SpaceX And Anthropic Partnership | The Brainstorm EP 131

Anthropic has been in crisis mode. Surging demand for Claude exhausted available compute, forcing the company to impose usage restrictions and token limits—cutting off not only customer access but also internal research capacity. Enter SpaceX and XAI. The Colossus 1 data center, a 300-megawatt facility housing 220,000 GPUs, is now being leased to Anthropic. The deal lifts capacity restrictions immediately and signals Anthropic's interest in «multiple gigawatts of space-based compute when the time is right».

For SpaceX, the economics are compelling. Colossus 1 is optimized for inference, not the training workloads XAI prioritizes for Grok. Leasing it out monetizes an underutilized asset while demonstrating infrastructure-as-a-service capabilities ahead of a hotly anticipated IPO. The arrangement also marks a thaw in relations: Anthropic had previously cut XAI off from accessing Claude models through third-party tools like Cursor, citing competitive concerns. Now, compute scarcity has brought the two back to the negotiating table.

The deal underscores a broader trend. Vertical integration—from chip fabrication to launch services—gives SpaceX optionality that pure-play model companies lack. If Grok demand explodes, SpaceX can redeploy GPUs in months, not years. Meanwhile, Anthropic gets the compute it needs to serve customers today, but remains structurally dependent on external infrastructure providers.

Can you build a gigawatt in orbit for less than $20 billion? That's the terrestrial facility benchmark—concrete, steel, labor, power infrastructure. In space, those costs vanish. Instead, you're manufacturing satellites at scale and launching them on Starship. The math hinges on reusability. At zero reuse, Starship costs roughly $1,000 per kilogram to orbit—putting a gigawatt at $25 billion, above the Earth baseline. But reuse five times, and launch costs collapse to $300/kg, or $7.5 billion for a gigawatt.

The broader $60 billion data center cost includes $30 billion in GPUs and $10 billion in networking, storage, and other IT. Those costs persist in space, bundled into the satellite bus, solar panels, and radiators. The real question is whether SpaceX can manufacture satellites more efficiently than terrestrial contractors can pour concrete and run power lines. The hosts argue yes: modular, serial production lines beat bespoke, site-specific construction. Anecdotally, data center developers are now piecing together 100-megawatt sites instead of single gigawatt campuses, adding operational complexity.

But space doesn't need to be cheaper to win. If Earth-based compute is scarce, AI companies will pay 50–100% premiums just for capacity. Anthropic's desperation for Colossus 1 illustrates the point. Velocity to market and predictable timelines may matter more than marginal cost—especially as revenue per gigawatt climbs due to surging enterprise demand and improving model efficiency.