Project Suncatcher's first TPU test flies on SpaceX

Google’s Project Suncatcher is scheduled to put its first Tensor Processing Units in orbit, launching a prototype satellite on SpaceX’s upcoming Transporter-18 rideshare mission. The satellite was developed in partnership with Planet and will gather in-orbit data on how TPU hardware handles the physical stress of spaceflight. Announced last year as a long-term research moonshot, the question is whether low Earth orbit can host scalable machine learning infrastructure.

What’s actually new

• A dated flight, on a named launcher. This is a scheduled first orbital test rather than a concept. The ride is a rideshare mission, so the hardware is a passenger rather than the primary payload.
• The power argument. Satellites in low Earth orbit get near-constant sunlight, generating up to eight times more solar power than on the ground. Longer term, Google wants to link constellations of satellites with lasers to handle larger AI workloads in orbit.
• Radiation testing already done on the ground. TPUs were run under AI workloads in a proton beam facility at UC Davis’s Crocker Nuclear Laboratory while the team watched for bitflips. Google says its Trillium TPUs survived a total ionizing dose greater than a five-year space mission.
• Vibration and cooling are the other two hurdles. The launch climb lasts about ten minutes with sustained acceleration up to 10g, and individual TPU chips seeing 50 to 100g; the team vibration tested the satellite on all three axes. Cooling is the harder problem, since a vacuum has no airflow and only radiators work, so heat pipes plus radiators are being validated in a thermal vacuum chamber.

Why it matters

Worth being precise about what has and has not happened: no data center is in space, and none is scheduled to be. This is one prototype satellite carrying a handful of chips, and Google’s own framing is that the launch is about finding out what breaks. The milestone that would actually change the conversation is 2027, when the team plans two satellites in orbit to test high-bandwidth laser interconnect, since dozens of TPUs per satellite only matter if the cluster can talk to itself fast enough. Treat orbital data centers as a research program with a real flight date, not a roadmap item.

(Original source: Google)