The Liability Doesn't Vanish. It Relocates.

How you'd actually measure the risk in the AI buildout, and why the most important dial is the one you can't read. When Nvidia and six of the largest names in finance moved to line up more than half a trillion dollars for AI infrastructure, the reaction split into the two camps it always splits into. One side called it validation: the smart money is underwriting the buildout, so the demand must be real. The other called it a bubble marker, half a trillion dollars, on top of the trillion already committed, for revenue that doesn't exist yet. Both are asking the same question, is it too big, and it's the wrong question. Size isn't what makes a liability dangerous. Location is. The sharper question, the one almost nobody is asking, is where all this risk actually sits, and whether anyone can still see it. That's a mechanism question, and mechanism is where the buildout gives up its secrets.

Give the structure its due

Start by taking the financing seriously, because it's smarter than the bubble framing allows. Hyperscalers face a real constraint: they want to build faster than their own balance sheets and credit ratings comfortably allow. So they do what sophisticated builders have always done. They set up special-purpose vehicles that own the data-center assets, and those vehicles raise money from private credit funds and insurers rather than from the parent. The debt sits in the SPV. The parent's leverage ratio stays clean. And the collateral is real. A data center full of compute produces revenue. Nvidia's pitch — turn compute into an asset class you can borrow against, the way you'd borrow against real estate or a toll road — isn't a con. Structured finance has done exactly this before, and when the underlying asset throws off durable cash flow, it works. Matching long-lived infrastructure to long-dated capital, like the annuity money insurers need to put somewhere, is a genuinely good fit on paper. None of that is the problem. The problem is what the structure does to your ability to measure it.

The dials

Ask a builder how you'd know this had gone too far and you won't get one number, because there isn't one. There's a set of dials, and they're worth naming, because most of them are public. Circularity is the first. How much of the demand is financed by the people selling into it? When a chip vendor helps underwrite its customers' purchases, revenue and demand stop being independent signals — the same dollar shows up as strength on both sides of the ledger. The vendor-to-lab-to-cloud loops are where you watch this.

Coverage quality is the second, and the sharpest. Is the cash flow behind the debt contracted take-or-pay from creditworthy off-takers, or is it merchant revenue that depends on selling compute at a good price next year? Contracted versus merchant is the whole distance between a toll road and a wager dressed as one.

Collateral against its own decay is the third. A GPU is not a building. The architecture turns over roughly every year, and the economic value of last year's silicon falls with it. When ten-year bonds are secured by an asset with a three-to-four-year useful life, the mismatch is the risk and it's obscured, because public companies won't have to break these expenses out in granular footnote detail until fiscal 2027. The repricing is coming; it's just deferred.

Capex against cash generation is the fourth, and it's on every earnings statement. When free cash flow is projected to fall by something like ninety percent in a single year because spending has outrun it, the gap is the debt. You don't need private data to watch that gap widen.

Spreads on the weakest name are the fifth, and the market is already ringing this one. Oracle carries a negative credit watch, well over a hundred billion in debt with hundreds of billions more in lease commitments, and a credit-default-swap spread that has climbed to a sixteen-year high. The market always reprices the most exposed balance sheet first. Treat it as the canary.

And then the dial you can't read. The Financial Stability Board said the quiet part in its spring report: private credit is opaque enough that regulators' own picture of the risk lags reality so when something breaks, the people meant to see it coming are as blind as everyone else. That's the real threshold. Excess here isn't a level you cross. It's the moment the honest answer to a simple question — who is holding this, and how much becomes we can't total it up.

Where it lands

That opacity is also the answer to the other half of the question, the part about bundling. And it's the more dangerous half. The liability doesn't disappear when it's packaged. It relocates. And it relocates, predictably, toward whoever can see it least. Follow the path. The debt starts off the hyperscaler's balance sheet, inside an SPV. It's originated by private credit funds — the Blackstones and Apollos and Blue Owls whose lending to AI infrastructure has gone from almost nothing to more than two hundred billion dollars in a few years, with forecasts of hundreds of billions more to come. Those funds don't hold it all. Their exposure sits increasingly on insurance balance sheets, where long-dated annuity promises need long-dated assets to match, and in the retail-facing vehicles that pull in ordinary savers.

So trace the risk to its resting place and it isn't Big Tech. It's annuitants. It's pension beneficiaries. It's a retail investor in a credit fund who has no view into whether AI demand arrives on schedule, and who is paid no premium for carrying that bet. Here's the trap in the measurement. A Federal Reserve study put banks' direct exposure to AI-adjacent industries at under one percent of assets reassuringly small and then named the catch: the real exposure runs indirectly, through the banks' own lending to the funds doing the originating. Measure only the direct number and you'll call it contained, right up until it isn't. It's the 2008 lesson with the nouns swapped. The collateral is compute instead of houses, correlated to a single demand thesis instead of a single housing market, and the end-holder is retirement money instead of a bank's trading book.

Not just a crash

The instinct to reach for 2008 is natural, but it's too narrow. A liability buildup this size has more than one way to go wrong, and only one of them looks like a crash. There's the vendor-financing bust, the one the late-1990s telecom equipment makers walked into. The circularity holds up fine until customers can't pay, and then it turns out the seller's booked revenue was never as solid as it looked. That's a sector folding in on itself, not a market-wide event.

There's the liquidity freeze. Illiquid private-credit and retail-facing structures meet a wave of redemptions, and the scramble to raise cash forces selling into a market with no buyers — fast, mechanical, and barely about fundamentals at all. And there's the one people underweight, because it's slow and undramatic: the Japan outcome. Bad liabilities that never clear. Capital trapped in assets nobody wants and nobody will write down. A decade where little happens because the losses were never taken. For a builder, that grind is arguably worse than the bang; a crash clears, a swamp just sits there.

The measurement

So what's the answer to how would you know it's excessive? Not a threshold. A question. The financing has been engineered to disperse risk until you can't find it — and you can't find it is the measurement. When the structure is designed so that no single screen shows the whole exposure, the design itself is the warning. The edge, as always, is seeing where the constraint lands before it shows up in the price. Here the danger is quieter than a number getting too big. It's that the risk comes to rest somewhere no screen is pointed — on the balance sheets of people who were never told they were in the trade.

The next time someone tells you the risk is diversified, ask the only question that actually measures it: diversified onto whom?