It is up
Google’s first piece of Project Suncatcher hardware reached orbit on 1 October aboard a SpaceX rocket. The refrigerator-sized prototype, built with the satellite imagery company Planet, carries four of Google’s tensor processing units. NPR reported that the chips will run a version of the company’s open-weight Gemma model to answer simple queries — for fifteen minutes at a time, because of heat.
That constraint is the story. On Earth a data centre dumps heat into air or water. In vacuum there is nothing to dump it into, so the chips run, get hot and stop. Travis Beals, who leads the project, told NPR the radiators are among the heaviest components on the spacecraft, which is a direct cost problem: heavier payloads cost more to launch.
What the mission is for
The satellite sits in a sun-synchronous orbit where its panels are almost never in shadow, which removes the need for heavy batteries. Planet will commission the spacecraft, after which Google will power up and test the TPUs. The mission is meant to run for a year, and it is measuring survival rather than throughput: whether accelerators designed for a climate-controlled building tolerate launch loads, radiation and thermal cycling.
Two more satellites are planned for next year to test the laser links between spacecraft, which any real cluster would need. Google’s sketched end state is clusters of satellites each carrying dozens of TPUs, talking to each other and to the ground by laser.
The economics nobody has solved
Beals was unusually direct with NPR about cost. He said he does not see space compute being cheaper than ground compute in the next five years, and that it will take longer than that. He also said he does not expect orbital data centres to replace terrestrial ones for a long time, and that some workloads will make more sense in each place.
Brandon Lucia, a professor of electrical and computer engineering at Carnegie Mellon, put the maintenance problem plainly to NPR: when something fails in a terrestrial data centre a technician walks in, and in orbit the cost and complexity of every such task is multiplied by ten or a hundred. TechCrunch, reporting on a Google white paper due in Joule, said the modelling implies roughly 1,800 Starship launches over the next decade to bring launch costs near $200 per kilogram by 2035, the point at which Google’s figures reach parity with the ground.
What to watch
The first published telemetry: how much power the panels actually deliver, how hot the chips run, and whether fifteen minutes turns out to be optimistic. Whether the 2027 laser-link mission holds its date. And whether anyone outside Google and SpaceX models the launch cadence independently, because a parity date that depends on 180 Starship flights a year is a claim about rockets, not about chips.
