Anthropic Is Paying Elon Musks SpaceX $15 Billion a Year. The Math Is Terrifying.
When the AI company that built its brand on safety and existential risk needs $45 billion in compute to stay competitive, something has to give.
Anthropic has committed to paying SpaceX $1.25 billion per month through May 2029, according to the companies joint SpaceX IPO filing, published this week. The total commitment: roughly $45 billion over three years, making Anthropic SpaceX's largest known AI compute customer. The deal gives Anthropic access to the Colossus 1 and Colossus 2 data center clusters, powered by more than 220,000 Nvidia GPUs, including H100s, H200s, and newer GB200 accelerators [Silicon Republic].
The numbers on Anthropic's side look impressive. The company told investors it expects Q2 revenue of at least $10.9 billion, more than double its $4.8 billion in Q1 [Reuters]. That would drive an operating profit of $559 million for the quarter. These are not the figures of a company in trouble.
But look at what it's paying for that growth.
SpaceX's AI segment lost about $2.5 billion in Q1, on revenue of $818 million [WIRED]. Anthropic's monthly payment of $1.25 billion is not merely buying GPU time. It is, by arithmetic, large enough to have materially offset SpaceX's AI division losses in Q1 — a relationship the filings describe as a customer contract, but whose scale raises questions about how cleanly the competing AI interests are separated. SpaceX confirmed in its filing that it is in discussions with other companies about offering AI compute as a service at scale [Axios]. The filing says the company believes its "dual monetization strategy provides multiple pathways to generate returns on invested capital." The xAI merger means SpaceX now controls infrastructure used by both xAI and its Anthropic customer — the degree to which those compute pools are separated or shared is not addressed in the public filing, and Anthropic declined to comment on governance or conflict-of-interest safeguards.
Either party can exit the deal with 90 days' written notice [Reuters]. That exit ramp is the detail that matters most. At $15 billion per year, Anthropic is making one of the largest infrastructure commitments in tech history, with a competitor who can walk away in three months. Or who can be walked away from.
The ownership structure makes this stranger still. Elon Musk, who controls SpaceX, has called Claude "misanthropic and evil" and accused the model of hating people of certain races and heterosexuals. Musk's SpaceX merged with his AI company xAI earlier this year. Anthropic, whose stated mission centers on building safe AI systems and whose founding team includes former OpenAI safety researchers, is now writing monthly checks to the infrastructure of Musk's competing AI ambition. Anthropic's board did not respond to questions about whether it formally assessed the Musk conflict before approving the deal.
Tom Brown, Anthropic's co-founder and chief compute officer, confirmed the deal's expansion this week, posting that the company would be "scaling up on GB200 capacity in Colossus 2 throughout June". He thanked Musk and the SpaceX team for "helping us find good homes for the Claudes." The personal warmth of the post is notable precisely because the structural relationship is so cold. Anthropic declined to comment on how conflicts of interest between its SpaceX partnership and xAI competition are governed. Whether the human relationships are cordial and the structural separation real are separate questions, and the filings do not answer the second one.
Anthropic's revenue trajectory is what makes this survivable. At $10.9 billion per quarter, $1.25 billion per month is roughly 34 percent of quarterly revenue going to a single vendor, one whose majority owner has publicly contemptuous views of Anthropic's product. The $559 million operating profit in the same quarter means the company is making money, but not enough to absorb this expense painlessly. Growth is paying for the infrastructure. If growth slows, the math tightens fast.
The deeper question is whether Anthropic had alternatives. Colossus is one of the largest GPU clusters in the world — and at the gigawatt scale Anthropic now requires for frontier model training, the practical substitutes are scarce. AWS, Google, and Microsoft can offer cloud compute, but not at the raw GPU density, procurement lead times, or contractual flexibility that a frontier lab burning through hundreds of millions of GPUs requires. Meta's $21 billion CoreWeave deal [Silicon Republic] shows the industry-wide pattern: the compute bottleneck is rewriting competitive dynamics, and whoever controls a gigawatt-scale cluster holds structural leverage over anyone who needs it.
That leverage is what makes the collapse scenario worth spelling out. If SpaceX exits the deal in 90 days — or if Anthropic needs to walk — there is no equivalent fallback at the GPU density and power scale the company now requires. Training runs mid-flight would be stranded. API rate limits for paying customers would collapse. The $10.9 billion quarterly revenue depends, in part, on infrastructure that can vanish on 90 days' notice. One AI infrastructure analyst who declined to be named cited the concentration risk as "structurally unusual" for a company Anthropic's size, noting that frontier labs have historically avoided single-vendor compute dependency at this scale precisely because the fallback options are so limited.
What the deal means for the competitive landscape is equally consequential. If the arrangement holds, SpaceX becomes the world's first vertically integrated AI infrastructure company — a launch company competing directly with AWS, at margins shaped by rockets and satellites rather than data center leases. If it doesn't hold, Anthropic faces a compute crisis with no equivalent fallback, and the SpaceX AI division returns to losing $2.5 billion a quarter. Either outcome reshapes the map: the industry is moving from a world where compute is a service you buy, to one where the physical substrate — the land, the power, the GPUs, the cooling — is the competitive moat, and whoever controls it makes the rules.
Anthropic is writing checks large enough to decide which version of that future gets built. Whether that is a sign of strength or a symptom of a deeper dependency problem is, as the numbers make clear, a matter of time.