Analysis
Anthropic's confidential IPO prospectus discloses compute agreements with SpaceX worth up to $84.5 billion through 2029, Reuters reported Sep 29 after reviewing the filing -- nearly double the roughly $45 billion figure implied by SpaceX's own earlier IPO paperwork. The agreements give Anthropic access to Nvidia-based hardware capacity run by SpaceX's compute infrastructure business, a relationship that has grown well beyond what either company had previously signaled.
The SpaceX number is a fraction of a much bigger disclosure: Anthropic's prospectus separately shows at least $518 billion in total infrastructure and compute spending obligations over the next decade, alongside nearly $42 billion in net losses for 2025. Roughly 80% of that $518 billion is non-cancelable -- minimum commitments Anthropic has to pay whether or not it actually uses the capacity, the kind of fixed-cost structure that amplifies both the upside and downside of a demand forecast miss.
The SpaceX agreements are the exception to that pattern. Unlike most of the other infrastructure deals in the $518 billion total, the SpaceX contracts can mostly be cancelled with 90 days' notice -- flexibility Anthropic doesn't have on the bulk of its other compute commitments. That distinction is likely to matter to IPO investors modeling downside scenarios: a company that can walk away from a chunk of its compute spend on a quarter's notice carries different risk than one locked into a decade of fixed payments.
The disclosure also underscores how central SpaceX has become to the AI infrastructure buildout beyond its core rocket and satellite business -- a role that's drawn less public attention than Anthropic's own compute deals with Google, Microsoft and Nvidia, but now rivals them in size.