Musk's Most Expensive Roommate

How SpaceX leased its biggest supercomputer to the competitor that's beating it, and what the deal reveals about who actually wins the AI race

Author

Justin Wright
May 11, 2026 • Est. Reading Time: 5 minutes

Happy Monday!

Three months ago, Elon Musk posted on X that Anthropic was "evil," "misanthropic," and that the AI lab hated Western civilization. Last Tuesday, he leased them his most valuable compute asset: the entire Colossus 1 supercomputer.

The deal gives Anthropic access to all of SpaceX's Colossus 1 data center in Memphis. That is 220,000 Nvidia GPUs, more than 300 megawatts of capacity, coming online within the month. Anthropic immediately doubled Claude Code's rate limits for paid plans, removed peak-hour usage caps, and raised API limits for Claude Opus models.

This is week three of a story we have been tracking. Three weeks ago, Musk paid $10 billion for a call option on Cursor because Grok could not crack the AI coding market. Two weeks ago, OpenAI missed its own revenue targets while Anthropic passed it at $30 billion in annualized revenue. Now, Musk is leasing his supercomputer to the company that's beating both of them. The pattern is not subtle: when you cannot win the product war, become the landlord.

SpaceX signed a deal giving Anthropic all of the compute capacity at Colossus 1, the supercomputer Musk built to train Grok. Grok generates less than $1 billion in annual revenue; Anthropic is tracking toward $40 billion. The deal generates $3-4 billion annually for SpaceX with $2.5 billion in profit, and positions the company as a hyperscaler ahead of its June IPO. Musk posted that SpaceX "reserves the right to reclaim the compute" if Anthropic's AI harms humanity. One of three frontier AI labs now runs on infrastructure controlled by a competitor's CEO.

TL;DR

Why Musk Is Renting Out His Supercomputer

Colossus 1 was built in 2024 to train Grok, but Grok has not filled it. According to New Street Research analyst Antoine Chkaiban, Grok generates less than $1 billion in annualized revenue. Anthropic is on track for more than $40 billion. The math is simple: Musk has too much compute and not enough demand. Anthropic has the exact opposite problem.

Chkaiban estimates the deal will generate $3 billion to $4 billion in annual revenue for SpaceX, with more than $2.5 billion in cash profit. The margins seem extreme, but the data center is already built. The capital expense is now a sunk cost. The only meaningful operating costs are electricity and staffing.

"He's not going to want multiple billions of dollars of GPUs sitting idle," Chkaiban told Fortune. "It's a very good business decision."

But filling empty racks is only half the story. SpaceX filed for a June IPO targeting a $1.75 trillion to $2 trillion valuation. The Anthropic deal gives the prospectus something it lacked: a marquee AI customer for a credible cloud business. A SpaceX that competes with AWS earns a hyperscaler multiple, not a rocket company multiple. Alphabet, Microsoft, and Amazon trade at roughly twice the forward earnings of Boeing and Lockheed Martin.

The Kill Switch Nobody Is Talking About

In a reply on X after the announcement, Musk wrote that SpaceX "reserves the right to reclaim the compute" if Anthropic's AI "engages in actions that harm humanity." The clause was not in the formal press release, and it is unclear whether it appears in the actual contract.

If enforceable, it gives Musk a leash on one of three leading AI labs while he simultaneously sues OpenAI's leadership in federal court. That is a lot of power he did not have two weeks ago.

Andrew Moore, former head of Google Cloud AI and now CEO of defense AI startup Lovelace AI, put it directly: "He who controls the data center really does control the application of artificial intelligence right now."

Gene Munster, managing partner at Deepwater Asset Management, put the odds the deal still exists in two years at 80%. The other 20% is a bet on Musk himself. "What makes it unique is Elon's history. He can change his mind."

Anthropic's Compute Empire

The SpaceX deal is one piece of a compute strategy that is incomparable in scale. Anthropic now has infrastructure agreements with Amazon (5 GW, $100B+ over ten years), Google and Broadcom (5 GW starting 2027), Microsoft and Nvidia ($30B of Azure capacity), Fluidstack ($50B in US infrastructure), and now SpaceX (300+ MW, all of Colossus 1).

The company also expressed interest in developing "multiple gigawatts of orbital AI compute capacity" with SpaceX. Data centers in space is now a line item in a compute agreement between two of the most valuable private companies on earth.

This multi-cloud strategy is deliberate. Anthropic trains Claude on AWS Trainium, Google TPUs, and Nvidia GPUs. No single provider controls enough of the stack to become a chokepoint. That diversification matters more now that one of those providers potentially has a clause in the contract allowing them to change their mind on a whim.

The Three-Week Pattern

Step back and look at what has happened in three weeks. Week one: Musk paid $10 billion for Cursor because Grok could not build a coding product developers would use. Week two: OpenAI missed revenue targets while Anthropic passed it at $30B ARR spending four times less on training. Week three: the company winning the product war is renting compute from the company losing it.

Musk built one of the world's largest supercomputers to train Grok; now it runs Claude. He is earning $2.5 billion in annual profit from the competitor that beat him. The irony is uncomfortable but the logic is sound. If you cannot monetize your own models, monetize the infrastructure they run on. The landlord collects rent regardless of which tenant wins.

What This Means for Practitioners

For AI teams, the immediate benefit is real. Doubled Claude Code rate limits and removed peak-hour caps mean more capacity for developers who depend on Claude daily. If your team has been hitting Opus rate limits, those constraints just loosened considerably.

For anyone evaluating infrastructure risk, the kill-switch clause matters. Even if it is not in the contract, Musk posting it publicly signals the power dynamic. Anthropic has contingency plans, and Moore predicted fallback options within three to twelve months. But right now, a meaningful share of Claude's capacity runs on hardware controlled by a competitor.

For investors, this deal is the SpaceX IPO proof point. The company is positioning as the fourth hyperscaler, capturing 30%+ margins that AI labs currently pay to AWS, Azure, and Google Cloud. Whether the market buys that narrative at $2 trillion is the biggest pricing question of the summer.

The Bottom Line

Musk called Anthropic evil in February. In May, he became their compute landlord for $3-4 billion a year. The supercomputer he built for Grok now powers Claude, because Grok could not generate enough revenue to justify keeping it.

Three weeks, one shared lesson: the product is the moat. And when the model cannot win, the smartest play is to rent the infrastructure to the company whose product can. Musk is not winning the AI race in the traditional sense, but he is making sure he gets paid by whoever does.

In motion,
Justin Wright

If the company that called your competitor evil three months ago is now your compute landlord with a self-declared kill switch, how diversified does your infrastructure actually need to be, and what does it mean for the industry that one person now holds leverage over two of the three frontier AI labs?

Food for Thought

If you haven’t listened to my podcast Mostly Humans: An AI and business podcast for everyone yet, new episodes drop every week!

Episodes can be found below - please like, subscribe, and comment!