Last week Anthropic quietly signed a lease for every chip in Elon Musk's Memphis supercomputer. All 222,000 of them. A 300-megawatt facility. The whole thing, locked up by one tenant for an undisclosed sum.

The buyer was the AI lab that built its brand on being the cautious one. The landlord was Musk, who has spent the past two years calling Anthropic a doomer cult. His public reaction to renting them the keys: "No one set off my evil detector."

What actually changed hands

Colossus 1 is the cluster Musk slapped together in 122 days to train Grok. It packs over 220,000 Nvidia H100, H200, and GB200 GPUs into a converted appliance plant outside Memphis. Anthropic now controls the entire site for its own training and inference workloads, alongside its existing AWS and Google TPU footprint.

This deal happened because Anthropic ran out of compute. CEO Dario Amodei has been blunt about it: Q1 2026 demand grew 80x against an internal plan of 10x. Claude Code rate limits were getting throttled. Opus API customers were complaining loudly. Something had to give.

The day the deal was announced, Anthropic doubled Claude Code's five-hour rate limit for every paid tier, removed peak-hour throttling for Pro and Max users, and raised Opus API limits by up to 1,500%. Subscribers got the upgrade overnight. The compute came from a guy whose own AI lab is suing Anthropic's largest investor.

Worth pausing on the optics. xAI and Anthropic are direct competitors in the frontier model race. Grok 4 and Claude Opus 4.6 sit within points of each other on most public benchmarks. Renting an entire training cluster to a rival is the kind of move investment banks invented golf for. It only happens when the landlord is convinced the rent is worth more than the lead.

The orbital part is real

Buried in the joint announcement is a line about Anthropic and SpaceX exploring "multiple gigawatts of orbital AI compute capacity." Gigawatts. In orbit. As in, data centers on satellites, beamed power, the whole science-fiction stack. Neither side put a date on it, but neither side denied it either.

This is what compute scarcity looks like at the frontier. Amodei testified to Congress two months ago that AI labs are pushing against the physical limits of US grid capacity. Maine just banned new data centers. The cheapest remaining electron is apparently the one you generate above the atmosphere.

My Opinion

I'll be blunt. Anthropic spent three years selling itself as the lab that wouldn't compromise. They wouldn't take Pentagon offensive-AI contracts. They wouldn't ship Mythos to the EU without safety audits. They publish responsible-scaling policies and ask competitors to follow along. Then 80x growth hit and they wrote a check to the guy who once tweeted that they should be regulated out of existence. Principles are easy when you have spare capacity.

And here's what bugs me. Musk's "no one set off my evil detector" line is the most honest thing anyone has said about AI safety this year. It's a joke, but it's also a confession: the safety framework everyone advertises collapses the moment a competitor pays in dollars. Nobody walked away. Nobody flagged anything. Anthropic's board approved sending hundreds of millions to a man whose company is currently in active litigation against OpenAI, who is in turn Anthropic's chief rival. The check cleared. Everyone moved on.

The orbital data center talk gets dismissed as a stunt. I think people are missing the actual story. Anthropic is signaling, in writing, that it expects to need more electricity than the United States can produce on the ground within the next decade. They are willing to bet on space-based power before they are willing to slow down. That is not a hedge. That is the new floor for what "frontier AI" requires, and every regulator, utility, and competitor should be reading this announcement as a flare.

Author: Yahor Kamarou (Mark) / www.humai.blog / 12 May 2026