AI model evaluation company PatronusAI co-founder Anand Kannappan posted on X that the outside world is misreading the SpaceX-Cursor deal as an acquisition. At its core, he argues, it's an experiment: is developer behavior data the real bottleneck for coding models?
https://twitter.com/anandnk24/status/2046784367456841969
According to his analysis, xAI's Grok has been trailing behind Claude Code and Codex in coding ability. Cursor, meanwhile, has the world's largest collection of developer interaction traces. The deal lets Cursor train its own Composer model on SpaceX's Colossus supercomputer, while xAI trains Grok on the same data — together, they'll test whether Cursor's data can make a difference.
Kannappan sees the option structure as a direct price on that uncertainty: if the data proves valuable, SpaceX buys Cursor and its pipeline for $60 billion; if not, it pays $10 billion as an experiment fee. His conclusion: a quasi-public company just put a price tag on the question "Is the developer trace the scarce resource for coding agents?" — $10 billion for the experiment, $60 billion if the answer is yes.