Space Exploration Technologies Corp.’s IPO registration statement, filed with the US Securities and Exchange Commission on May 20, revealed that artificial intelligence lab Anthropic agreed to pay $1.25 billion a month through May 2029 for AI compute capacity, according to reports that reviewed the document. The agreement could approach $45 billion if it runs its full term, but reduced fees apply during a May–June 2026 ramp and either party has the right to terminate with 90 days’ notice.
The filing put a hard number on a partnership that Anthropic and Elon Musk’s xAI, operating under the name SpaceXAI, announced two weeks earlier. On May 6, the pair said Anthropic gained access to all compute capacity at SpaceX’s Colossus 1 data centre — more than 300 megawatts and over 220,000 Nvidia H100, H200 and GB200 GPUs — and that the added horsepower would help raise usage limits for Claude Code, the Claude API and paid Claude Pro and Max subscribers.
The two sides also said Anthropic had expressed interest in developing multiple gigawatts of orbital AI compute capacity, a vision that remains aspirational and not a binding agreement. No timeline was given for such a project.
The S-1, however, showed the arrangement was larger than the initial disclosure suggested. The cloud services agreements cover not just Colossus 1 but also Colossus II, an expansion project flagged in the filing, though precise capacity figures for the second site were not independently confirmed. The monthly payments begin with reduced fees during the two-month ramp, which the reports did not detail. Either Anthropic or SpaceX can walk away with three months’ notice, meaning the headline figure assumes the contract survives nearly three years without early cancellation.
Anthropic’s spending dash reflects acute compute scarcity across frontier AI. The same lab has publicly committed to up to 5 gigawatts each of compute capacity with Amazon and with Google via Broadcom, a $30 billion Azure commitment with Microsoft and Nvidia, and a $50 billion infrastructure build-out with Fluidstack. Those agreements amount to a multi-gigawatt expansion strategy, and the SpaceX contract, the company has stressed, is diversification, not a pivot away from hyperscale suppliers. Unlike those multi-year infrastructure pledges, the SpaceX contract’s 90-day termination gives Anthropic unusual flexibility to adjust its capacity mix.
The total theoretical value is contingent. Bloomberg characterised the deal as “nearly $45 billion,” while The Information said it was “up to $40 billion.” With the 90-day escape hatch, SpaceX cannot book the full sum as guaranteed revenue, and Anthropic retains flexibility to reallocate spending if its capacity needs shift or if other projects come online. The reports also did not specify the size of the fee reduction during the ramp, making it difficult to gauge Anthropic’s total outlay even if the contract runs its course. The S-1’s full text was not directly reviewed by Graupel; the financial terms are drawn from credible secondary reports citing the filing, and a reporter should pull the document from SEC EDGAR before publication. It was not immediately clear from those reports whether the exact legal counterparty is SpaceX, xAI or SpaceXAI. How Anthropic plans to finance the monthly payments and the effect on its cash position were not addressed in the filing or reports.
For SpaceX, the disclosure supplies a marquee anchor tenant as it courts public-market investors, demonstrating that AI compute can be a standalone business line independent of Musk’s AI ambitions at xAI. For Anthropic, the deal underscores how far labs will go to lock in capacity — even from an organisation that houses a direct competitor — while the termination clause keeps an exit ramp open if the multi-gigawatt strategy starts to deliver.