A couple of weeks ago, in an Oakland federal courtroom, Elon Musk took the stand under oath and spent three days testifying against Sam Altman in the trial of his lawsuit against OpenAI. A fortnight later, on Wednesday, May 6, his company SpaceX signed an agreement to lease Anthropic its entire Memphis supercomputer.
Yes, that Anthropic. The one Claude makes. A direct competitor of xAI, Musk’s own AI company. And if that wasn’t enough, he’s leased the data center he originally built for xAI to stand up to OpenAI.
Let’s see, this has to be told calmly.
It’s one of those moves that when you first read it, you think there’s a mistake in the headline. But there isn’t. It’s exactly what it sounds like. And to my mind, it’s one of the most telling things to happen in the AI industry so far in 2026.
The agreement
Let’s take it one step at a time. The agreement announced on May 6 gives Anthropic full access to Colossus 1, the data center that SpaceX operates in Memphis, Tennessee. We are talking about more than 220,000 Nvidia GPUs between H100, H200, and the new GB200, totaling around 300 megawatts of computing capacity that will come online in less than a month. To give you an idea, 300 megawatts is the power consumed by an entire small city, just to run Claude.
Anthropic has taken advantage of the announcement to to double the limits Claude Pro and Claude Max subscribers had on Claude Code, its programming assistant. It has also removed the “peak hour” restrictions that cut off access when there was congestion. And in the API, the lowest tiers have seen a beastly increase: +1,500% in input tokens per minute and +900% in output for Tier 1.
Come on, it’s not an ornamental move. It’s Anthropic accelerating full throttle just as the AGI race is tightening.
(By the way, this fits in like a glove with what we were saying two weeks ago about the the race for the AGI reduced to three players in the West.. Anthropic is not playing to lose, far from it).
From "evil" to "no one set off my evil detector".
This is where the story really gets interesting. In February of this very year, literally three months ago, Musk had called Anthropic “misanthropic and evil” in a public response on X. Misanthropic and evil. Those were his exact words.
And what did he say this week when announcing the deal? I quote him verbatim: “Everyone I met was highly competent and cared a great deal about doing the right thing. No one set off my evil detector. So long as they engage in critical self-examination, Claude will probably be good” (quote). Which could be translated as “no one set off my evil detector”.
Three months, same person, different story.
What happened in between, according to Axiosis that Musk met in person with senior members of the Anthropic team the week before the announcement. That… and that someone probably did some very specific numbers on a whiteboard.
Why Anthropic needs this: the watt economy
If there is one lesson from the last twelve months of AI, it is that the real currency of the AGI race is not the idea or the model. It’s the watt. The raw computational capacity to train, fine-tune, and serve inference to millions of simultaneous users.
And Anthropic, until a few months ago, was clearly behind OpenAI and Google in that dimension. Dario and Daniela Amodei’s company has been racking up infrastructure deals like there’s no tomorrow: five gigawatts with Amazon, five more gigawatts with Google and Broadcom by 2027, a partnership with Microsoft that includes 30 billion in Azure capacity, a 50 billion investment with Fluidstack in American infrastructure, and now this extra 300 megawatts via SpaceX.
Add it up in your head. It’s hovering around ten gigawatts committed. For a company that barely existed in the public conversation two years ago, these are truly significant numbers.
And why so hungry for compute just now? My reading is that it has a lot to do with Mythos and the Glasswing program that Anthropic has started to deploy in cybersecurity, the rebound in Claude downloads after the Pentagon episode we experienced in February, and the sustained pressure to keep Claude Code up to speed in a market where GitHub Copilot, Cursor, and OpenAI itself are pushing very hard.
When you double users every few months, either you get more compute, or the service goes down. It’s that simple, and that disturbing at the margins.
Why Musk really does it
And here comes the part that seems to me the juiciest part of the matter. Because Musk’s official version is that of the convert, impressed by the Anthropic team’s human competence. But analysts who have looked at the numbers tell another, rather more prosaic story.
First up, SpaceX is preparing an IPO for next month; having 220,000 GPUs parked without generating revenue is not the best picture for the day of the roadshow. Converting those GPUs into contractual revenue, with analyst estimates of $3-4 billion per year for SpaceX, is.
Second, and this is pure Silicon Valley, this move allows Musk to do two things at the same time to Sam Altman: remove OpenAI’s relative weight in the computing economy and strengthen the competitor that makes him most uncomfortable. A classic board move. If you can’t win the game on your own, strengthen your rival’s rival.
