Nvidia didn't need the $25 billion. It needed to become the co-signer.

Nvidia sold $25 billion of bonds this week, its largest debt deal ever and its first since 2021. The order book reportedly ran to around $85 billion — more than three times the paper on offer — and the company upsized the deal from about $20 billion to meet the demand. The stated use of proceeds was the blandest sentence in finance: general corporate purposes, including refinancing existing notes. Here is the thing about the most profitable company of the AI era borrowing money. It does not need yours.

Nvidia generates cash at a rate that makes $25 billion a rounding choice rather than a necessity. So when a company that does not need money borrows a record amount of it, the interesting question is not 'why does it need the cash.' It doesn't. The question is what the borrowing is for. And the answer, I think, has very little to do with the $25 billion on Nvidia's own balance sheet and almost everything to do with the far larger pile of debt sitting on everyone else's.

The real thing, first

Let me separate the real thing from the trade built on it, because that is the only honest way to read a deal like this. The real thing is that Nvidia's chips are the most sought-after industrial input on the planet right now, and the company's cost of borrowing is, accordingly, about as low as a corporate borrower's gets. Reuters reported that one of the deal's actual objectives was to establish a liquid benchmark for Nvidia's cost of credit — in plain English, to put a clean, heavily-traded yield curve on the board so that everything Nvidia and its orbit borrow later can be priced against it. That is not a distress signal. That is a company building plumbing it expects to use a great deal.

So this is not a story about a great company in trouble. Conflating a bad price with a bad company is how people lose money in both directions, and Nvidia is not a bad company. The question I want to sit with is narrower and, I think, more important: what happens to the AI trade when its strongest balance sheet starts lending its strength to everyone else's borrowing.

The borrowing that isn't on the balance sheet

Here is the part of the week's news that got a fraction of the attention the $25 billion did. Nvidia is increasingly not just a borrower but a credit enhancement — the reason other people's riskier debt gets bought at all. The widely-cited example is a data-centre financing that raised billions in below-investment-grade bonds to build a single large facility, on terms that investors were willing to accept in large part because of a long-dated lease with Nvidia as the anchor tenant. The reported structure: a roughly $4.59 billion junk-bond deal for a 200-megawatt build, leaning on a 16-year Nvidia lease to make the cash flows look dependable.

Read that mechanism slowly, because it is the whole column. A speculative project borrows money it could not otherwise borrow cheaply. The thing that makes the lenders comfortable is not the project's own prospects — it is a promise from Nvidia to rent the capacity for sixteen years. Nvidia's creditworthiness is doing the work that the borrower's own balance sheet cannot. The chipmaker is, functionally, co-signing the loan. It does not appear as Nvidia's debt. It appears as Nvidia's demand.

The chipmaker is functionally co-signing the loan. It does not appear as Nvidia's debt. It appears as Nvidia's demand.

And that is the elegant, dangerous part. A lease commitment from Nvidia turns into a building full of Nvidia chips, financed by a bond that exists because Nvidia stood behind it. The demand for the product is being manufactured by the maker of the product, using its own balance sheet as the collateral. When you hear that AI infrastructure is attracting record debt issuance, understand that some meaningful slice of that debt is being made fundable by the same company that sells the picks and shovels. The circularity is not a glitch. It is the design.

Compared to when?

I keep a file of every arrangement that was supposed to be different this time, and this one has a clear ancestor. At the turn of the millennium, the great telecom-equipment makers — Lucent, Nortel, the names that were going to wire the new economy — discovered that the fastest way to sell more gear was to lend their customers the money to buy it. It was called vendor financing. The customers, many of them newly-minted competitive carriers with thin balance sheets, took the loans and bought the switches and the routers, and for a couple of glorious years the revenue line and the demand line went up together because they were, in a real sense, the same line.

Then the customers could not pay. The carriers went bankrupt, the equipment came back or simply went dark, and the vendors discovered they had booked sales to entities that only existed because the vendors had funded them. The receivables turned to ash. Lucent's vendor-financing book became one of the more instructive craters of that cycle. The lesson was not that the technology was fake — the internet was extremely real, the fibre got laid, the world did change. The lesson was that when a supplier underwrites its own demand, it stops being able to tell the difference between a customer and an extension of itself, and the market loses its only honest signal: whether anyone would have bought the thing without the financing attached.

I want to be careful here, because being early is a way of being wrong and I have done it. Nvidia's lease commitments are not Lucent's receivables. A 16-year lease on power-hungry capacity from a company swimming in cash is a far sturdier thing than a loan to a carrier with no revenue, and Nvidia is structuring exposure, not booking phantom sales. The analogy is not 'this ends the same way.' The analogy is 'this is the same mechanism,' and the mechanism has a known failure mode: it works beautifully right up until the underlying demand it is propping turns out to be partly its own reflection.

What's actually priced in

So distinguish the company from the price, as ever. Nvidia the company is making a shrewd, even conservative move — locking in cheap long-term funding, building a benchmark curve, using its balance-sheet strength as a strategic asset. If you are Nvidia, this is good management. The risk does not sit with Nvidia. It sits with everyone who has bought, or will buy, the debt that Nvidia's lease commitments make investable — the pension funds and insurers and credit funds reaching for yield in data-centre bonds that look investment-grade because of who the anchor tenant is.

What is priced into those bonds is the assumption that Nvidia's demand for compute capacity is permanent and that Nvidia's willingness to keep signing 16-year leases is unconditional. Both are plausible today. Neither is a law of nature. The history of vendor-supported booms is that the supplier's appetite is the most pro-cyclical thing in the whole structure — enormous when the cycle is up, and exactly the thing that vanishes first when it turns, because a chipmaker facing a glut stops signing leases long before it stops making chips.

None of this means the AI buildout is a mirage, and I am not going to cry bubble at a balance-sheet manoeuvre. The chips are real. The data centres are real. The demand, for now, is real. But the $25 billion Nvidia borrowed for itself is the least interesting debt of the week, because Nvidia can clearly pay it back. The debt worth watching is the much larger amount that other people borrowed, that only got cheap because Nvidia's name was on the lease — and that gets repaid only if the demand Nvidia is currently manufacturing turns out to have been demand all along. The question, as it always is, was never whether the technology is real. It was always what you paid, and who is holding the paper when the appetite changes.