The Vendor Is Now the Underwriter
Nvidia's quarter was reported as a supply story. The more consequential work was on the other side of the transaction — making someone else willing to lend against the chips.
Nvidia reported its second quarter on August 26 as a supply story, and on the surface it is one. Revenue of $96.2 billion, up 106 percent from a year earlier. A data center segment of $89.0 billion, up 117 percent. A third-quarter guide of $108 billion. Colette Kress, the chief financial officer, told analysts the company expects to grow revenue by roughly 70 percent in fiscal 2028, then attached the qualifier that did the real work: “This is a supply-constrained outlook.” Jensen Huang said the unconstrained figure “would be a lot higher.” A company that cannot make enough of what it sells has a pleasant problem to narrate. It is also not the problem Nvidia spent the summer solving. The harder one sits on the other side of the transaction: where the money to buy all of it comes from.
The scarce input is no longer silicon
Sixteen days before the print, Nvidia announced financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, intended to mobilize more than $500 billion of third-party capital for AI compute infrastructure over time. The release is careful in the ways that matter: the platforms are described as independent, final agreements are noted as pending execution, and no guarantee from Nvidia is disclosed. What the company asserts instead is a claim about the asset — that “NVIDIA compute is an investable asset,” with the lowest token cost, the highest revenue and the longest life. Jim Zelter of Apollo called modern compute “a scarce, mission-critical asset class with compelling investment characteristics.”
Strip the adjectives and this is a problem being solved for lenders, not for buyers. Nobody needed convincing that GPUs are in demand. What a project lender needs is a residual: a defensible view of what the collateral is worth in year four, in the downside where the borrower’s revenue did not arrive. Six of the largest alternative asset managers in the world do not commit capital to a new category on the strength of an adjective. They commit on a structure.
Kress described one on the earnings call. For the neoclouds — the specialist providers that rent compute rather than consume it — Nvidia has “introduced a revenue-sharing structure” under which it provides “a take-or-pay commitment,” a minimum revenue guarantee “that gives lenders the confidence to underwrite the project.” Huang, separately, said the company has invested nearly $50 billion in the frontier AI labs. And the receivable line moved: days sales outstanding rose to 60, which Kress attributed to longer payment terms extended to certain investment-grade customers.
What a minimum revenue guarantee actually does
Read together, those three disclosures describe a business that does not appear in the segment table.
The chips are still the product. What Nvidia has started supplying alongside them is the credit enhancement that makes someone else willing to lend against them.
A take-or-pay commitment does not decide whether a data center gets built. It decides who absorbs the loss if the racks sit idle. The lender is no longer underwriting a neocloud’s ability to fill a cluster at an assumed price; it is underwriting Nvidia’s willingness and capacity to make up the shortfall. Project risk — hard to price, harder to syndicate — becomes single-name corporate exposure to a company earning a 75 percent gross margin. That is genuinely useful financial engineering, and it is the reason $500 billion is a plausible number rather than a press release.
It carries an accounting asymmetry worth naming plainly. The revenue such a structure unlocks books now, at full margin, in the segment investors value most. The obligation it creates is contingent, lives in commitments and contingencies, and books only if things go wrong. There is nothing improper in that; it is how guarantees work in every industry that uses them. But it means the income statement and the risk profile are moving in the same direction at different speeds, and only one of them is in the headline.
The case that none of this matters
The strongest counter-argument is a scale argument, and it deserves to be made properly.
Nvidia generated $21.3 billion of free cash flow in the quarter and returned $26.0 billion to shareholders. Against that, a portfolio of revenue floors on projects that will mostly get built and mostly get let is a rounding error — and one that costs nothing in the states of the world where demand holds. The counterparties on the extended payment terms are described as investment grade. Sixty days of DSO is unremarkable for capital goods of this size; it would be short for an aircraft.
The collateral argument is stronger still. Huang’s claim that the platform is “fungible and durable” and can be redeployed to other customers is not spin. A repossessed cluster is a general-purpose asset with an active rental market, which is precisely what the vendor-financing failures of the last telecom cycle lacked: that gear was purpose-built for one buyer and had no second life. And the per-gigawatt economics are improving rather than decaying. Huang put the revenue opportunity per gigawatt at roughly $18 billion on Hopper, $25 billion on Blackwell, and $40 billion on Vera Rubin. Rising asset productivity is the friend of any residual-value assumption a lender has to defend.
All of that holds. It establishes that the exposure is small and well-collateralized in the ordinary case. It does not speak to the case the structure exists to insure against.
The exposure is correlation, not size
Guarantees are not priced by their notional. They are priced by when they get called.
The event that causes a neocloud to miss its revenue floor is a slowdown in demand for rented AI compute. That same event compresses Nvidia’s own order book. It also lowers the resale value of the collateral behind the loan, because the buyer of a used cluster is another operator working from the same demand assumption. Three exposures that look independent on a schedule converge in the one scenario that matters. Nvidia has, in effect, written a put on its own end market, taken payment for it in the form of higher current revenue, and secured it with collateral whose value is a function of the same variable.
The cushion is thinner than 75 percent suggests, too. Kress guided gross margin to a trough of 71 to 72 percent in the fourth quarter, settling at 72 to 73 percent in fiscal 2028, on component costs whose increases have “exceeded our prior expectations and are headed even higher into next year.” Inventory stood at roughly $32 billion ahead of the Vera Rubin launch. None of this is distress. It is a reminder that the margin which would absorb a called guarantee is being negotiated downward by suppliers at the same moment the guarantees are being written.
We argued in July that the build-out was being financed as though compute were a fixed asset that lasts forever. The question has since moved one step earlier in the chain. It is no longer only how long the asset lasts. It is who stands behind the loan that bought it — and whether that party’s own results are correlated with the reason the loan goes bad.
Huang’s formulation in August was that compute is revenue. It is a good line and, on the current numbers, a true one. The more consequential sentence of the quarter was Kress’s, and it was addressed to lenders. For a growing share of the compute now being installed, Nvidia is not merely the seller. It is the reason someone else was willing to finance the purchase.