The Compute Crunch

Anthropic just signed a deal for 300+ MW of SpaceX compute. The AI compute race just shifted to orbital.

Anthropic has a compute problem. Not a software problem, not a talent problem — an infrastructure problem. The solution they just landed: SpaceX's Colossus 1, a 300+ megawatt data center that was supposed to power xAI's Grok models. Instead, it's going to run Claude.

The Deal That Reshapes Everything

The announced partnership gives Anthropic access to "all of the compute capacity" at Colossus 1, with the deal closing within the month. This isn't a small pilot — this is a full-tilt commitment to solving what's been the industry's worst bottleneck: GPU access.

Musk called Anthropic "evil" less than a year ago. Now his data center is running their models. The irony isn't lost on anyone who's been paying attention to how the AI sausage gets made — principles are negotiable when compute is scarce.

300 MW is roughly equivalent to powering 200,000 US homes. Anthropic just got all of it.

Why This Matters

The AI industry runs on compute, and compute runs on two things: chips and power. NVIDIA can't make H100s fast enough. Every major player — OpenAI, Google, Meta, xAI — has been scrambling for capacity. The hyperscalers are burning through billions building data centers. Energy constraints are hitting real limits in Virginia, in Oregon, everywhere there's been easy grid access.

SpaceX's Colossus went live in 2025 with a stated purpose: power Grok. Eighteen months later, it's powering the competitor that's suing Elon Musk. That's the market for you.

The broader signal: compute is the new oil, and whoever controls the infrastructure calls the shots. Anthropic just bought themselves a year of breathing room. What happens to the startups and smaller players who can't land these deals? They're at the back of the queue — again.

What's Next

Anthropic mentioned interest in "orbital AI compute capacity." They want to put compute in space. That's either brilliant or absurd, and probably both. Either way, the playbook is clear: secure infrastructure first, build models second. Everything else is just prompting.

This deal changes the calculus for everyone. If the compute-constrained can make deals with whoever has spare capacity, the entire supply chain gets rebuilt. The question isn't whether more deals like this happen — it's who ends up with the keys when the music stops.

Data via TEXXR