The AI compute wars just went vertical—and infinite solar might finally break the energy ceiling.
The Summary
- Google is in talks with SpaceX for a rocket launch deal to support Project Suncatcher, its orbital AI data center initiative, while startup Cowboy Space closed a $275M round to build its own solar-powered orbital infrastructure
- The play: bypass Earth's energy grid entirely by tapping continuous solar power above the atmosphere, no day/night cycle, no weather, no transmission loss
- Cowboy's bet opens new markets in space technology while addressing AI's insatiable power demands—training runs that would brown out Texas could run indefinitely in orbit
The Signal
We're watching two parallel tracks converge. Google's Project Suncatcher discussions with SpaceX signal that hyperscalers see orbital compute as more than science fiction. Meanwhile, Cowboy Space's $275M raise proves venture capital agrees. The timing isn't coincidence. AI training runs now consume gigawatt-hours. GPT-4 training reportedly cost over $100M in compute, much of that energy. The next generation models will cost more.
The physics are simple. A solar panel on Earth catches maybe 6 hours of useful sun per day, filtered through atmosphere. The same panel in low Earth orbit gets 24-hour exposure at 35% higher intensity. No clouds. No night. No NIMBYs blocking your power plant permits. The economics only work if launch costs keep falling, which SpaceX has made plausible.
"The AI compute wars just moved to an arena with infinite solar and no energy regulators."
Here's what makes this moment different from past space infrastructure hype:
- Launch costs dropped 97% since 2000, from $65,000/kg to under $1,500/kg on Starship
- AI compute demand is genuinely bottlenecked by power availability, not just capital
- Starlink proved you can operate thousands of satellites profitably with current technology
Cowboy Space is betting orbital data centers revolutionize AI infrastructure, but the real tell is that Google is exploring this seriously enough to negotiate launch contracts. Google doesn't talk to SpaceX about speculative science experiments. They talk when they're modeling capex and power purchase agreements.
The catch: latency. Light takes 4 milliseconds to reach low Earth orbit and back. That's fine for batch training jobs but brutal for inference serving user requests. So the initial use case is obvious: training foundation models in orbit, serving them on Earth. Train where power is free and infinite. Serve where users are.
The Implication
Watch for which AI labs announce orbital training runs first. That's your signal this is real. If Google moves forward with SpaceX, expect Microsoft to counter with Blue Origin, Anthropic to announce partnerships, and a land grab for orbital slots that makes the domain name rush look quaint.
For anyone building in the agent economy: your models might soon train in space. The companies that crack orbital compute logistics will define the next decade of AI capability curves. Energy just stopped being the bottleneck.