In early May 2026, a seemingly contradictory deal surfaced in the AI world: Anthropic, the maker of Claude, rented the entire compute capacity of SpaceX’s Colossus 1 data center. The contradiction? Anthropic’s Claude is the direct competitor of Grok, which SpaceX owns. A rival renting a data center from a rival looks, to an outsider, like Coca-Cola buying production lines from Pepsi.

But the deal is far from foolish; it precisely reveals an industry truth: on today’s AI battlefield, compute itself has become a business you can buy and sell. This piece walks you through how this “friend and foe at once” works. To get to know SpaceX and its AI map first, see What Is SpaceX?.

First, what the deal looks like

Start with the confirmed parts. On May 6, 2026, xAI (now SpaceX’s AI division SpaceXAI) and Anthropic jointly announced that Anthropic would use the entire compute of the Colossus 1 facility, at a scale of over 220,000 GPUs and over 300 megawatts of power, mainly to support Claude’s service capacity.

On price, things are relatively clear: SpaceX’s IPO filing (the S-1) states about US$1.25 billion per month, and Anthropic confirmed the figure to the media. Because it is written into a listing document and backed by one of the parties, this monthly figure is relatively solid.

One key point, however, is in dispute and worth reading side by side:

  • The S-1 text: the monthly fee runs through May 2029, with a 90-day mutual notice termination clause. On that basis, the total could reach tens of billions of dollars.
  • Musk’s account (on X in late May): it is actually a “180-day short lease,” with 90-day notice termination thereafter, and the short term was SpaceX’s own request, not a long-term commitment.

Both sides acknowledge the 90-day termination right; the difference is whether the base commitment is six months or three years. As of now, neither SpaceX nor Anthropic has publicly clarified the gap. So when you see figures like “US$45 billion” or “locked through 2029,” remember they rest on the longer version and may not be settled.

What Colossus is, and why it can be leased out

To understand the deal, you first need to know what makes Colossus special.

It is a large AI compute cluster that xAI self-built in Memphis, Tennessee; the first generation, Colossus 1, is said to have been built at remarkable speed and is equipped with over 220,000 GPUs. The key words are “self-built”: most AI companies rent compute from cloud providers, but xAI built its own facility and owns the hardware itself.

Precisely because the facility is its own asset, xAI can lease the whole thing to someone else. Had it merely rented from the cloud, there would be nothing to sublet. That is why players who can build data centers hold special leverage in this AI race. For how the whole compute supply chain works, read on at The AI Hardware Supply Chain, End to End.

Why Anthropic would rent from a rival

The answer is very practical: there is not nearly enough compute to go around.

Over the past two years, AI models have grown ever larger, and demand for GPUs and data centers far outstrips supply. Advanced-chip lead times run into quarters, and the big cloud providers locked up capacity early. For a fast-growing AI company, whether it can secure enough compute is often more urgent than whether the supplier is a rival.

And Anthropic does not rent from SpaceX alone. It signs compute contracts with several providers at once, spreading orders across suppliers, and renting from a competitor is just one of them. Placed in that context, the deal stops being strange: when compute is a scarce resource, you grab it where you can, and whether the source is a rival is secondary.

Why SpaceX would rent to a rival

From SpaceX’s side, there are two very practical reasons too.

First, monetizing capacity it isn’t fully using. Per reporting, Colossus 1 mixes different GPU generations and was inefficient for training SpaceXAI’s own latest models, so SpaceXAI moved its main training to a newer-generation facility. With this one’s capacity freed up, leasing it out beats leaving it idle; the hardware cost was already spent, so extra rental is nearly pure margin.

Second, supporting the IPO story. SpaceX is filing to go public, and if it can show that compute can be sold as a service, it is no longer just a rocket and satellite company but one with a “selling compute” revenue line resembling a cloud giant’s. For a listing narrative chasing a high valuation, that helps. For how to read this line within the overall financials, see How to Read SpaceX’s IPO Financials.

The industry truth this deal reveals

Put the pieces together and a bigger picture appears: AI’s “compute layer” is becoming a business separate from the “model layer.”

At the model layer, Grok and Claude are open rivals, each chasing users and comparing capability. But at the compute layer, the underlying data centers, power, and chips are scarce resources shared by all; whoever builds them first and has spare capacity can sell it, and even rivals are potential customers.

This structure has appeared in tech before: phone brands fight each other, while a parts maker may supply all of them. In this AI wave, compute is that key part. And it brings a new risk: when your compute is rented from a rival, the other side in theory holds the leverage to “cut off your utilities.” Musk has said that if the other party’s AI does something that harms humanity, SpaceX reserves the right to reclaim the compute. That kind of dependency is a side effect of this new industry structure.

In short

Anthropic renting Colossus from rival SpaceX is not a misstep by either side but a rational choice under a compute shortage. It makes one thing clear: on the AI battlefield, models can be enemies while compute can be a business, and whoever can build the data centers has even rivals knocking at the door.

To place this facility back in the whole rocket empire’s context, see What Is SpaceX?; for how Grok and Claude spar at the model layer, a future piece will go deeper.