SpaceX’s $44 Billion Nvidia Chip Deal (2026): Why Lenders Don’t Want the GPUs Launched Into Orbit

Jump to comments

A SpaceX Falcon 9 rocket lifts off at dusk from a Florida launch pad between four lightning towers, with a bright orange exhaust plume and billowing white clouds reflected in the water in front
A SpaceX Falcon 9 lifts off with 23 Starlink satellites from Cape Canaveral, Florida, on March 10, 2024.
Photo: U.S. Space Force/Joshua Conti, 2024, via Wikimedia Commons (Public domain)

On Thursday, Nov. 19, 2026, bond investors bought about $32 billion of debt backed by Nvidia chips that have not been delivered yet. The next morning, Apollo and Goldman Sachs announced that the full package had priced at roughly $44 billion, calling it "the largest GPU financing executed to date." The money buys next-generation Nvidia Vera Rubin hardware for SpaceX. Then came the detail that sent the deal around X: according to the Financial Times, the lenders had tried to stop Elon Musk from launching their collateral into space.

The short answer: SpaceX is not buying these chips itself. A trust owned by Apollo borrows the money, buys the Nvidia racks and leases them to SpaceX for about five years, starting in the first half of 2027. SpaceX’s rent pays the bondholders. If SpaceX ever stopped paying, the lenders’ backstop would be to take the chips and sell them, which only works if the chips sit somewhere a repossession crew can reach. SpaceX has said it plans to fly space-optimized versions of the same Vera Rubin NVL72 systems on orbital data-center satellites starting late next year, so lenders asked for a promise that their collateral stays on the ground. Whether that promise made it into the final documents is still unknown.

Here is how the deal works, why a rocket company is financing chips like an airline finances jets, what the orbit clause is really about, and what the filings say about SpaceX’s ability to pay.

What SpaceX actually agreed to

The deal started with a short filing. In an 8-K dated Oct. 24, 2026, SpaceX told investors it had signed a binding commitment letter for a "master finance lease" with an entity managed by Apollo Global Management. That entity, the lessor, will acquire and lease to SpaceX "graphics processing unit pods and related equipment" for use at data centers SpaceX selects.

The filing spells out the mechanics. The equipment arrives in tranches, and each tranche gets its own lease schedule running about five to five and a quarter years. Rent is due every quarter, in advance, and is generally one-twentieth of the tranche’s financed cost plus interest at a fixed rate and some ongoing costs. Twenty quarterly payments of one-twentieth each means the hardware is fully paid off over the five-year lease. SpaceX put the total principal it expects to pay at about $40 billion and said the leases should begin in the first half of 2027, when the GPUs are delivered.

Apollo’s Nov. 20 announcement filled in the rest. The notes that fund the purchase were "oversubscribed and upsized" because of investor demand, with funds managed by BlackRock, Goldman Sachs Asset Management and accounts advised by KKR providing anchor commitments. Goldman Sachs was sole structuring agent. The release describes the hardware as "the next generation of NVIDIA Vera Rubin chips for SpaceX’s terrestrial compute buildout," and Apollo partner Jamshid Ehsani called the deal an example of the firm’s "focus on downside-protected, asset-based structures."

The deal, piece by piece

The trade publication IFR reported the individual pieces of the package, which the deal’s backers called "Project Saturn." The bonds were sold through Saturn Funding Trust, a special purpose vehicle owned by Apollo. A separate Apollo vehicle takes ownership of the chips if SpaceX defaults.

PieceAmountTermsLegal maturity
Senior secured class A3 notes$25.168 billion3.45 points over Treasuries, an 8.259% yieldApril 2033
Senior secured class A2 notes$4.366 billionSenior secured, shorter datedApril 2031
Class B subordinated notes$2.564 billionSecured, paid after the A notesApril 2033
Senior secured loanUp to $8.236 billionCould shrink because the bonds were upsizedNot disclosed
Class C, D and E notesUp to $2.4 billionPlaced privatelyNot disclosed
Upfront rent from SpaceX$7.1 billionPaid by SpaceX at the startNot applicable

The three public bond tranches add up to about $32.1 billion. IFR said all three are expected to be repaid more than a year before their legal maturity dates, and the FT reported that principal is paid off in less than four years, a schedule designed to cut lenders’ risk. Moody’s rated the debt A3, one notch above SpaceX’s own credit rating, and KBRA gave the class A and B notes a preliminary A-minus. Both agencies credited "hell-or-high-water" clauses that stop SpaceX from walking away from the lease. IFR described the underlying order as a $47 billion purchase of Nvidia chips.

Some quick math shows what SpaceX is signing up for. At one-twentieth of roughly $40 billion per quarter, principal alone comes to about $2 billion a quarter, or $8 billion a year, once all the equipment is delivered and leasing. Interest at a rate near the bonds’ 8% yield comes on top, falling as the balance shrinks.

