THE CRUNCH
Google's experimental Suncatcher satellite has generated its first AI response from orbit, answering a ground-station prompt with a haiku about being a TPU in space. Travis Beals, senior director of Google's Paradigms of Intelligence research team, called it the "Hello World" milestone for the project: with the satellite's host and TPU trays powered up and machine learning tests complete, the onboard Gemma 4 model produced an inference over the downlink during a pass over Svalbard, Norway.
The poem itself was modest, but the economics were not. The Register calculates that the roughly 20 output tokens would cost about $0.000012 through the Google Cloud API, excluding input costs, while putting a satellite of perhaps 500 kg into orbit at retail launch prices of around $7,000 per kg makes it a very expensive haiku indeed.
Google's bet is that launch costs will fall. Its own calculations suggest that if getting mass to low Earth orbit drops to $200/kg, the amortised launch cost per kilowatt per year could become comparable to what US terrestrial datacenters spend on electricity. Google predicts SpaceX could reach that price by around 2035, and argues it is worth prototyping well in advance.
Radiation is the other hurdle. Lab tests exposed Google's Trillium Cloud TPU and AMD host server to a 67 MeV proton beam, and the system's high-bandwidth memory showed irregularities after a cumulative dose of 2 krad(Si), about three times Google's minimum survival requirement. For typical inference workloads, researchers estimate a failure probability on the order of 1 per 3 million inferences at expected orbital doses, though they note the impact on training jobs needs further study.
In numbers
Launch cost per kg to LEO
Today
~$7,000
/kg
By around 2035
$200
/kg
Google predicts SpaceX could reach that price by around 2035
Figures from The Register, as quoted in this brief.


