The companies training the models that will run your life are running out of Earth.
The Summary
- SpaceX is deploying orbital AI data centers powered by Nvidia's Vera Rubin chips, targeting 10 gigawatts of compute capacity by 2027
- The collaboration leverages space-based infrastructure to bypass terrestrial energy grid constraints and thermal limits
- This isn't a proof of concept anymore. It's industrial-scale compute infrastructure where cooling is free and power density regulations don't exist.
The Signal
SpaceX isn't just launching satellites. They're building orbital data centers with Nvidia's Vera Rubin architecture, aiming for 10 gigawatts of AI compute by 2027. That's roughly the power consumption of a small country, now floating 340 miles above regulatory reach. The move sidesteps the two biggest bottlenecks choking AI scaling on Earth: energy grid capacity and heat dissipation.
In space, cooling is trivial. Radiate into the void. No chillers, no water infrastructure, no neighbors complaining about the noise. Power comes from solar arrays that never see night for half the year in the right orbits. The space-based approach fundamentally transforms global data capabilities by removing the physical constraints that have defined data center location strategy since the industry began.
"The companies training frontier models are no longer optimizing for proximity to fiber or cheap electricity. They're optimizing for escape velocity."
The timeline is aggressive. 10 GW by 2027 means SpaceX is launching compute payload at a pace that would make traditional satellite operators dizzy. For context:
- Current largest terrestrial data centers run 300-500 MW
- 10 GW is 20-33x that scale, distributed across orbital infrastructure
- Each launch carries compute density that would require city-sized campuses on the ground
Nvidia gets a captive customer for chips that might not even need to be optimized for Earth's atmosphere. This collaboration could significantly boost Nvidia's market influence beyond their already dominant position. More importantly, it locks in architectural choices for the models being trained up there. Whatever runs well on Vera Rubin in orbit becomes the de facto standard.
The second-order effects ripple fast. If training happens in space, where does inference happen? If your agent's brain was built in orbit, does it phone home through Starlink every time it needs to think? Latency to low Earth orbit is 25-35 milliseconds. That's workable for some tasks, terrible for others.
The Implication
Watch what happens to terrestrial data center real estate in the next 18 months. If SpaceX proves the unit economics work in orbit, the market will reprice every power-constrained facility on the ground. Coastal metros that bet big on attracting hyperscale compute may have built for a paradigm that's already obsolete.
For anyone building AI-native products, the question becomes: where does your compute live, and who controls the launch manifest? If SpaceX becomes the gatekeeper to orbital training capacity, they're not a launch provider anymore. They're infrastructure for the intelligence layer of Web4. That's a different kind of power than selling rocket rides.