SpaceX disclosed on May 20 that Anthropic will pay it $1.25 billion a month, roughly $15 billion a year, for AI compute through May 2029. The figure surfaced in SpaceX's S-1 registration with the SEC, the paperwork a company files ahead of going public. Total contract value clears $40 billion (about $41 million a day, if you enjoy that kind of arithmetic). The compute sits in the Colossus data centers near Memphis, Tennessee, which until a few weeks ago belonged to a different Elon Musk company entirely.
Because that is the wrinkle worth pausing on. Colossus was xAI's. Musk folded xAI into SpaceX this month, so the GPU clusters built to train Grok now sit on SpaceX's balance sheet, and the checks Anthropic writes land there too. A rocket company as one of the largest compute landlords in frontier AI? That's where we are.
Why rent compute from Musk?
The short version: Anthropic needed capacity now, and Colossus had it sitting half-used. Grok's user base never grew into the cluster xAI built. Anthropic has been buying every GPU it can get its hands on, stacking up compute deals with Amazon, Google, Microsoft, Nvidia and Broadcom. The catch is timing. Most of that capacity isn't expected online until late 2026 or 2027. SpaceX hands Anthropic roughly 300 megawatts and north of 220,000 Nvidia GPUs more or less immediately.
And Anthropic hasn't been shy about needing them. The company told users earlier this year that demand for Claude had put real strain on its infrastructure, denting reliability at peak hours. If you've watched Claude Code throttle you mid-session, there's your explanation.
Then there's the matter of who everyone is fighting. One framing, quoted by VentureBeat: "Elon's enemy is Sam. Dario's enemy is Sam. Enemy of my enemy is a compute partner." Crude. Also not wrong.
The compute wars make strange roommates.
The numbers, and what they leave out
$15 billion a year reads very differently depending on whose ledger you're holding. For SpaceX it nearly doubles annual revenue, which sits around $18 billion. For an AI segment that lost roughly $2.5 billion last quarter on a sliver of revenue, a locked-in multi-year customer is exactly the line you want near the front of an IPO prospectus.
For Anthropic the math is calmer than the headline number suggests. The company says its $30 billion run rate is up from about $9 billion at the end of 2025. So the SpaceX bill runs to roughly half of annualized revenue, not the 80% or more that some early summaries floated. Big. Not existential.
A few things the filing doesn't spell out. The reported terms let either side walk with 90 days' notice, which is loose for a deal this size and says something about how fast both companies expect the compute market to move. Pricing was reportedly discounted for the first two months while capacity ramps; the filing notes reduced payments for May and June. And the $15 billion covers Colossus 1. Whether the Colossus 2 expansion runs on the same terms is not clear from anything public.
Tom Brown, an Anthropic co-founder, announced that Colossus 2 piece on X, scaling up on Nvidia GB200 capacity through June and thanking Musk and SpaceX "for helping us find good homes for the Claudes." Make of the phrasing what you will.
What happens next
SpaceX is reportedly aiming for a NASDAQ debut around June 12 at a valuation somewhere between $1.75 and $2 trillion, and the Anthropic contract is carrying weight in that pitch. The open question is whether public investors read one dominant customer as ballast or as concentration risk. Anthropic gets its capacity in June while it waits on other suppliers due next year. I doubt this is the last odd compute pairing we'll see before the dust settles.
