The headline number from Nvidia's consortium with Goldman Sachs, BlackRock, Blackstone, KKR, Apollo and Brookfield has been picked over since it landed. The more useful line came later, when Jensen Huang described what happens when one of these AI factories fails.
The architecture keeps running. Someone else picks it up.
That single point is what converts a chip into a 'financeable asset', and it is the part of the structure that has had the least scrutiny.
A collateral problem nobody had solved
Project finance works when the lender can seize something and sell it. Power stations, aircraft, toll roads. The asset outlives the operator.
Compute has never passed that test. A specialised cluster built for one tenant is close to worthless once that tenant stops paying.
Huang's answer is that Nvidia specifies the system architecture itself, applying what he called a "stamp of approval" to particular customers. The design is meant to be universal. If the operator hits trouble, the platform still functions and another party can run it.
Fungibility is doing the heavy lifting here. CUDA is the reason a GPU bought for one workload can be repointed at another.
Nvidia is not writing the cheques
The financing is third party. Nvidia has been explicit that it is not committing the $500 billion and is not backstopping individual projects, including the reported $250 billion OpenAI site in Ohio.
The Korea deal shows the shape of the model. Total investment of $10 billion for Naver compute, of which Nvidia put in $1 billion and partners found the remaining $9 billion.
So the exposure is reputational and residual, not balance sheet. Nvidia is underwriting the resale value of the collateral rather than the credit of the borrower.
That is a smaller commitment than it first appears. It is also a harder one to walk away from.
1970s comparison cuts both ways
Data centre financing was described on the call as a new frontier for financial engineering, on a par with the birth of the mortgage-backed securities market.
The comparison is meant as a compliment about scale. Roughly $9 trillion sits in US money market funds and $100 trillion in equities, against a build cost of $50 billion to $60 billion per gigawatt and more than 70 gigawatts needed in the US alone.
It carries an obvious second meaning. Securitisation markets work until the assumption about recovery values proves wrong.
What to watch before 26 August
Demand is not the question. Requests for large language model capacity have risen sevenfold in six months, and Micron can fill about half of what data centre customers ask for.
The question is what happens to residual values when the first operator does fail. Nothing in the structure has been tested through a default.
Nvidia reports on 26 August. The financing platforms will matter more to the shares than the quarter itself.
