these structures from a quiet arrangement into a loud event?
Probably not one dramatic catalyst. More likely the plain arithmetic of higher-for-longer interest rates — every month a blocked project sits idle, the unpaid interest compounds — colliding with the refinancing dates that finally drag the made-up valuations into daylight, while the moratorium wave turns "might be delayed" into "is delayed." That's the genuinely unsettling part. This is a risk that grinds rather than snaps, and there may be no single morning when a headline tells you it has started.
That’s the thing about a shadow. It doesn’t mean nothing’s there.
It means the light hasn’t reached it yet.