The collateral behind the bonds is a shipment of processors that will not arrive until the first half of 2027. The borrower is a rocket company. The seller of the debt is a bank, and the entity that actually owns the hardware is a vehicle set up by Apollo.
Goldman Sachs and Apollo plan to begin marketing about $4 billion of SpaceX bonds as soon as this week, according to Bloomberg. The sale is part of a $40 billion financing package built around Nvidia’s next generation of artificial-intelligence chips, the Vera Rubin line.
The structure is worth following closely, because it is becoming the template for how large technology companies pay for computing without adding the debt to their own balance sheets.
Atlas SP Partners, the asset-backed finance arm Apollo built out of Credit Suisse’s securitized products business, is setting up a special-purpose company. That company buys the chips. It then leases them to SpaceX for five years, a term that runs to 2032. SpaceX has already paid $7 billion upfront, and the remaining obligations sit with the vehicle rather than with the rocket company.
Goldman is separately leading an $8.3 billion senior secured loan. Together with the equity cheque, the arrangement covers roughly $47 billion of chips.
The economics are simple in outline. Nvidia gets paid. SpaceX gets the machines without recording the borrowing. Bondholders get a claim on hardware that can be repossessed and, in principle, re-leased to someone else if SpaceX stops paying.
The risk sits in the residual value. A graphics processor bought in 2027 for a five-year lease has to hold value until 2032, and the history of this hardware is that each generation reduces the rental price of the one before it. Lenders in these structures typically assume a depreciation curve and size the debt against it. If the curve steepens, the equity in the vehicle absorbs the loss before the bondholders do, which is why the upfront payment matters.
There is a second risk, and it is not financial. The chips do not exist yet in volume. Delivery depends on Nvidia’s manufacturing partners, on memory supply and on packaging capacity, all of which have been the binding constraint on the industry for two years. A lease that begins on delivery cannot begin if delivery slips.
For SpaceX the appeal is straightforward. The company is private, does not publish full financial statements, and has an enormous capital program across launch, satellite broadband and now computing. Keeping $40 billion of hardware debt off its balance sheet preserves borrowing capacity for everything else and keeps the reported picture of the business cleaner than the underlying commitments.
That is a legitimate use of the structure and also the reason accountants and rating analysts examine it closely. Lease obligations are disclosed in footnotes even when they are not on the face of the balance sheet, and a lender comparing SpaceX with a public peer will add them back.
The $7 billion upfront payment is the part of the structure that signals confidence. A lessee with doubts about the hardware, or about its own need for it, would not put that much equity into a vehicle it does not control. SpaceX did, which effectively subordinates its own money to the bondholders’ claim and gives the sale something to point at when investors ask who is first in line if the lease breaks.
The template has been used this year to finance data-center construction for other large technology companies, and its spread tells you something about the state of the market. Investors want exposure to artificial-intelligence infrastructure but not necessarily to the credit of a single operator. A vehicle that owns the equipment and leases it out offers a claim on the assets with a defined term and a defined collateral package.
It also moves the risk to a place where it is harder to see. Ten years ago, a company that wanted $40 billion of equipment borrowed $40 billion and reported it. Now the same company can lease it, and the obligation appears as a payment schedule rather than as debt.
SpaceX’s own use of the capacity has not been detailed. The company runs satellite broadband, launch operations and a growing set of government contracts, and it has been buying computing for reasons it has not disclosed in public filings.
The term of the lease is the tell. Five years from delivery takes the contract to 2032, which is longer than most operators are willing to commit to a single generation of accelerators and shorter than the useful life lenders would prefer to underwrite. Someone negotiated a duration that matches an expectation about how long these chips stay competitive, and the coupon will be set against that expectation.
What the bond sale will establish is a price for that risk. The spread over Treasuries on a five-year lease of depreciating processors, guaranteed by nothing but the equipment, is the market’s estimate of how fast this hardware becomes obsolete. Everyone financing the buildout is using a number for that. This week one of them becomes public.


