A haiku from orbit
Google said on 30 October that its Project Suncatcher prototype satellite has returned the first AI response generated by a data-centre-class TPU in orbit. The team uplinked a prompt during a ground-station pass asking the onboard Gemma 4 open model for a haiku about being in space. The answer came down as the satellite passed over Svalbard, Norway:
“Blue marble below / Data streams in silent grace / Starlight feeds my core.”
Travis Beals, a senior director at Google working on what the company calls Paradigms of Intelligence, wrote the post. He described the milestone as a “Hello World” moment rather than a demonstration of useful capacity, which is the right framing: one prompt, one short poem, one satellite.
What the satellite has actually done so far
The prototype launched on 1 October. Google says the weeks since went on a commissioning checklist — powering up the host and TPU trays, then running early machine learning tests. The haiku is the first result from that sequence that the company has published.
After the poem, the team tried space-themed limericks and telemetry prompts that asked the payload to report on its own system status. Both are reasonable first workloads: short, cheap, and easy to check against what the hardware reports through its ordinary channels.
What Google has not published is the part that decides whether any of this scales. The post does not give the TPU generation on board, the number of accelerators, the power available to them, the latency of the ground link, or how many tokens the payload produced per second. For a project whose entire premise is that compute can be cheaper in orbit than on the ground, none of the economics are in the announcement.
The tests that matter have not run
Google says the next step is to evaluate how the TPUs behave under thermal stress and radiation exposure in orbit. Those are the two failure modes that have historically kept commercial silicon out of space. Heat has nowhere to convect to in a vacuum, so a dense accelerator has to radiate everything it generates. Charged particles flip bits and, over time, degrade transistors; terrestrial parts are not hardened against them.
Until those results exist, the haiku establishes that the chips powered on and completed an inference pass, not that they can hold a workload. Google’s own sequencing concedes the point: the company says the next milestone, a pair of interconnected satellites in 2027, depends on first verifying that TPUs can run workloads in orbit reliably.
Why anyone is trying this
The case for orbital compute is power. A satellite in the right orbit sees the sun almost continuously, which removes both the intermittency that constrains terrestrial solar and the grid connection that has become the binding constraint on data centre construction. Utilities in several markets are now the scheduling authority for new AI capacity; an orbiting array answers to nobody’s interconnection queue.
The case against it is everything else. Launch mass is expensive, heat rejection is hard, radiation shortens component life, servicing is impossible, and the data has to come down through a ground station network that does not yet exist at the bandwidth a real cluster would need. The Svalbard pass that carried a haiku would not carry a training run.
What to watch
Two numbers will tell whether Suncatcher is a research curiosity or a direction. The first is the thermal and radiation results Google says are coming, which will establish whether consumer-process accelerators survive long enough in orbit to be worth launching. The second is the 2027 twin-satellite test, which is where inter-satellite links get measured — the thing that would turn individual spacecraft into something resembling a cluster. Google has attached a year to the second and no date at all to the first.
