SpaceX has agreed to lease about $47 billion of Nvidia's newest Vera Rubin accelerators on terms that read like a car lease, and the roughly $40 billion of financing behind the purchase will never appear on SpaceX's own balance sheet. Semafor reported on Tuesday, citing people familiar with the matter, that Apollo Global Management is assembling the package through a special-purpose vehicle, with portions of the seven-part deal possibly offered to investors as early as this week. Apollo declined to comment and SpaceX did not respond to a request for comment.
The structure works like this: an SPV managed by Atlas SP, an Apollo-backed platform, buys the Nvidia chips and issues asset-backed securities against them, while SpaceX pays rent to use the hardware. Bloomberg, relayed by Yahoo Finance, reports a $7 billion down payment already made, with each batch of chips leased for roughly five years and some contracts running into 2032. Rent is owed even if the chips are damaged, the leases cannot be terminated early, and SpaceX does not guarantee what the silicon will be worth at the end, leaving the residual-value risk with lenders. Goldman Sachs is separately leading an $8.3 billion senior secured loan that gets repaid first and is backed directly by the chips.
The pricing is expensive, and deliberately so. Semafor says the bulk of the deal, $35.5 billion, will price at 8.3 percent and 7.75 percent, while the riskier tranches carry a blended rate that could exceed 9 percent on less than $3 billion of the total. Apollo plans to lend close to $20 billion of the package upfront and then syndicate or sell the rest. All of this lands in a market where AI and data-center debt has become measurably harder to place, as investors grow choosier about who is holding the risk of a very fast-depreciating asset class.
The off-balance-sheet trick is the point. Meta has used similar structures, and keeping the debt out of SpaceX's filings protects the investment-grade narrative the company built around its June IPO. The market is not fully convinced: SpaceX's own debt climbed from about $23 billion at the end of 2025 to nearly $39 billion by September 30, and its five-year credit default swaps have widened from roughly 110 basis points at the IPO to about 265 basis points on November 16, which Yahoo Finance's arithmetic reads as roughly 20 percent implied odds of default within five years.
The exposure has a specific shape. One customer, identified by Bloomberg as Anthropic, contributed about a third of SpaceX's third-quarter revenue on a contract that can be ended with 90 days' notice, while the rent on the chips runs five years. SpaceX finance chief Bret Johnsen has pointed to payback periods of less than a year on AI infrastructure, but that math sits awkwardly against non-cancellable five-year leases on hardware Nvidia plans to outdate with two newer chips by 2028. The lease alone already exceeds everything CoreWeave has borrowed against chips as collateral since 2023.
The wider context is that this deal is a hand-off, not an outlier. JPMorgan's Michael Cembalest estimates Nvidia's off-balance-sheet commitments doubled in a single quarter to $422 billion, and that big technology companies, Nvidia and Anthropic together carry more than $4 trillion of AI-related commitments that do not appear on their books, roughly six times what does. The Federal Reserve warned last week that AI-related default fallout could end up larger than current balance sheets suggest. Benchmark partner Bill Gurley said vendor financing has "gone too far," and Dan Niles of Niles Investment Management quipped that "Nvidia the bank" may be "closed for a while." Goldman's credit team, for its part, projects AI chip financing will grow nearly sevenfold next year.
What happens this week is a template test. Private credit is stepping in exactly as the chip vendor steps back, at 7.75 to 8.3 percent on the safe tranches and double digits on the risky ones. If the seven tranches clear at those levels, expect the SPV-plus-ABS pattern to spread quickly to the next oversized GPU order; if investors balk at owning five-year residual risk on silicon, financing, not demand, becomes the binding constraint on the AI buildout. Siebert Financial CIO Mark Malek put the bottleneck plainly: there are only so many dollars out there for these companies, and they need all of them.
Comments (0)
Log in to join the discussion
Log InNo comments yet