SpaceX in talks on $40 billion of debt for Nvidia chips, with the GPUs likely as collateral
The financing leans on a market belief that a GPU holds its value for about seven years. Spread over that life, $40 billion of chips is roughly $5.7 billion a year before a dollar of interest.
Apollo and several banks are in talks with SpaceX to finance a $40 billion purchase of Nvidia graphics processors, CNBC reported on Wednesday, citing a person familiar with the talks. The deal would rely mainly on the investment-grade bond market, and the GPUs themselves would likely serve as collateral. CNBC says the discussions are preliminary and that the lenders are not final. Apollo declined to comment and Nvidia had not responded when CNBC published. CNBC credits the Financial Times with the first report; a Bloomberg item on Tuesday carried the same $40 billion figure, attributed to the FT. We could not read the FT piece itself, and neither company has confirmed anything publicly.
The number behind the number: seven years
The detail that matters is not the $40 billion. It is the assumption under it. CNBC quotes a credit-industry source saying the market still works on the expectation that GPUs keep their value for about seven years, supported by a shortage of computing power that is not expected to ease soon.
Work that through. $40 billion divided by seven years is about $5.7 billion a year in value lost to age, if the chips fade evenly. That is our arithmetic, not a figure from either report, and real depreciation rarely runs in a straight line. If a new Nvidia generation makes the old chips worth less sooner, the collateral shrinks faster than the loan does. That is the risk a lender in this structure is underwriting.
What it costs to borrow right now
CNBC notes that SpaceX raised $25 billion of bonds two weeks after its mid-June IPO, to strong demand, but that AI-related bonds have since sold off and credit spreads have widened. Investors are asking a higher yield for AI debt, the sources said, and more sales are expected.
For scale, the 10-year Treasury yield closed at 5.27% on October 6, according to Treasury's daily curve. In September we reported AI borrowers paying about 115 basis points over Treasuries, in our story on AI bond spreads. As an illustration only, not a quote for this deal: 5.27% plus 1.15 points is 6.42%, which on $40 billion is about $2.6 billion of interest a year. Spreads have reportedly widened since then, and a secured deal could price differently, so the real figure could land on either side.
Put the two together and the illustration implies something like $8 billion a year of combined interest and chip wear on the purchase. The question for bond buyers is whether the compute the chips produce earns more than that.
Who it touches
- Investment-grade bond fund holders. A deal this size adds to the supply of AI paper that index funds absorb automatically.
- Anyone else borrowing. Heavy AI issuance competes for the same buyers. Companies outside the sector, including mid-sized firms that borrow floating or refinance at market rates, are paying a 10-year benchmark above 5% already.
- Traders in AI and chip stocks. Debt-funded chip orders support demand for Nvidia, but they also tie that demand to credit conditions. If lenders balk at pricing, the order book is where it would show.
For the benchmark behind all of it, see our 10-year Treasury chart.
Sources: CNBC; Bloomberg, citing the Financial Times; U.S. Treasury; Chronicle reporting. Interest and depreciation figures are our own illustrative arithmetic, not terms of any deal. This is market information, not investment advice.
Want your business to be the answer?
Get a full package of articles about your business, built so customers, Google and AI assistants can find you.