Google Project Suncatcher puts four TPUs into orbit for AI compute tests
Transporter-18 prototype with Planet confirms contact; four Trillium TPUs run ~15-minute workloads while engineers watch radiation, heat, and vacuum cooling.
GoogleProject SuncatcherTPUTrilliumPlanetTransporter-18SpaceXOnboard AIOrbital computeLaser commsRadiationThermal managementJouleSpace tech
Google’s Project Suncatcher prototype is in orbit. The satellite, built with Planet, rode SpaceX’s Transporter-18 rideshare on 1 October 2026. Google confirmed contact and said the spacecraft is operating as expected.
The flight is the first hardware check for a research moonshot: whether low Earth orbit can host scalable machine learning infrastructure. The refrigerator-sized “MVP” satellite carries four Trillium-generation Tensor Processing Units—the same TPU family used in Google Cloud—and solar arrays rated around one kilowatt. Compute runs in short bursts of roughly 15 minutes at a time so engineers can watch radiation effects, thermal behaviour, vibration survivability, and vacuum cooling without airflow.
Ground campaigns already pushed the chips hard. Google vibrated the satellite on three axes to mimic launch loads, and ran AI workloads on Trillium TPUs in a proton beam at UC Davis’s Crocker Nuclear Laboratory. Early results pointed to radiation hardness beyond a five-year shielded mission dose. In space, cooling relies on heat pipes and radiators rather than fans—something the team also exercised in thermal-vacuum chambers before launch. A peer-reviewed paper in Joule and earlier Google Research system-design notes spell out the longer bet: dawn–dusk sun-synchronous constellations with free-space optical links.
This prototype does not prove orbital data centres are economical. It is a learning flight. The next published milestone is a two-satellite mission around 2027 meant to exercise high-bandwidth laser links between closely spaced spacecraft—pointing precision Google compares to hitting a coin-sized target from miles away while both ends move. Clusters of dozens of TPUs per satellite remain a design goal, not a claim for this launch.
For Voyage’s tech desk, Suncatcher matters because it puts commercial AI accelerators, orbital thermal design, and inter-satellite optical networking on the same test article. Watch the telemetry: radiation bit errors, thermal throttling, and whether 15-minute Gemini-class workloads stay stable over a planned one-year life.
Desk note from public releases. Sources: