The collateral behind the $500 billion is not the GPUs

What keeps a $500 billion loan book safe is not the GPUs. It is the credit of the companies renting them.

Nvidia turns to Wall Street giants to raise $500bn for AI infrastructure (The National / Bloomberg, 2026-08-11)

출처: thenationalnews.com

Bloomberg reported that the six firms designed the structure around separate entities able to issue tens of billions of dollars of bonds at a time, with Goldman Sachs running the public bond sales.

Why a supermarket chain belongs in this story

Meru reads the structure as a sale-and-leaseback. You sell a building you own, pocket the cash, and rent the same building back on a long lease. Swap the store for a data center and the tenant for an AI buyer such as OpenAI, and only the cast changes. What the lender actually holds is not the building. It is the tenant's ability to keep paying every month.

The take-or-pay contract has holes in it

The tool used to close that risk is a take-or-pay contract. It says you pay for the agreed amount whether or not you actually use it. Revenue holds up even if demand softens, so credit committees treat such a contract almost as collateral. But in both the United States and South Korea, an administrator in a court restructuring can reject contracts that have not yet been performed.

When Homeplus filed for court receivership in March 2025, its leases were terminated. The property funds and REITs that had been paying bank interest out of that rent were left holding stores that had stopped trading. The life of the contract was set by the tenant's condition, not by the term written into it. Bankruptcy is not even required for this to bite. Once a tenant's credit wobbles, renegotiation on softer terms usually arrives before any default.

The 25% that was in the announcement

One side is looking at the quality of the collateral, the other at why the guarantee was needed at all.

The rate OpenAI ends up paying

This is still a memorandum stage, so what can be observed now is the price rather than the size. Meru put it at roughly 6% for an investment-grade borrower like Google, and questioned whether OpenAI could raise money even at 15%. If the first bonds price in investment-grade territory, the structure settles in as infrastructure finance; if they price in high-yield territory, the market is still reading it as money supplied by the seller. Samsung Electronics and SK hynix, which simply hand over chips and take payment, will read the size of the next order off that same rate.