DEV Community

Cover image for $43B of GPUs, unrecoverable in orbit: SpaceX's lenders balk
frank chu
frank chu

Posted on

$43B of GPUs, unrecoverable in orbit: SpaceX's lenders balk

SpaceX just sold more than $30 billion of bonds at an 8.3% yield — for paper rated A3, a notch above the company's own debt. The market is charging more for the "safer" paper than for SpaceX itself, and the reason is in the loan documents: lenders want the $43 billion of Nvidia chips behind it barred from ever leaving Earth.

That is the story the Financial Times reported on Friday, and it is the first time orbital compute has shown up not as a keynote slide but as a loan covenant.

What actually happened

Project Saturn is SpaceX's GPU-backed debt package, announced October 26 as a roughly $40 billion commitment to buy Nvidia's next-generation Vera Rubin chips. The structure is a lease, not a purchase: an Apollo-backed special purpose vehicle buys the chips and leases them back to SpaceX for five years, keeping the debt off SpaceX's balance sheet. I wrote about the announcement in October — the lease math, not the headline number, was the story then too.

On Thursday, more than $30 billion of the package sold to large investors. The biggest piece is a $25 billion investment-grade bond backed by SpaceX's five-year lease payments for Vera Rubin NVL72 systems. The terms, per the FT and the rating agencies:

  • Yield of about 8.3%, or 3.45 percentage points over Treasuries — versus roughly 7.2% on SpaceX's own 2031 bonds
  • Notes mature in 2032, but principal is repaid in under four years
  • Moody's rated them A3, one notch above SpaceX itself; KBRA gave A-minus, citing "hell-or-high-water" lease clauses — SpaceX pays quarterly whether the hardware works or not
  • Apollo and Goldman Sachs led the deal; BlackRock, Goldman Sachs Asset Management and KKR anchored it, with $4 billion from Pimco

Then the clause. Lenders sought a provision barring the chips pledged as collateral from being launched into space for orbital data centers. Their concern is exactly what it sounds like: if SpaceX defaults, chips in orbit cannot be reclaimed and resold. Whether the provision made it into the final documents is not clear. SpaceX says it plans to use these chips on Earth — Moody's says they will sit in US data centers majority-owned, controlled and operated by SpaceX — and SpaceX did not respond to the FT. Apollo, Goldman Sachs and Pimco declined to comment.

The math that matters

Three numbers in this deal tell you more than the orbit clause does.

1. The 1.1-point inversion. The new bonds yield 8.3%; SpaceX's own 2031 bonds yield about 7.2%. The SPV paper carries a higher rating (A3, one notch above the company) and still costs 1.1 percentage points more. Ratings say the collateral makes it safer. The market says the collateral is the problem. Both can be true: the rating agencies are pricing the lease payments, the bond market is pricing the chips.

2. The four-year clock. Principal is repaid in under four years on notes that mature in 2032, against a five-year lease. Lenders are not underwriting the chips' useful life — they are underwriting an exit before obsolescence eats them. A Vera Rubin system bought in 2027 is not a 2032 asset. The structure admits that.

3. The resale arithmetic. The whole premise of GPU-backed debt is that the collateral can be sold if the borrower stops paying. On the ground that is already a depreciating proposition — H100s now trade at roughly half their 2023 cost basis, per Compute Credit Research Index — and in orbit the resale value is not low, it is undefined. There is no repo market for a satellite. No auctioneer, no forklift, no second buyer. The orbit clause is lenders pricing in the difference between "worth less" and "worth nothing to us."

That is also why the "hell-or-high-water" clause is doing the real work in this deal. If the chips fail, SpaceX still pays quarterly. The lease payments are the collateral; the GPUs are the story told to justify the rating.

What this means for orbital compute

Musk has said SpaceX will launch Nvidia-powered computers into orbit, with the first Starmind satellites targeted for the third quarter of 2027, and Sequoia's Shaun Maguire argued this week that almost all inference compute will eventually be in orbit. Polymarket gives a qualifying orbital AI data center before 2028 about a 20% chance. The market reaction on Friday was skeptical in the other direction: SpaceX shares slid below their June IPO price in pre-market trading, and Blockonomi reports them down 6.76% to $129.07.

The honest comparison is Google. Suncatcher — the TPU satellite Google launched on October 1 — is self-funded, runs Google's own chips, and nobody had to convince a credit committee that a satellite is a recoverable asset. SpaceX has to buy its chips from Nvidia, its Terrafab chip plant is not built yet, and Starship still has not flown a fully reusable mission — Flight 15 caught the ship and lost the booster to the Gulf. Every one of those gaps is a reason a lender would rather the collateral stay on the ground.

So orbital compute now has a financing fork, and this deal drew it clearly. Either you self-fund it, like Google, or you pay a financing premium that prices in the fact that your collateral can fly away — literally. The 8.3% yield is the market's first quote on that premium. It is not a verdict on whether orbital data centers work. It is a verdict on whether anyone else will pay for them if they don't.

What is confirmed and what isn't

Confirmed: the deal size, pricing, ratings and investors, per the FT and the rating agencies. Unconfirmed: whether the orbit ban made it into the signed documents — that comes from people familiar with the talks, and the FT is explicit that it is unknown. Two numbers to keep separate: $43 billion is the whole debt package; $40 billion is the chip purchase announced in October. They are different scopes, not a discrepancy. And the H100 resale figure is per Compute Credit Research Index, not my own measurement.

Sources: Financial Times, via Dealroom (Nov 20, 2026); TipRanks (Nov 20, 2026); Benzinga via TradingView (Nov 20, 2026); Blockonomi (Nov 20, 2026); TechCrunch, "Chips can survive the space environment. Can their financing?" (Nov 20, 2026). All ratios in this piece are my own arithmetic from those figures: 8.3% − 7.2% = 1.1pp; $25B is the largest tranche of the $30B+ sold Thursday.

Top comments (0)