A satellite in orbit wrote a haiku about starlight feeding its core. Cute. But the haiku is the least interesting thing about Google’s Suncatcher prototype, which is now the first working test of whether AI chips can survive and do useful work in space.
Google’s prototype is online and communicating with the ground after launching earlier this month on a SpaceX Falcon 9, alongside more than 100 other spacecraft, according to TechCrunch. The satellite carries a single Tensor Processing Unit and generates up to 1 kW of power. That is a tight budget. The TPU can only run for about twenty minutes at a time, which makes the already awkward job of space-to-ground communication harder. After about three weeks of commissioning, engineers sent their first prompt to the onboard model, Google DeepMind’s Gemma 4. The result was the haiku. More usefully, Google also asked the spacecraft to report on orbital conditions, and a spokesperson said it is operating nominally.
The poetry is a demo. The test campaign is the point. Google has blasted its TPUs in particle accelerators on Earth, but orbit is different. Radiation can cause bit flips, which change stored values in memory and can corrupt data without tripping normal monitoring tools. Ground tests suggested this would be a minor issue for inference and a bigger one for training. Now Google gets to check that against real hardware under real thermal stress, not a simulation.
If the chip holds up, two more satellites with advanced chips are planned for 2027. They will generate more power and demonstrate inter-satellite optical links. That matters because Google’s approach is unusual. Its project is largely alone in planning high-speed links between close-flying satellites to run coherent distributed workloads. SpaceX’s data centers are also expected to use laser links, but they will reportedly orbit too far apart to share processing well. Starcloud and Cowboy Space are betting on inference on individual satellites, which is simpler but less ambitious.
The economics still hinge on a rocket that doesn’t fully exist yet. Both Google and SpaceX need Starship to become fully and rapidly reusable. Suncatcher researchers, in a peer-reviewed analysis published this month, estimated SpaceX would need to fly Starship 1,800 times to get the learning-driven cost reductions it saw with earlier vehicles. That is a staggering number, and it shows how much of this idea depends on launch costs falling.
There’s also a corporate tangle. Alphabet holds a significant investment stake in SpaceX, so Google’s orbital ambitions and its launch provider are closely linked. Whether that is convenient or risky depends on how Starship progresses.
So does this matter? Yes, but modestly for now. One TPU running for twenty minutes at a time is not a data center. It’s a proof of survival. For developers and founders, nothing changes today. Terrestrial AI infrastructure is still where the work happens, and power and cooling constraints on the ground are the pressure driving this research. If you’re watching the space compute race, the radiation results and the 2027 optical-link demo are the milestones that count.
Suncatcher is a classic Google moonshot, aimed at infrastructure that makes sense in the 2030s. Getting there may depend on the near term, though. Alphabet’s stock fell sharply this week after it reported that efforts to put AI into search are hurting ad revenue growth. Moonshots need patient funding, and that patience gets tested when the core business wobbles.
