Elon Musk has confirmed that SpaceX closed $40 billion of mostly callable debt to buy Nvidia GPUs, with the chips due to arrive in the first half of 2027, according to a TIKR report that surfaced through Yahoo Finance. Investor Jim Chanos read the same news and asked the question I would have asked, had anyone invited me to the meeting: how much of the 10 gigawatts SpaceX is building in 2027 will sit in data centers on the ground.
There is a lot of vocabulary packed into those two sentences, so let me slow down. A GPU, pronounced letter by letter, is a graphics processing unit. It was designed to draw video game explosions and turned out to be unreasonably good at the arithmetic that trains and runs AI models. Nvidia makes the ones everyone is fighting over, and buying them by the tens of billions of dollars has become the standard opening move for anyone who wants to be taken seriously in AI.
Callable debt is the part that made me put down my coffee. A callable loan or bond gives the borrower the right to repay it before the scheduled maturity date, which is useful if cheaper money shows up later. TIKR reports that most of the $40 billion is structured that way. Who lent it, at what interest rate, and against what collateral are all unknown.
Now the gigawatts. A gigawatt is a billion watts, roughly the output of a large power plant, and the AI industry has taken to measuring data centers by how much electricity they can swallow rather than how many chips they hold. Ten gigawatts is the kind of number that gets electric utilities to answer the phone on the first ring. It is also, to be clear, a power figure and not a chip count, so the $40 billion of GPUs and the 10 gigawatts of capacity are two different measurements of the same ambition.
The word that matters in Chanos's question is ground. SpaceX is a rocket company, and his phrasing points at the possibility that some of that capacity is not meant to stay on Earth. Whether any of it would go somewhere other than terrestrial sites, and how much, is unknown. TIKR's headline has Chanos calling the whole thing "a pretty material $500B question," and what that $500 billion figure counts is also unknown. I would treat it as his framing rather than a measured number.
Here is why the location question is not just a curiosity. A GPU that arrives in the first half of 2027 needs a building, a cooling system, and a grid connection waiting for it, or it sits in a crate depreciating. If SpaceX intends to run those chips in conventional data centers, it is competing with every hyperscaler for the same sites, transformers, and interconnection queues. If it intends something else, then the engineering questions multiply faster than I can list them, and the timeline could depend on hardware that does not yet exist.
So the bet, as far as the available reporting shows it, is this: borrow $40 billion now against chips that arrive next year, and have 10 gigawatts of somewhere ready to plug them in. The chips are the easy part. Somewhere is the whole story.
Leave a Reply