Google’s first space data center satellite is online and talking to the ground, after launching earlier this month alongside more than 100 other spacecraft on a SpaceX Falcon 9 rocket. The prototype, part of Google’s Project Suncatcher, carries four of the company’s Tensor Processing Units and generates up to 1 kilowatt of electricity, though power limits mean only one chip can run at a time. Google says the spacecraft is operating nominally, and two more satellites are planned for 2027.
A haiku from orbit
Engineers spent three weeks commissioning the satellite, testing its subsystems before it was ready to work. For the first prompt, the ground team asked Gemma 4, the Google DeepMind model aboard, to write a haiku about being in space. The model answered with a poem about the blue marble below, data streams in silent grace, and starlight feeding its core.
The poem was a demonstration, not the mission. Google also asked the spacecraft to report on the orbital conditions around it, a practical test of operating a computer far above Earth. A company spokesperson confirmed the satellite is operating nominally, and the test campaign now turns to heat, radiation, and sustained operation.
One chip at a time
The power budget shapes everything the prototype can do. With four TPUs aboard but only about 1 kilowatt to share, the satellite can switch on a single chip for roughly 20 minutes at a time, which complicates sending results back to Earth. That constraint is part of the test, since the point is learning how hardware behaves in orbit.
Google’s frontier model work, by contrast, remains tightly held: Gemini 4 Carbon is limited to selected cyber defenders for now, while the satellite flies the open Gemma 4 model. The contrast shows two tracks inside Google, one pushing the most capable models under close control, and another testing whether ordinary accelerator hardware can survive and work in space at all.
Radiation is the central worry for the ongoing tests. In orbit, charged particles can flip bits in memory, corrupting values without triggering the usual monitoring tools. Google has tested its TPUs in particle accelerators on Earth, where results suggested radiation would matter little for inference and more for training, but the team now measures those effects in reality, not simulation.
A constellation that flies in formation
Suncatcher’s longer ambition is not a single satellite but a compact constellation of solar-powered spacecraft, flying close together and connected by high-speed optical links. The idea is to share work across satellites as if they were racks in a data center, with sunlight providing near-continuous power and no need for heavy batteries.
Other companies are chasing ideas. Starcloud and Cowboy Space expect to run inference on individual satellites, while SpaceX’s orbital data centers are expected to use laser links but orbit too far apart to share processing. Google’s bet is that tight formation flying makes a space data center work, and SpaceX’s AI infrastructure ambitions show how crowded the race is becoming.
Launch economics gate the vision. In a peer-reviewed analysis published this month, Suncatcher researchers estimated that SpaceX would need to fly its Starship rocket 1,800 times to achieve the same cost reductions it saw with earlier vehicles. Both companies will need that rocket to be reusable before orbital data centers can scale, which puts the timetable in years, not months.
Why It Matters
Suncatcher matters because it moves AI compute from paper to hardware. Google’s research argues space may be the place to scale AI, since a solar panel in orbit can be up to eight times more productive than on Earth. If launch prices fall below 200 dollars per kilogram by the mid-2030s, orbital compute could rival data centers on cost.
The caveat is the timescale. This prototype runs one chip for minutes at a time, not a product, and the 2027 satellites are a test phase with higher power and optical links to prove. Treat the haiku as a milestone marker, not evidence that space data centers are near, and watch the radiation and thermal results before judging the concept.
Background
Project Suncatcher is Google’s research moonshot for solar-powered, satellite-based AI infrastructure, and the prototype now in orbit is its first step from analysis to flight hardware.
Source: TechCrunch report and Google research post
0 Comments