The cloud is running out of Earth ↗

Google is putting AI hardware in space next week. On October 1, a fridge-sized satellite carrying four Google TPUs rides a SpaceX Falcon 9 into low Earth orbit. Project Suncatcher is the first real test of an idea that sounds absurd until you look at a utility bill: run the data centers in space, where the sunlight barely stops and nobody complains about the water.

The honest version is smaller than the vision. The satellite’s solar arrays make about a kilowatt — a microwave’s worth. The TPUs can only run around 15 minutes before heat builds up and forces a shutdown, because in a vacuum there’s no air to cool anything. Google already proved the chips survive radiation doses beyond a five-year mission at UC Davis. The part no ground test can cover is everything failing at once.

This is what the AI buildout looks like when the easy constraints are gone. It’s not chips anymore. It’s not money — Brookings puts total AI infrastructure spend at $10.3 trillion through 2032. It’s the grid connection that takes years, the cooling water, the town that doesn’t want a gigawatt next door. SpaceX put orbital compute in its IPO pitch. A startup already flew an H100 up there. Everyone’s looking up for the same reason.

Here’s the kicker: space solves the power problem by creating a worse one. Unlimited free sunlight, infinite cold — and nowhere for heat to go. The future of AI might not be decided by who trains the biggest model, but by who figures out how to air-condition a computer with no air.