Overview
- Google’s Project Suncatcher prototype, which launched Oct. 1, completed commissioning and on Friday returned the first verified AI output from a data‑center class TPU when the onboard Gemma 4 model downlinked a haiku.
- The team will now run a sequence of experiments to measure how thermal cycles and space radiation affect TPU reliability, testing for problems such as radiation‑induced ‘bit flips’ that can silently corrupt memory and data.
- The satellite carries a single TPU and generates roughly 1 kilowatt of electrical power, which limits active TPU runtime to short bursts of roughly 15–20 minutes and shapes how workloads can be scheduled in orbit.
- Google says the next hardware milestone is a pair of optically linked satellites planned for 2027 to demonstrate higher power, inter‑satellite links, and coordinated inference across nodes, but wider economics depend on much lower launch costs and high flight rates of heavy‑lift rockets.
- Project Suncatcher builds on prior ground tests, including proton‑beam radiation experiments, and represents a long‑term research effort to map technical tradeoffs for space‑based AI compute rather than an immediate commercial product.