Let’s talk about the GPU your product runs on.
It isn’t yours. It isn’t really your vendor’s either. It’s the thing a lender takes if the vendor stops paying.
That’s what collateral means. The asset the bank seizes when the loan goes bad.
Software didn’t used to have collateral. Your average SaaS vendor had a lease, some servers, a few hundred salaries, and a renewal email once a year. Now the thing under your API is a rack of chips that somebody borrowed against, and the loan is priced on the assumption that those chips keep earning for six years.
Picture a founder in October. He runs a support product. Four thousand tickets a day go through two models, and one of them sits on a neocloud because it was forty percent cheaper than the big cloud and twice as fast. His rep offers a three-year committed spend at a further discount. Sign now, lock the rate, and the unit economics finally work on the spreadsheet.
He signs. Of course he signs.
What he didn’t ask, because nobody asks, is why the discount existed. Not because the chips got cheaper. Because the vendor needed three years of contracted revenue on the books. Contracted revenue is what makes a loan bankable. His signature went into somebody’s collateral package, and he will never see that document.
That’s the pattern. Not a crash story. A structure story.
Oracle’s five-year credit default swaps — that’s insurance against the company not paying its debt — sit at 295 basis points, up 287 percent in a year. S&P rates it BBB-, the bottom of investment grade, after a downgrade in July. It’s carrying $118 billion of long-term debt and another $39 billion in capital and lease obligations. Consensus forecasts put free cash flow at minus $23.7 billion this year and minus $36 billion next. And it’s the main builder behind a $300 billion commitment to OpenAI.
CoreWeave is the cleaner example. One of its notes, 9.625 percent due 2032, has traded under eighty cents on the dollar at a yield above fifteen. That isn’t a company in trouble. That’s a company the bond market has started pricing like it might be.
Now follow the money back, because where it came from explains how it behaves.
Most of this debt isn’t in banks. It’s in private credit — Blackstone, Blue Owl, Apollo, Pimco, BlackRock. Outstanding lending to AI-related companies went from close to zero to over $200 billion in a few years, and Morgan Stanley projects another $800 billion of data-centre financing over the next two. It gets routed through special purpose vehicles, which is a polite term for a company that exists to hold one building and one loan. Meta’s Hyperion deal in Louisiana was $30 billion: a vehicle called Beignet Investor, Meta owning twenty percent, Blue Owl owning eighty. More than $120 billion of data-centre spending moved off balance sheets in about eighteen months.
A February study from the Chicago Fed found bank exposure to AI-adjacent industries at 0.8 percent of total assets. Then it added the sentence that matters: banks most likely have more, through lending to the non-bank lenders. Which means the other end of this trade is somebody’s insurance float and somebody’s pension. Not a villain. Just a person who never chose it.
Here’s the mismatch underneath all of it. These chips get depreciated on the books over five to six years. Their competitive life for frontier work runs closer to two or three. Take the 2026 server cohort — roughly $425 billion. At six years it costs about $71 billion a year to carry. At three, it’s $142 billion. Michael Burry’s number is that the big buyers understate depreciation by $176 billion across 2026 to 2028.
Microsoft’s CFO said something in January that ties the whole thing together. Two-thirds of the capex is short-lived assets, mostly GPUs, depreciated over six years, and the capacity is — his words — already sold for the entirety of their useful life.
Already sold. That is the pre-sale. That is the founder’s signature. The six-year book life only holds if somebody has promised to pay for years four, five and six in advance, and the thing that promises is a contract with a customer who has less information than the lender.
So watch the chain of how this reaches a person who has never bought a bond in his life. GPU resale value falls. The tenant can’t service the loan. The tenant misses a lease payment to the building. The building’s cash flow breaks. The asset-backed paper behind the building breaks. Nobody’s model got worse. The machine didn’t fail. The financing did.
Three things people are getting wrong about this.
The first is that they’re reading the credit stress as a story about investors. It’s a story about product. A vendor with a 2028 maturity needs renewals more than it needs your affection, and that shows up as pricing, capacity allocation, and the shape of every “commit to save” offer you get this quarter. The balance sheet is now upstream of the roadmap.
The second is that cheap tokens get treated as an engineering achievement. Some of it is. Some of it is a financing decision — a subsidy with an expiry date, and the expiry date is on a credit calendar, not a product calendar. You will not get a warning email. You’ll get a new rate card.
The third, and the ugly one: the loss doesn’t land on the people who underwrote it. It lands on whoever signed the longest contract at the best rate, and on the pension fund holding the paper. Same trade, two ends, and the customer is the end with no seat at the table.
So what do you actually do, on Monday.
Find out who owns the metal and who financed it. Ask your account manager where the GPUs sit and when the debt matures. A vendor who won’t say is telling you something.
Keep every commitment shorter than your ability to switch. If your own testing says you could move in a quarter, don’t sign for three years to save fifteen percent. The discount is the price of your options, and your options were the thing worth having.
Get a credit-event clause into the contract. One you’ve actually read, not one you assume is standard.
Run one real workload on a second provider every month, small, so the switch is a reflex and not a project. Keep your prompts and your test set in your own repository, because that is the only part of the migration that’s genuinely hard.
And price your own margin at twice and three times today’s inference cost. If the business only works at the subsidised rate, then the business is a bet on somebody else’s refinancing, and you are not the one who gets to make that call.
The principle here isn’t complicated. When a machine gets built on borrowed money, the promise at the bottom of the stack isn’t a feature roadmap. It’s an obligation, and obligations always get handed down until they reach somebody who can’t refuse.
You should be able to see who’s holding the other end of that promise. And you should be able to read the clause that says what happens on the day they can’t.