Google runs an AI model on a satellite using one of four TPUs in orbit

Google's Project Suncatcher prototype is online and communicating with the ground, but its compute is tightly limited. The satellite carries four Google Tensor Processing Units and generates up to 1 kW of electricity. TechCrunch reports that the chips' power demands mean only one can be switched on at a time, and for about 20 minutes per run. That constraint adds to the difficulty of space-to-ground communications.

The spacecraft launched earlier this month on a SpaceX Falcon 9 alongside more than 100 other spacecraft. After about three weeks of commissioning, during which engineers tested the satellite's subsystems, the team sent its first prompt. The model preloaded on the satellite, Google DeepMind's Gemma 4, was asked to write a haiku about being in space. Google also had it write space-themed limericks and report on the orbital conditions around the satellite. A spokesperson told TechCrunch the satellite was operating nominally.

The ongoing test campaign will examine how the chips perform under thermal stress and in the space radiation environment. Radiation can cause bit flips, which change a stored value in computer memory and can corrupt data without triggering typical monitoring tools. Google had already tested its TPUs in particle accelerators on Earth. That testing suggested radiation would have minimal effect on inference workloads but more of a problem for training a model. TechCrunch's report does not say what the orbital tests have found so far.

The ground results Google has published come from a test of Trillium, its v6e Cloud TPU, in a 67 MeV proton beam. The high bandwidth memory was the most sensitive component, but it showed irregularities only after a cumulative dose of 2 krad(Si). That is nearly three times the expected five-year mission dose of 750 rad(Si) for a shielded chip. No hard failures were attributed to total ionizing dose up to the maximum tested dose of 15 krad(Si) on a single chip.

Google's design research envisions compact constellations of solar-powered satellites. The proposed system would likely use a cluster of satellites in a dawn-dusk sun-synchronous low Earth orbit, where they would be exposed to near-constant sunlight. Data center-scale links between satellites require very high received power. Because received power falls with the square of distance, the satellites would fly in a very close formation, kilometers or less apart. Google's bench-scale demonstrator reached 800 Gbps each way using a single transceiver pair. In an illustrative 81-satellite configuration at a mean altitude of 650 km, the cluster radius is 1 km and next-nearest neighbors are roughly 100 to 200 m apart.

TechCrunch describes Suncatcher as largely alone among space-based data center plans in calling for high-speed links between satellites to support distributed workloads. It says SpaceX's data centers are also expected to use laser links, but to orbit too far apart to share processing effectively. Starcloud and Cowboy Space, by contrast, expect to run inference on individual satellites. Alphabet, Google's parent, holds a significant investment stake in SpaceX, and both companies would need Starship, the heavy-lift rocket still in development, to be fully and rapidly reusable. A Suncatcher analysis estimated that SpaceX would need to fly Starship 1,800 times to achieve the learning-driven cost reductions it saw with earlier vehicles.

Google's design research argues that launch cost is a primary economic barrier to large-scale space-based systems. With a sustained learning rate, it suggests, launch prices may fall to less than $200/kg by the mid-2030s. At that price, a space-based data center could be roughly comparable to the reported energy costs of an equivalent terrestrial one, per kilowatt per year. The research also notes that significant engineering challenges remain, including thermal management, high-bandwidth ground communications and on-orbit reliability.

The next steps are a learning mission with Planet and a pair of follow-on satellites. Google's research says the mission, slated to launch two prototype satellites by early 2027, will test how its models and TPU hardware operate in space and validate optical inter-satellite links for distributed ML tasks. TechCrunch says the next satellites would launch in 2027 with the advanced chips, supporting higher power generation and demonstrating those links, if the TPUs onboard this spacecraft prove capable. TechCrunch also noted that Alphabet's stock fell the same week it reported that its efforts to insert AI into search were hurting ad revenue growth.