Google received a first response produced by its artificial intelligence hardware in orbit. The Project Suncatcher prototype ran the Gemma 4 model and sent a short poem to Earth. This result opens the tests in real conditions of a project that, ultimately, aims at spatial computing infrastructures.
A first exchange with a data center chip in space
The satellite took off on October 1. Crews have since gotten it up and running, turning on the onboard equipment and beginning to check the calculation equipment. The experience reported by Google takes place during a passage within the range of a ground station: an instruction is sent to the Gemma 4 model, responsible for composing a haiku about space. The answer comes when the satellite passes near Svalbard, Norway. Other essays follow, with poems and requests related to telemetry, that is, the technical information that describes the status of the payload. Google places this first exchange in the week of October 26, without specifying the day of its execution.
We presented the Suncatcher prototype in September before its release. This time, the processors performed an AI task after reaching orbit. TPUs, designed by Google to speed up model calculations, are typically used in data centers. The test should help engineers determine how this hardware behaves outside of its usual environment.
Space partner Planet had already announced that the satellites launched on October 1 on the Transporter-18 mission, including Google’s prototype, were in good health and were making progress in their calibration. Producing a response from Gemma now adds a result directly related to the built-in calculation function.
The motivation of the project is energy. Google is studying orbits that allow almost continuous exposure to the Sun, with fewer periods without electricity production and fewer batteries to transport. In the scenario presented by its researchers, a panel could produce up to eight times more energy than on land. This estimate depends on the choice of orbit; it does not measure the performance of the currently tested prototype.
This research accompanies the difficulties of expanding terrestrial infrastructures, particularly their electrical connection and water consumption. Other players are exploring the same direction, including SpaceX and its orbital projects. Google, however, is still in the experimental phase and has not announced any commercial spatial calculation service on the occasion of this first result.
The next hurdle is getting the satellites to work together.
A data center is more than just a processor capable of responding to a request. Large models mobilize numerous accelerators that continuously exchange information. For this reason, Suncatcher projects nearby satellites, connected by laser beams, to share the work between several machines.
Research published by Google in November 2025 described links requiring speeds of several tens of terabits per second. Engineers achieved 800 gigabits per second transmission in each direction on the test bench with a pair of transceivers. In space, maintaining the alignment of these connections between moving vehicles is another project. Therefore, the architecture studied favors compact formations more than very distant satellites.
It is also necessary to evacuate the heat from the fleas. In a vacuum, fans cannot move air around components. Google is experimenting with heat pipes and radiators that transport and then dissipate this heat. The next tests of the prototype will focus on the thermal and radiation limitations actually encountered in orbit, after the tests already carried out on the ground.
The cost of launches remains an important economic condition. In its initial study, Google anticipated a move closer to certain costs of Earth’s energy if the price of orbiting it fell below $200 per kilogram in the mid-2030s. This is a research hypothesis, not a transportation rate obtained for this mission.
The next step announced provides two satellites interconnected in 2027. They should make it possible to test optical exchanges and the distribution of calculation tasks between several machines, a capacity that the first exchange with Gemma has not yet tested.
