The SpaceX AI chips deal is now on the record rather than in the rumour column. Elon Musk posted on X on Sunday 25 October that financing had closed, and a regulatory filing the next day set out a binding commitment for roughly $40bn of GPU financing with an entity managed by Apollo.
Musk later described it as “primarily consisted of callable debt for graphics processing units”. The hardware is expected to land in the first and second quarters of 2027.
How a launch company ended up making an AI chips deal
The company’s centre of gravity has moved. SpaceX’s AI business, which it calls Super Intelligence, is now its biggest unit, with third-quarter revenue of $7.56bn against $813m in the same quarter a year earlier. That is roughly ninefold growth, and it is more than Starlink and the rocket business produce together.
Musk’s argument for why lenders would write a cheque this size is that SpaceX is investment grade and has Starlink and launch revenue that does not depend on anyone’s AI spending. It is a collateral argument as much as a credit one.
Growth of that shape changes what the company is. A division that went from $813m to $7.56bn inside four quarters is not an expansion of the existing business, and the GPUs are the input it consumes.
A deal and a warning in the same breath
Musk framed the financing as the exception rather than the weather. CNBC paraphrased him saying debt markets have largely closed to companies without meaningful revenue outside AI data centres, and that the 200-plus neocloud providers are heading for widespread failure and consolidation down to a handful.
The week supported the reading. Microsoft drew $115bn of orders for a $25bn bond sale, while companies without that kind of underlying business were not raising anything. Capital is plentiful and extremely selective at the same time.
There is also a structural point in Musk’s framing. Lenders are no longer pricing AI exposure as one category; they are splitting borrowers into those who can service debt from unrelated revenue and those who cannot, and only the first group is being served.
The parts still missing
A binding commitment with a named counterparty is a strong document, and it is why this is reportable at all. What has not been published is the borrowing entity, the repayment schedule, any guarantees, or the final pricing. Earlier coverage described a proposed structure and indicative rates, and separate reporting suggested pricing talks were still live after the close was announced.
Those are not small gaps on a $40bn obligation, and they are the figures that will determine whether this looks prudent in three years.
Why the 2027 delivery date matters
Chips arriving in the first half of 2027 meet a market that is being actively rearranged. By then Amazon plans to sell its own accelerators to outside buyers and several non-Nvidia options will have reached general availability on major clouds. Committing now is a judgement that demand for compute stays ahead of all of that new supply. CNBC reported the filing and the segment numbers.
