Google Project Suncatcher runs its first in-orbit TPU inference, with Gemma 4 writing a haiku

Background and pain points

AI compute growth is running into limits on power, cooling and land on the ground. Google's Project Suncatcher is a long-term research moonshot that asks whether scalable machine learning infrastructure could live in space, where low Earth orbit has abundant solar power but air cooling is impossible and hardware must survive radiation and launch vibration. The prototype satellite, built with Planet, reached orbit on SpaceX Transporter-18 on Oct 1, which we covered at launch.

Architecture highlights

On Oct 30 Sundar Pichai posted on X that the team sent a command up to orbit, booted the TPU, ran a prompt and received the first complete response back on Earth, a haiku. According to TechCrunch, after about three weeks of subsystem commissioning, Gemma 4 preloaded on the satellite handled this inference, and the team also asked the satellite to report on orbital conditions around it. TechCrunch says the satellite generates up to about 1 kW and that power limits mean the TPU can run only about 20 minutes at a time, which tightens space-to-ground communication windows.

Benchmarks

Google has published no throughput, latency or error-rate data; this is a functional check. The next tests focus on thermal stress and bit flips in the real radiation environment. Ground accelerator tests suggested radiation matters less for inference than for training, and that now has to be confirmed in orbit. Note that the TPU count differs by source: Google's launch post on X said four, while TechCrunch reports one.

Getting started

This is research hardware with no public endpoint, so developers cannot call it. Worth watching are the 2027 launch of two more satellites to test inter-satellite optical links, and Google's peer-reviewed paper in Joule. Production work should stay on Cloud TPU or the Gemini API. For small-model builders, Gemma 4 running in orbit under tight power limits is a signal that open small models suit edge and extreme-environment deployment.