The AI compute crunch has gotten so absurd that Anthropic is now running on Elon Musk's GPUs. The same guy who called their AI "evil" three months ago. And both companies are now actively planning something wilder: data centers in orbit, because Earth's power grid literally can't keep up.

The Deal

On May 6, 2026, SpaceX signed a deal to lease its entire Colossus 1 supercomputer — all 220,000+ NVIDIA GPUs and 300+ megawatts of capacity — to Anthropic. The numbers are staggering:

Metric Value
GPUs 220,000+ (150K H100, 50K H200, 30K GB200)
Power 300+ MW
Location Memphis, Tennessee
Revenue for SpaceX $3-4B/year
Cash profit $2.5B+/year
Access timeline Full capacity within one month

Anthropic needed this badly. Their Q1 revenue grew 80x — far exceeding their own internal 10x projection. They were already sitting on compute commitments from Amazon ($25B investment), Google/Broadcom ($40B), Microsoft Azure ($30B), and NVIDIA ($15B). None of that capacity comes online for another six to twelve months. Colossus 1 is plug-and-play.

The Irony

Three months ago, Elon Musk called Anthropic "Misanthropic," accused them of stealing data "at massive scale," and claimed they "hate Western Civilization." On May 6, he walked it back:

"No one set off my evil detector. So long as they engage in critical self-examination, Claude will probably be good."

Translation: $3-4 billion in annual revenue changes everything. SpaceXAI had already migrated training to Colossus 2 — a 550,000+ GPU monster at 1-2 GW — so Colossus 1 was sitting partially idle. Leasing it to a competitor for $4B/year ahead of SpaceX's June 8 IPO roadshow is just smart business.

The community noticed. One HN commenter put it bluntly:

"SpaceX is the only unallocated NVIDIA GPUs on earth."

Another Reddit user: "Elon called Anthropic 'evil' 3 months ago. Now taking $4B annually. Pretty good pivot."

The Colossus Stack

Colossus 1 Colossus 2
GPUs 230,000+ 550,000+
Power ~500 MW ~1-2 GW
Architecture H100, H200, GB200 GB200, GB300
Investment ~$8B ~$18B
Status Leased to Anthropic Primary xAI training cluster

Colossus 2 is being built as the first gigawatt-scale AI facility. For comparison, a typical nuclear reactor produces ~1 GW. The cluster itself consumes more power than a small city.

The Orbital Play

Here's where it gets genuinely sci-fi. The joint press release casually mentions "multiple gigawatts of orbital AI compute capacity." SpaceX filed an FCC application in January 2026 for a million-satellite orbital data center constellation at 500-2,000 km LEO, connected via optical laser links at 6+ Tbps through Starlink.

The thesis is straightforward but audacious: terrestrial power, land, and cooling can't scale fast enough. A data center in space gets near-constant solar energy, no land acquisition battles, and theoretically unlimited geographic reach.

The technical challenges are severe. Andrew McCalip's analysis pegs a ~400% cost mismatch vs terrestrial deployment. Without atmosphere, there's no convective cooling — heat rejection is radiative only, which is far less efficient. Google's internal projections suggest cost parity by 2035, but that assumes launch costs continue dropping, which they have been, thanks to Starship reusability.

What This Means for You

If you use Claude, you already benefited. Rate limits for the five-hour window doubled for all paid plans. Peak hour restrictions were removed for Pro and Max subscribers. API rate limits for Opus models increased "considerably."

But the bigger story is what this deal signals. SpaceX anchoring SpaceXAI's IPO at $1.75-2 trillion by showing the Colossus cluster generates real revenue — from a competitor. Anthropic solving their immediate compute problem. And both companies betting that the next generation of AI infrastructure won't be built on Earth at all.

So What

The deal is smart for both sides. SpaceX monetizes idle hardware ahead of an IPO. Anthropic gets immediate compute when their other $100B+ in commitments won't deliver for another year. But the orbital data center plan is what sticks with me.

The fact that two of the most resourceful companies in AI are seriously pursuing space-based compute tells you something about the scale of this bottleneck. We've gone from "can we build better chips" to "can we build enough data centers" to "the entire planet's power grid isn't enough." That's not a prediction about the future — that's a press release from this week.

The environmental angle is the tension I can't shake. SpaceX filing for a million satellites in LEO for compute isn't a side note. It's the largest single proposed expansion of orbital infrastructure in human history. The regulatory, environmental, and space debris implications are enormous and largely unaddressed.

For now, your Claude limits doubled. But the deal that made it possible says more about where AI infrastructure is heading than any benchmark ever could.

Coverage: Forbes, CNBC, Tom's Hardware, Financial Times, Silicon Republic, Gizmodo, xAI official announcement, HN discussion, Reddit, Introl analysis, Data Center Dynamics, Andrew McCalip

Sources