← Back to feed

Google is launching its first space hardware this week: a prototype satellite carrying Trillium TPU chips aboard SpaceX’s Transporter-18 rideshare mission, developed in partnership with Planet. The test marks the first flight validation of Project Suncatcher, a long-term moonshot announced last year to explore whether low Earth orbit could serve as a platform for scalable machine learning infrastructure.

The logic behind Suncatcher starts with power. In LEO, satellites receive near-constant sunlight — up to eight times more solar energy than ground-based data centres. Google’s hypothesis: link enough satellite clusters together and you have a platform that can manage large AI workloads without the thermal and grid constraints of terrestrial infrastructure.

What the Hardware Had to Survive

Before flight, engineers ran the TPU hardware through two gauntlet tests.

Vibration: The rocket trip to orbit lasts roughly 10 minutes and subjects components to sustained loads of up to 10g, with individual chips experiencing peak forces of 50 to 100g. Google shook the assembled satellite on all three axes to mimic rocket-launch frequencies. The hardware held.

Radiation: Solar events and cosmic rays can corrupt memory through bitflips, destroying ongoing workloads. The team tested Trillium TPUs at UC Davis’s Crocker Nuclear Laboratory by running AI workloads under a proton beam. The chips survived a radiation total ionising dose greater than what they would accumulate over a five-year space mission. Initial results are described as “remarkably well” — rare language in engineering test reports.

The Cooling Problem

AI chips generate heat in dense concentrations. On Earth, that means fans and liquid cooling. In a vacuum, there is no convective airflow. The only heat escape path is radiation — requiring a completely different thermal architecture.

Google’s current approach combines heat pipes and radiators, tested in a thermal vacuum chamber before flight. The actual performance in orbit will validate or revise the design before the next hardware generation.

2027: The Harder Problem

Putting one satellite in orbit is the easy part. Making multiple satellites work together as a compute cluster requires each to know its precise position relative to its neighbours, and to maintain high-bandwidth links between them.

Google is using laser communication for the inter-satellite links — a technology that already exists in space, but typically optimised for low bandwidth over large distances. The Suncatcher design inverts that: very high bandwidth over very short distances, with the precision required to hit a coin-sized target from miles away while both ends are in motion.

A two-satellite laser-interconnect test is scheduled for 2027. That test will determine whether the fundamental networking architecture for orbital compute clusters is achievable at the required specs.

Where This Fits

Project Suncatcher is not a near-term infrastructure play. Google’s framing — “like early research into autonomous driving and quantum computing” — signals a decadal horizon. The immediate deliverable from this week’s launch is flight data on TPU behaviour in radiation and thermal conditions that cannot be fully replicated on the ground.

Travis Beals, Google’s Senior Director of Paradigms of Intelligence, is leading the project. The satellite is flying on Transporter-18 as a rideshare passenger alongside Planet’s Earth observation hardware — a common low-cost path for experimental payloads.

If the Trillium chips return good data from orbit, the 2027 twin-satellite test becomes the first step toward the real goal: constellations of TPU-carrying satellites that can be networked together to handle inference and training workloads from anywhere on — or above — Earth.