The third thing, and this is what makes me think the most, is that Colossus 1 was built specifically for xAI to stand up to OpenAI. That three months later, Musk leases it in its entirety to Anthropic says a lot about how xAI is doing in the model race. Grok is already on its fourth generation (Grok 4.3 just landed in the API this week), but the market noise still isn’t keeping up with the pace Musk promised. If your own AI doesn’t need 220,000 GPUs turned on full blast, better to charge for them than watch them turned off.
Hypocritical? Maybe. Clever from SpaceX’s point of view? Undoubtedly.
The Martian detail: Anthropic looking into space
There is one paragraph in the release that is worth highlighting even if it is not the main course. Anthropic and SpaceX have confirmed that they are exploring together the development of “multiple gigawatts” of on-orbit computing. Yes, in space. According to SpaceNewsSpaceX has already filed with the FCC a plan for a constellation of up to one million satellites loaded with compute payloads, capable of adding about one hundred gigawatts of orbital capacity per year once Starship achieves reliable reusability.
Those of us who have been doing this for a while are not entirely surprised. If you remember, I discussed this topic in detail in my blog “Data centers in orbit: the ultimate cloud for the AI era? in which companies like Starcloud and Nvidia were in the pilot phase.
What then seemed like a fancy experiment by Nvidia to get attention is now starting to have serious customers lined up. And SpaceX is probably the only company in the world that has both the rockets to launch, the financial muscle to build and a CEO with enough narrative to sell this to investors. Whether it’s economically viable in the near term… that’s another question.
What I believe
The first thing is that inference is becoming the decisive lever of this entire industry. Whoever controls gigawatts will be able to decide which models are served to millions of users and at what speed. And that’s real power, not just market power. When energy is the bottleneck, the big deals are no longer made between AI companies, but with power companies, nuclear operators, and now space companies.
The second thing is that, in Silicon Valley, alliances are made out of interest, not affinity. Musk’s “no one set off my evil detector” is a good headline, but the real decision is that his shareholders, on the verge of a SpaceX IPO, would rather see compute generating cash flow than GPUs shut down. And the shareholders are right. Sometimes it pays to read public moves for what they are: corporate communication with economic sense, not personal convictions.
And the third thing, which to me is the most significant, is that Anthropic is repositioning itself as much more than a modeling lab. Mythos for defense, Glasswing for cybersecurity, Claude for enterprise, Claude Code for developers, and now an infrastructure footprint comparable to hyperscalers. It’s a pretty ambitious strategic move. The open question is whether a company of its size can sustain all those fronts at once without losing focus, don’t you think?
It’s an interesting company to follow, honestly. He’s doing everything at once and, so far, it’s going pretty well. You don’t see that every day.
Three months ago Elon Musk called Anthropic “misanthropic and evil”. Today he writes them a check for four billion dollars to lease the supercomputer he had built to take them on. Incredible, don’t you think?
Leave me your comments, I’d love to read them.
Have a good week!
INTERESTING RESOURCES
- CNBC – “Anthropic, SpaceX announce compute deal that includes space development” (6 May 2026)
- Anthropic – “Higher usage limits for Claude and a compute deal with SpaceX” (6 may 2026)
- Bloomberg – “Anthropic Inks Computing Deal With SpaceX to Meet AI Demand” (6 May 2026)
- Tom’s Hardware – “Musk’s SpaceX has rented out access to its supercomputer’s 220,000 Nvidia GPUs and 300 megawatts to rival Anthropic” (7 May 2026)
- Fortune – “Elon Musk called Anthropic ‘evil’ 3 months ago. Now he’s taking $4 billion to become its data landlord” (7 May 2026)
- Axios – “How Elon grew to love Anthropic” (7 May 2026)
- Al Jazeera – “SpaceX backs Anthropic with data center deal amidst Musk’s OpenAI lawsuit” (6 May 2026)
- TechCrunch – “Elon Musk’s lawsuit is putting OpenAI’s safety record under the microscope” (7 May 2026)
- Marketplace – “Anthropic’s new powerful allies: Elon Musk and SpaceX” (8 May 2026)
- Yahoo Finance – “Anthropic to rent all AI capacity at SpaceX’s Colossus data center” (6 May 2026)
