SpaceX pays twenty times Cursor's revenue to restock the talent xAI lost

Twenty times revenue is the figure worth holding onto: $60 billion of purchase price against the $3 billion of annual recurring revenue Cursor passed in early 2026 ^[1][8][13]. For a subscription business that got from four MIT dropouts in 2022 to an enterprise B2B book in four years ^[7], that multiple is arguable on ordinary software logic. But the buyer's stated shortage was talent, not revenue, which makes ordinary software logic the wrong lens ^[4].

Divide the price by the eleven xAI co-founders who walked out at once in late March 2026 ^[3] and you get roughly $5.5 billion a head ^[18], a deliberately silly number that still tracks the sequence the report lays out: the computing resources were in place, the talent that turned them into software had gone ^[4], and about five months later the parent of an AI coding agent was a wholly owned SpaceX subsidiary ^[1][14].

What the money buys operationally, in the framing the report offers, is an agent that opens the relevant files, edits the code and debugs on a natural-language instruction rather than autocompleting the next line ^[9], aimed at orbital calculations for Falcon 9, Starship's flight control systems, and a Starlink constellation of more than 10,000 satellites whose interconnected software is already past the point where people can check it one item at a time ^[15]. Set the price against that constellation and Anysphere cost SpaceX under $6 million per satellite now in orbit ^[19]. There is a duller second reason to own the tooling rather than rent it, which is that on defense work an outside vendor seeing core internal source code is exposure ^[17].

The counter-thesis sits in the same report. Forty-five people have left Cursor since April, mainly from product and engineering ^[10], and the senior list runs from the head of reinforcement learning to the general counsel, taking in design, brand and the senior vice president for global strategic accounts ^[11]. If the goal is the one Musk gave Cursor employees directly, that Grok is behind and has to catch Anthropic ^[12], then the reinforcement learning departure is the one that bites, because an editor is a distribution surface and RL is the training lever. My read is that the $60 billion functions as a hiring bill with a product attached, not a product purchase with a hiring bonus attached.

Several endings are available. Cursor works as internal tooling and the return shows up as launch vehicles and satellites that were never lost, which is a real payoff that never appears in a revenue line ^[16]. Or the absorption into SpaceX's existing organisation ^[6] keeps costing the engineers who were the asset, and the buyer is left holding a $3 billion subscription book at twenty times ^[13]. Or Anthropic's IPO, which the report puts as early as next month and expects to fund a widening lead ^[5], settles the frontier question before any coding agent gets the chance to.

The test of the scarce-input thesis is cheap to run: $60 billion would buy an enormous amount of compute if compute were still the binding constraint. SpaceX spent it on twenty-somethings and their harness instead.