Why lenders care where the chips end up

Chip-backed loans have become a standard way to pay for AI data centers. Cloud companies borrow against their GPUs and repay from the revenue the chips earn. The chips are the safety net: if the borrower defaults, lenders can take the hardware and resell it. That backstop is real even for older parts. TechCrunch cited data showing some Nvidia H100s from 2023 now sell for about half their original cost, while some H200 setups still sell for what they cost when introduced in 2024.

Even on Earth, that safety net has limits. KBRA warned about it in a note before the sale, quoted by IFR:

GPUs are rapidly depreciating assets, and their value at the time of a potential default, as well as the extent to which such value would be sufficient to cover the then outstanding debt, is inherently uncertain.

KBRA, in a pre-sale note quoted by IFR

Now put the collateral on a satellite. Nobody is going to repossess a server rack moving at about 17,500 miles per hour, hundreds of miles above the ground, and nobody is going to buy one there either. Once launched, the chips stop being something a lender can seize and sell, and become a bet on SpaceX’s credit alone. According to the FT, lenders worried that if SpaceX later decided to send the financed chips to orbit, they could not reclaim and sell them after a default. The proposed covenant, the FT added, "presented potential legal challenges" for SpaceX’s plans and showed how new this kind of financing still is.

Some investors argue the collateral never mattered much in the first place. "This is really SpaceX-secured equipment, dressed as an ABS," one portfolio manager told IFR, meaning the bonds look like asset-backed securities but really depend on SpaceX paying its rent.

SpaceX plans to fly the same hardware

The concern is not hypothetical, because SpaceX has been specific about what it wants to put in orbit. On Jan. 30, 2026, it asked the FCC for permission to launch and operate up to one million satellites that would work as orbital data centers, in narrow shells between 500 and 2,000 kilometers up. Days later, a SpaceX update laid out the math: launching a million tons of satellites a year, at 100 kilowatts of compute per ton, would add 100 gigawatts of AI compute annually. "My estimate is that within 2 to 3 years, the lowest cost way to generate AI compute will be in space," the update says.

The hardware choice followed. In June, SpaceX described its first data-center satellite, AI1, as 20 meters tall with a 70-meter wingspan and about 150 kilowatts of peak power, Light Reading reported. On Aug. 4, Nvidia posted that SpaceX’s "Starmind AI1 satellite compute payload is powered by NVIDIA Vera Rubin NVL72." On Aug. 24, Musk wrote that SpaceX and Nvidia had designed "a space-optimized Vera Rubin NVL72 system for launch to orbit in Q4 next year, with significant scale in 2028." A month later he added that each satellite "is 250kW and will have a SpaceX-designed Nvidia Vera Rubin NVL72 computer."

So the racks the bondholders are paying for are the same model SpaceX wants to launch, in a modified form. SpaceX has said the financed chips will be used on Earth, and Moody’s wrote that they will sit in U.S. data centers majority-owned, controlled and operated by SpaceX. The covenant fight was about making sure that stays true for the life of the debt.

What a Vera Rubin NVL72 is

Vera Rubin NVL72 is Nvidia’s newest rack-scale AI computer. According to Nvidia’s spec sheet, one liquid-cooled rack holds 72 Rubin GPUs and 36 Vera CPUs linked into a single system, with 20.7 terabytes of high-bandwidth HBM4 memory and 216 terabytes per second of sixth-generation NVLink bandwidth between the chips. It succeeds Nvidia’s Blackwell-based GB200 and GB300 NVL72 racks.

A liquid-cooled Nvidia GB200 compute tray on a black display table, showing GPU and CPU packages connected by black coolant hoses and cables, with a GB200 label at left
A liquid-cooled Nvidia GB200 compute tray, the building block of the NVL72 racks that Vera Rubin succeeds, on display at Computex 2024.
Photo: Geekerwan, 2024, via Wikimedia Commons (CC BY 3.0)

In its Nov. 17 earnings release, Nvidia said it began shipping Vera Rubin NVL72 in its third quarter, "the fastest product ramp in the company’s history," with racks running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius, and "coming online soon at SpaceXSI." Nvidia does not publish rack prices, and neither Apollo nor SpaceX has said how many racks the deal buys.

Questions people are asking

Did the no-orbit clause make it into the final deal?

Nobody has said. The FT reported that "it is uncertain whether the provision was included in the finalized documents." Apollo, Goldman Sachs and Pimco declined to comment to the FT, and SpaceX did not respond. The closest thing to an answer is Apollo’s own wording, which ties the chips to SpaceX’s "terrestrial compute buildout."

Wouldn’t lenders still own the chips in orbit?

