Anthropic disclosed agreements to pay SpaceX up to $84.5 billion for computing capacity through 2029, nearly double the roughly $45 billion in potential spending SpaceX had previously disclosed for the relationship.
The fuller scope emerged in Anthropic's confidential IPO prospectus filed September 29. The original agreement, announced in May, called for payments of about $1.25 billion per month and was structured for a three-year term. The newly disclosed agreements extend through 2029, though most can reportedly be canceled with 90 days' notice.
Anthropic is one of the largest independent developers of large language models and has been among the biggest consumers of GPU compute capacity in the industry. The company competes with OpenAI, Google's Gemini and other rivals for access to scarce training resources. Computing power for large-scale AI training typically costs tens of millions of dollars per month, and Anthropic's ability to secure long-term capacity commitments from infrastructure providers is central to its business plan.
SpaceX's Starshield division has been building out compute infrastructure, though the company is primarily known for launch services and satellite internet through Starlink. The compute agreements represent a significant revenue stream for SpaceX outside its core aerospace business. The deals give Anthropic contractual certainty around hardware availability while allowing the AI company flexibility to reduce commitments if training needs change.
The $84.5 billion potential value assumes all agreements run to their full term and Anthropic maintains spending at disclosed rates. At the stated $1.25 billion monthly figure, that annual run rate would total $15 billion per year. The prospectus indicates most agreements include early termination rights, which reduces the likelihood that SpaceX will collect the full amount.
Anthropic is pursuing a public offering and must disclose material contractual obligations in SEC filings. The broader compute market has become a focal point for AI infrastructure investors as training costs for frontier models continue to climb. Anthropic's disclosed spending with SpaceX is now larger than publicly disclosed compute agreements from most other AI labs, though the company also works with other providers including cloud platforms and chip makers.
At that scale, a single computing provider represents a material dependency. The 90-day cancellation clause allows Anthropic to exit if circumstances change. The number to watch is whether other AI developers disclose similarly large forward commitments in their own IPO filings, which would indicate whether the industry is moving toward long-term infrastructure exclusivity deals or maintaining optionality across multiple providers.
