Google Project Suncatcher Gets First AI Response From a TPU in Orbit: Gemma 4 Downlinks a Haiku, Twin Satellites Planned for 2027

Background

AI compute growth is increasingly limited by power, land and cooling on the ground. Google's Project Suncatcher is a long-term research moonshot exploring whether machine learning infrastructure could run in space on direct solar power. The prototype satellite, built with Planet, reached orbit on October 1 on SpaceX's Transporter-18 rideshare, and the research paper is published in the journal Joule. On October 30, Sundar Pichai announced a "hello world" milestone on X, and Google Research published a blog post with details the same day.

How it works

According to the official blog, the team spent a few weeks on its commissioning checklist: powering up the host and TPU trays and running early machine learning tests. During a ground-station pass it uplinked the first prompt, asking the onboard open Gemma 4 model for a haiku about being in space; as the satellite passed over Svalbard, Norway, the downlinked poem confirmed the inference worked. It then tested space-themed limericks and telemetry prompts reporting payload status. Per TechCrunch, the satellite carries four TPUs and generates up to about 1 kW, so it can run only one TPU at a time for about 20 minutes.

Benchmarks

Google has not published in-orbit inference speed, power or throughput figures. The Register cites ground-test data from Google's paper: high-bandwidth memory showed irregularities only after a cumulative 2 krad(Si), about 3x the 750 rad(Si) minimum requirement; at one inference per second, silent data corruption works out to about one in 3 million inferences, acceptable for inference, while the impact on training still needs study. These are vendor figures.

What developers can do

This is a single-satellite inference demo, not a usable cloud service, so developers cannot call it. Next Google will test how TPUs handle thermal stress and radiation in orbit, then launch two interconnected satellites in 2027 to test optical inter-satellite links and distributed computation. To follow along, read the official blog and the Joule paper; the same open Gemma 4 model can already be tried locally or on Google Cloud.