Google's Orbital AI Satellite Returns a Haiku From Low Earth Orbit

Google announced on Friday that its first experimental orbital AI datacenter satellite has produced a poem in space, according to The Register. The Suncatcher craft, which launched earlier in the month and now circles in low Earth orbit, answered a prompt relayed from a ground station asking for a haiku written as if by a TPU watching Earth turn below.

Travis Beals, a senior director at Google's Paradigms of Intelligence research group, described the milestone in a blog post. He wrote that the satellite's host and TPU trays had been switched on and machine learning checks finished before the onboard Gemma 4 open model returned its first inference from orbit. Beals framed the result as the project's Hello World moment, with the reply arriving over the downlink as the satellite passed above Svalbard, Norway.

The text itself was tiny. Beals put the output at roughly 20 tokens, which at Gemma 4 26B's published API rates would come to about $0.000012, before any input charges are counted.

The hardware carrying that reply is where the expense lies. Using the mass of SpaceX's Starlink satellites as a rough guide, a craft might weigh around 500 kg, and at a retail launch price near $7,000 per kg the poem becomes a costly piece of verse.

Google has run the numbers on bringing those launch costs down. Its calculations suggest that if reaching low Earth orbit could cost $200/kg, the amortized launch expense over a satellite's working life, put at roughly $810 to $7,500 per kW each year, might land near what US terrestrial datacenters currently pay for electricity, about $570 to $3,000 per kW annually. Google expects SpaceX could get there around 2035 and argues for prototyping well ahead of that point.

Beals said the team followed the haiku with space-themed limericks and prompts requesting telemetry on payload systems. Thermal stress and radiation testing of the TPUs comes next.

Laboratory work has already exposed a Trillium Cloud TPU and an AMD host server to a 67 MeV proton beam. Google researchers reported that high-bandwidth memory is the most dose-sensitive part, showing irregularities once cumulative exposure reaches 2 krad(Si), about three times the company's 750 rad(Si) survival threshold. In a paper published earlier this month they estimated a silent data corruption rate near one event per 17 rad, which against an orbital dose of about 150 rad(Si) a year works out to roughly one failure in 3 million inferences at one inference per second. They noted that this is probably tolerable for inference but that effects on training jobs need more study.

Google plans to test two linked satellites next year and envisions many small satellites talking to each other rather than a few large ones.