The agreement scale doubles, far exceeding the level disclosed in May

As Anthropic moves toward an IPO, a regulatory filing has brought its computing power transaction with SpaceX to the forefront. Materials obtained by Reuters show that both parties signed a cooperation agreement with an upper limit of 84.5 billion U.S. dollars, under which Anthropic will purchase AI computing power for data centers from SpaceX, owned by Musk, over the next few years. This is double the figure known to the public a few months ago.

This deal first emerged this year in May. At that time, SpaceX mentioned the collaboration with Anthropic for the first time in its IPO filing, stating that it covered computing power for two data centers, Colossus 1 and Colossus 2, aiming to monetize idle GPU resources. According to the previous statement, Anthropic would pay 1.25 billion U.S. dollars per month from May 2026 to May 2029, with lower rates in the first two months due to production ramp-up, totaling approximately 42.5 billion U.S. dollars. The newly exposed prospectus now reveals the actual scale—84.5 billion U.S. dollars, about twice the previous number.

Musk Expands "Computing Super Factory," Anthropic Keeps an Exit Route

For SpaceX, this revenue comes at a crucial time. Musk is currently pushing forward AI infrastructure expansion, and Colossus 2 is called a "computing super factory," planning to bring more than one million GPUs online by the end of the year, most of which will use the NVIDIA Blackwell architecture. After integrating xAI in January, SpaceX is not only building rockets and laying satellite networks but also engaging in computing power rental business, with Anthropic being one of the most important external clients.

The agreement still has flexibility: most terms allow Anthropic to terminate the agreement with 90 days' notice, and even if long-term arrangements are locked in, they can adjust strategies according to market conditions.

A broader context is the expensive competition. Anthropic expects to invest at least 518 billion U.S. dollars in AI infrastructure within ten years, with partners including SpaceX, Amazon, Alphabet, and Microsoft. When training costs reach unprecedented heights, securing GPU access means holding the entry ticket for the next generation of AI. Data centers and computing power are becoming the hardest strategic assets on the global technology landscape.