The number was sitting in plain sight, on a page almost nobody will read.
When SpaceX filed its S-1 with the Securities and Exchange Commission on May 20, the document ran to hundreds of pages of risk factors, segment accounting, and rocket economics. Buried inside was a single line that priced the strangest compute deal of the year. Anthropic, the company Elon Musk had publicly branded "evil" two weeks earlier, will pay roughly $1.25 billion every month to rent the supercomputer his engineers built to train Grok.
The payments run through May 2029. Add them up and the contract is worth more than $40 billion to Musk's side of the table, with a discounted rate for the first two months while the cluster finishes ramping. Either party can walk with 90 days' notice.
For data scientists and ML engineers, the headline is not the soap opera between two rival founders. It is the price tag. This filing is the clearest public look yet at what frontier-scale compute actually costs, and it confirms a quiet truth of the 2026 AI economy: the constraint is no longer talent or even data. It is power, GPUs, and the buildings to hold them. Anthropic just signed one of the largest compute leases ever disclosed, and it signed it with a competitor.
The Filing Put a Real Number on Frontier Compute
The deal itself was not a secret. Anthropic secured the output of the Colossus 1 data center earlier in May, taking the entire 300 megawatts of capacity at the Memphis, Tennessee site that xAI had built to train its own models. What nobody knew until the S-1 landed was the cost.
SpaceX wrote that the arrangement "allows us to monetize unused compute capacity in our infrastructure." The accounting is blunt: $1.25 billion a month, about 15 billion dollars a year at full run rate. Colossus 1 went fully operational in December 2024 after a 122-day build that Musk's team treated as a point of pride. Eighteen months later, the company is renting the whole thing to the rival it spent the spring attacking.
The 90-day termination clause matters more than it looks. A compute contract that either side can exit in a quarter is not the decade-long anchor tenant that data center economics usually depend on. It is a stopgap, useful to both sides for now, and unwound the moment either one's needs change.
The Same Document Shows Why Musk Needed the Deal
A few pages over, the filing answered the obvious question: why would the company that built the world's largest AI supercomputer hand it to a competitor?
Because the numbers underneath are ugly. xAI, now folded into SpaceX as the SpaceXAI division after a February all-stock acquisition, lost $6.4 billion from operations in 2025. Spending grew faster than sales, and the filing signals it is not slowing down.
| xAI operations | 2024 | 2025 |
|---|---|---|
| Revenue | $2.62 billion | $3.2 billion |
| Operating loss | $1.56 billion | $6.4 billion |
Revenue grew, but the loss more than quadrupled in a single year. A company hemorrhaging at that rate does not casually idle a 300-megawatt supercomputer. It finds a tenant.
The product side is no kinder. Grok, the assistant all that compute was supposed to power, is losing users. Mobile daily active users fell from 13.9 million in March to 12.2 million in April, a drop of nearly 13 percent in a single month. In the United States the decline was steeper, closer to 16 percent. By April, Grok had slipped to fifth place among AI assistants, passed by Claude, Gemini, and DeepSeek.
That is the real story the S-1 tells. xAI built Colossus for a level of Grok demand that never showed up. The servers were sitting partly idle. Renting them to Anthropic turns an overbuilt liability into a revenue line, one large enough to matter as SpaceX walks toward a public listing.
The Neocloud Trade Is Now an Industry Strategy
The label for what xAI is doing has a name: a neocloud, a company that builds AI infrastructure for itself and then sells the spare capacity to others like a cloud provider. Until recently, AI labs picked one lane. They either built data centers to train their own models or rented from the hyperscalers. Doing both at once was rare.
SpaceX described the approach as a "dual monetization strategy" and told investors it expects to sign more contracts like the Anthropic one. The logic is straightforward once the demand math turns against you.
| Approach | Who does it | The bet |
|---|---|---|
| Build for yourself | Most frontier labs historically | Your own model demand fills the cluster |
| Rent from hyperscalers | Smaller labs and startups | Avoid capital risk, pay a margin |
| Neocloud (build, then resell spare capacity) | xAI now, increasingly others | Offset overbuild by renting idle GPUs to rivals |
For practitioners, the neocloud shift has a practical edge. When even direct competitors are willing to sell each other capacity, raw GPU access becomes more of a commodity and less of a moat. The differentiator moves up the stack, to the quality of the models, the data, and the engineering, rather than to who happened to pour the most concrete in 2024.
The Other Side Sees a Smart Use of Idle Iron
Not everyone reads the deal as a distress signal. SpaceX framed it as disciplined capital management, and the math supports the spin. The company has directed roughly $13 billion toward AI infrastructure and data centers. A contract worth north of 40 billion dollars over its life could cover that outlay several times, turning a speculative build into a financed one ahead of an IPO that analysts peg near a 1.75 trillion-dollar valuation.
Bulls also point out that Anthropic gets something real and scarce: 300 megawatts of working, GPU-dense capacity it did not have to wait years to build, at a moment when compute is the binding constraint on every frontier lab. For a company reportedly targeting its own IPO later in 2026, locking in capacity now is worth paying a premium for, even to a rival.
The skeptics' counter is the termination clause and the user numbers. A 90-day exit window suggests neither side sees this as permanent, and a supercomputer that has to be rented out to pay for itself is, by definition, a supercomputer its owner could not fill.
The Bottom Line
Strip away the feud and the filing says something simple. Frontier AI now costs more than the people building it can spend alone, and the bill is so large that yesterday's enemies will rent each other the means of production to keep the lights on.
Anthropic pays $1.25 billion a month for compute it did not have to build. SpaceX collects revenue on a cluster its own product could not justify. Both companies look stronger for it on the way to the public markets, which is precisely why the deal exists. The open question is what happens in 90 days, and the next 90 after that, when the demand math shifts again for whichever side blinks first.
Musk spent the spring calling Anthropic a danger to the world. His IPO filing calls it a customer.
How the Deal Came Together
Sources
- Anthropic will pay xAI $1.25B per month for compute — TechCrunch, May 20, 2026
- xAI burned $6.4B last year, SpaceX's IPO filing shows why the spending is far from over — TechCrunch, May 20, 2026
- SpaceX IPO filing reveals Anthropic set to pay Musk's firm $1.25bn a month — Data Center Dynamics, May 21, 2026
- SpaceX's $45 Billion Anthropic Deal Could Fund Its AI Buildout Three Times Over — Benzinga, May 20, 2026
- The SpaceX IPO filing is filled with AI bets, Starship dreams, and Elon Musk at the center — TechCrunch, May 20, 2026
- Elon Musk's Grok loses users in 2026 as rivals surge ahead — AAJ / Forbes data, May 2026
- SpaceXAI — Wikipedia (background on the xAI–SpaceX merger), May 2026