Yesterday we talked with our future AI overlords in heaven, and heaven haiku’d back.
Google sent a command into orbit, booted up a TPU on a satellite, prompted an LLM, and got back a haiku as an answer.
The tiny TPU chip that could. Survived launch, survived a few weeks in vacuum and radiation, and still ran a model when asked. Pichai is calling it the project’s hello world moment, for a first data center in orbit proof-of-concept.
Google’s Project Suncatcher’s premise is simple: low Earth orbit gets unfiltered sunlight nearly all the time, up to eight times what the same panel gets on Earth, so run the compute where the power already is instead of building another gas plant next to a data center. The prototype launched October 1st carrying four TPUs, one operable at a time, built on a standard Planet Labs satellite bus just to see if the hardware survives the trip. It did. Google’s own chips took a harder radiation dose in testing than a five-year mission would deliver and kept working.
Scaling the feat is a different conversation. Google’s own peer-reviewed paper argues that getting launch costs down to around $200 a kilogram by 2035 requires SpaceX’s Starship flying roughly 1,800 times over the next decade, about 180 launches a year. Starship has never flown more than five times in a year. Their vision is a cluster of 81 satellites talking to each other via super-precise laser links in motion.
So the actual milestone this week is very small considering the end game. A haiku back from orbit means the TPU survived, the command chain worked, and the model ran. We are just missing 1.8k trips to heaven.
