Everyone thinks SpaceX is building an AI compute business to compete with Amazon and Google. Look at the numbers.
The company just signed a deal to supply $1.25 billion a month in compute power to Anthropic through May 2029. That sounds like scale. Elon Musk says they're in discussions with other companies. The S-1 filing mentions "AI Infrastructure" as a market worth $2.4 trillion. The obvious question is whether SpaceX can become a hyperscaler.
But the more interesting question is what they actually built. The way to understand this isn't to look at the market size. It's to look at what happened first.
SpaceX merged with xAI in February, creating a combined entity valued at $1.25 trillion. They built Colossus 1 in Memphis - 300 megawatts of capacity, roughly 200,000 GPUs - to train their own AI model, Grok. Then usage of Grok dropped. So they had all this compute sitting there.
Meanwhile, their AI operations were burning money. AI losses ballooned fourfold last year to more than $6 billion, driven by higher cloud costs and GPU depreciation. For the first quarter of this year, the losses more than doubled to almost $2.5 billion. They're filing for an IPO in June.
So they had overbuilt capacity, massive losses, and an IPO deadline. Then Anthropic needed compute. The deal lets SpaceX "monetize unused compute capacity in our infrastructure". Either side can terminate with 90 days' notice. Anthropic gets a discounted rate for the first two months.
This isn't a planned hyperscaler strategy. It's iterative discovery: build for your own needs, discover you can sell the excess, then call it a business. The S-1 even says the arrangement "allows reallocation of the capacity for our own internal initiatives if needed". This is capacity they might take back.
There's another contradiction that matters. SpaceX uses Google Cloud through its subsidiary, Starlink. They announced a partnership in 2021 where SpaceX installed Starlink ground stations at Google data centers. Now they're competing with Google Cloud by selling hundreds of megawatts of compute to Anthropic. You can't be both a customer and a competitor unless you're figuring this out as you go.
The orbital data center plans sound ambitious - space-based computing clusters powered by solar - but those are years away. Right now, the business is selling excess capacity from an overbuild.
I suspect the real test comes after the IPO. If this were a deliberate hyperscaler play, you'd see more than one major customer. You'd see longer-term contracts without 90-day outs. You'd see profits, not $6 billion in losses. What you have instead is a company that built something for itself, found someone else who wanted it, and is now calling it a business model.
The way to evaluate this isn't to ask whether SpaceX can compete with AWS. It's to ask when selling your excess capacity becomes a real business versus just financial cleanup. One data point: they have one customer paying $1.25 billion a month while losing $2.5 billion a quarter. The math suggests they need more customers, or lower costs, or both.
Watch what happens after June. If they sign more deals like this with other companies, maybe they discovered a real business. If not, maybe they just cleaned up their balance sheet for the IPO. The numbers will tell you which it is.