Google Runs Gemma 4 on a TPU in Orbit in Suncatcher Test

A test bed, not a data center

The prototype carries four TPUs but can use only one at a time, for short periods of about 20 minutes. It is a test bed rather than an orbital data center. Google introduced TPUs in 2016 as an alternative to NVIDIA GPUs, although Broadcom helps develop them and TSMC manufactures them.

Running AI in orbit is not new. In 2020, the European Space Agency's Φsat-1 used AI to filter cloudy images before transmitting them. In November 2025, Starcloud launched a satellite with an NVIDIA H100 GPU and said the following month that it had run Gemma on board. Google's claim is narrower: getting a response from a data-center-class TPU in space.

The energy bet and the obstacles

Project Suncatcher envisions distributed satellites with processors, linked by optical connections to run AI workloads. The main draw is energy: in certain orbits, near-constant sunlight could yield, by Google's estimates, up to eight times the electricity of equivalent solar panels on Earth.

The obstacles are significant. In a vacuum there is no air to carry heat away from processors, so Google is studying heat pipes and radiators, which it has tested in vacuum chambers. Space radiation can disrupt circuits, and the company has run ground tests for it. It still has to show how the hardware performs over long periods and whether deployment costs are acceptable.

After the haiku, researchers tried other space-themed poems and requests about the satellite's technical status. They are now preparing tests on how the TPUs withstand thermal stress and radiation in orbit. The next announced goal is an experiment with two interconnected satellites in 2027.