Google Just Put a TPU in Orbit
October 2, 2026 · 5 min read
On October 1, the first prototype satellite of Alphabet's "Project Suncatcher" rode a SpaceX Falcon 9 into orbit on the Transporter-18 rideshare mission, lifting off from Vandenberg. Onboard: a Google-designed TPU, powered by solar panels. The mission is a reality check on one of the wildest ideas in computing — running AI chips in space.
Google wants to answer a brutally practical question: can an AI chip survive up there and do useful work?
Space is a terrible place for a chip
Everything a datacenter takes for granted, orbit takes away. Radiation flips bits — a cosmic ray can quietly corrupt a calculation, which is a minor annoyance in a cat photo and a catastrophe in a model training run. Temperatures swing wildly between sunlight and shadow. And worst of all: there's no air. In a datacenter, fans and liquid cooling carry heat away; in vacuum, you can only shed heat through heat pipes and radiators, glowing it off into the void. A TPU running full tilt in space is a space heater with nowhere to dump the heat.
That's exactly what Suncatcher is testing. Not "can we get a chip to space" — SpaceX solved delivery — but "can it run stably, correctly, and for a long time once it's there."
Why bother? The pitch is actually simple
Down on Earth, AI datacenters are fighting three wars: electricity, land, and cooling water. Some training clusters now consume as much power as small cities. In orbit, the pitch goes: unlimited space, and sunlight that never sets — free, endless power with no grid to build and no neighbors to complain.
If the physics works out, the economics start to look interesting. Launch costs have collapsed thanks to rideshares like Transporter-18 — you're not buying a rocket, you're buying a seat. Solar power in orbit is free forever. The remaining question is whether chips can be made space-tough cheaply enough that the whole equation beats building another datacenter in the desert.
Don't hold your breath — but don't laugh either
This is one prototype satellite, not a datacenter constellation. The realistic near-term wins are narrower: processing satellite imagery in orbit instead of beaming raw data down, running models where the data is collected. Nobody is training GPT-7 in space next year.
But the history of computing is full of ideas that looked absurd until the cost curves crossed. Ten years ago, "AI chips designed by a search company" sounded odd too — now TPUs power a decent chunk of the world's AI. Suncatcher is Google buying a cheap option on the possibility that the next datacenter isn't a building. It's an orbit.