Google’s Suncatcher satellite writes its first poem in orbit

Google says its Project Suncatcher prototype produced its first AI response in orbit on October 30. A ground station sent a prompt to an onboard Google DeepMind model, which returned a haiku. The satellite launched on October 1 as a test of AI hardware in space. The result marks an early research milestone, not a working orbital data center.

Google’s prototype rode SpaceX’s Transporter-18 mission into low Earth orbit aboard a Falcon 9 rocket. The launch took place from Vandenberg Space Force Base in California. Space.com reported that the rocket carried 130 payloads, including Google’s satellite.

Google says it built the satellite in partnership with Planet. Google says the mission began with weeks of commissioning work. Engineers powered up the satellite’s host systems and TPU trays before trying early machine-learning tests.

The first prompt asked Gemma 4, Google DeepMind’s open model, for a haiku about space. It returned: “Blue marble below / Data streams in silent grace / Starlight feeds my core.” — Gemma 4, Google DeepMind’s open model. The Google update says the response came down as the spacecraft passed over Svalbard, Norway. It demonstrated the basic chain: a ground prompt, onboard inference and a response sent back.

Google also tested space-themed limericks and prompts asking the satellite to report on its own technical status. Those tests suggest the satellite can run a model and send results back to Earth. They do not show that it can serve ordinary users or process large commercial workloads.

The payload includes four Google Tensor Processing Units, or TPUs. A correction to TechCrunch’s report clarified that the satellite can operate only one TPU at a time. The chips are Google’s data-center hardware, but this prototype tests a limited workload.

The prompt was an inference task. Inference means using a trained model to generate an answer, rather than training a model from scratch. Gemma 4 was already onboard, and the satellite generated a short poem after receiving a prompt from Earth.

This distinction matters because the mission’s goal is still basic. Google wants to learn whether its hardware can operate reliably in orbit. The first response does not demonstrate a distributed AI system or a space-based cloud service.

Space creates engineering problems that ground-based data centers do not face in the same way. Radiation can disrupt electronic memory through bit flips, while launch vibration and temperature swings can stress components. Google says it will use the mission to measure those effects directly. Ground tests cannot reproduce every aspect of orbital radiation or thermal exposure.

Cooling is another unresolved challenge. A vacuum has no circulating air to carry heat away from chips, so engineers must use radiators and heat pipes. Google’s Project Suncatcher facts page describes those systems as part of its research into operating TPUs in space.

TechCrunch reports that power limits keep the onboard TPU from running continuously. The current experiment therefore falls well short of the scale needed for a conventional data center. Google has not presented the satellite as a commercial service.

Project Suncatcher explores whether groups of satellites could eventually handle larger machine-learning workloads in orbit. Google says low Earth orbit offers near-constant sunlight and could provide more solar power than panels on Earth. That is a future possibility, not a capability demonstrated by this single satellite.

The company’s plans also depend on linking satellites together. Google describes future systems that would use laser communications to connect nearby spacecraft at high bandwidth. Its next stated milestone is a pair of interconnected satellites in 2027.

For now, the prototype can test one chip at a time and has returned a poem and telemetry prompts. A future constellation would need reliable power, cooling, radiation protection and fast links between satellites. The current mission can supply data on those hurdles, but it cannot prove that orbital data centers will be practical or economical.

Ultimately, Suncatcher has shown that a data-center-class TPU can generate a small AI response in orbit. Google will continue testing the hardware against radiation and thermal stress. The next milestone is two linked satellites in 2027. Until then, the project remains a research experiment, not a new cloud service.