On paper, yes. Article VIII of the Outer Space Treaty says ownership of objects launched into space is not affected by their presence there. In practice, ownership is worth little if the asset cannot be reached, switched off or resold. The legal system built for this problem is not working yet. The Space Assets Protocol to the Cape Town Convention, adopted in Berlin on March 9, 2012, would create an international registry for security interests in satellites, much like the one lenders rely on for aircraft. It is still not in force. Only two countries have joined: Paraguay in November 2024 and the Democratic Republic of the Congo in June 2026.

Can SpaceX afford the payments?

Its latest numbers are strong but capital-hungry. In the 10-Q for the quarter ended Sept. 30, 2026, SpaceX reported revenue of $14.58 billion, up from $5.58 billion a year earlier, and $2.79 billion of operating income across its three segments. Its AI unit, now called the Super Intelligence (SI) segment, brought in $7.56 billion of that, up from $813 million, almost all from "SI Solutions & Infrastructure." Backlog stood at $62.96 billion. On the other side, capital spending reached $48.07 billion in the first nine months of 2026, and $40.0 billion of it went to the SI segment. Nameplate compute, SpaceX’s measure of installed GPU power, hit 1.7 gigawatts, up from 0.6 gigawatts a year earlier. Putting the chips in an off-balance-sheet trust helps SpaceX keep its investment-grade rating while it keeps spending, the FT reported.

Who will use the chips?

SpaceX increasingly rents out compute rather than using it all for its own models. Anthropic agreed to pay SpaceX $1.25 billion a month for capacity at the Colossus and Colossus II clusters, with either side able to end the arrangement on 90 days’ notice, according to a SpaceX letter to the SEC. The 10-Q shows one SI customer, listed only as Customer B, made up 34.5% of SpaceX’s total third-quarter revenue. IFR called Anthropic "one of the likely end customers" for the new chips. When it signed its Colossus deal in May, Anthropic also said it was interested in working with SpaceX on "multiple gigawatts of orbital AI compute capacity."

Why did investors demand more than 8%?

Timing and trust. The bonds launched hours after Anthropic priced its record IPO, but AI stocks and bonds sold off through the day, pushing the Nasdaq Composite to its lowest level in seven months, IFR reported. The A3 notes priced at an 8.259% yield, while SpaceX’s existing bonds maturing in 2031 traded around 7.2% that day, according to the FT. Some investors passed. One told IFR that talk of a possible SpaceX merger with Tesla made it hard to know where the debt would sit; SpaceX’s 10-Q lists "the evaluation of a potential combination with Tesla" among its forward-looking risks.

What happens next

SpaceX said it expects the lease to close before the end of 2026, with deliveries and lease payments starting in the first half of 2027. Its orbital timeline runs right behind that: a first space-rated Vera Rubin system in the fourth quarter of 2027 and "significant scale" in 2028, if Starship can fly often enough. On its most recent test, Flight 15 on Nov. 13, SpaceX caught a Starship upper stage at the launch tower for the first time, but the Super Heavy booster was not recovered, and TechCrunch noted it is unclear how quickly a ship can be readied to fly again.

The bigger lesson reaches beyond SpaceX. Whoever wants to build data centers in space, including Google with Project Suncatcher and startups such as Starcloud, will need to pay for chips, and the cheapest way to pay for chips today relies on being able to take them back. Until there is a way to lend against hardware in orbit, orbital compute may have to be paid for with equity, cash flow or pricier debt. For a deeper look at the engineering side, see our explainer on AI data centers in space, and for how ground clusters handle heat, our guide to how data centers stay cool.

Further reading

Cover of Liftoff by Eric Berger
Liftoff: Elon Musk and the Desperate Early Days That Launched SpaceX

Eric Berger’s history of SpaceX’s first four Falcon 1 launches, the near-failures that shaped the company now financing AI chips.

Cover of Chip War by Chris Miller
Chip War: The Fight for the World’s Most Critical Technology

Chris Miller’s prize-winning history of the semiconductor industry and why advanced chips became the most fought-over product on Earth.

NVIDIA Jetson Orin Nano Super Developer Kit board with cooling fan
NVIDIA Jetson Orin Nano Super Developer Kit

Nvidia’s palm-sized AI computer for hobbyists and students, the cheapest way to run models on the same CUDA software stack as a data center.

Celestron StarSense Explorer LT 80AZ refractor telescope on a tripod with a smartphone dock
Celestron StarSense Explorer LT 80AZ Telescope

A beginner refractor that uses your phone to find targets, handy for spotting planets and bright satellites passing overhead.

Amazon Echo Dot smart speaker
Amazon Echo Dot

Amazon’s small smart speaker, an everyday example of cloud AI answering by voice in a second or two.

Aglena Best Of product picks

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top
Aglena