Google's Project Suncatcher satellite returns its first AI output from orbit

Google said on October 30 that its Project Suncatcher prototype satellite has returned its first AI output from orbit. In a post on The Keyword, Travis Beals, senior director of Google's Paradigms of Intelligence team, wrote that engineers uplinked a prompt during a ground-station pass earlier this week, asking the onboard Gemma 4 open model for a haiku about being in space. The answer came back as the satellite passed over Svalbard, Norway. Google calls this the first AI response from a data-center-class TPU in orbit; we have not seen an independent check of that "first" claim.

The poem itself is beside the point. What the milestone shows is that the chip, its host computer and the power and thermal systems around it survived launch and several weeks of commissioning well enough to run inference and send results to the ground.

What has been tested so far

The satellite was launched on October 1 on SpaceX's Transporter-18 rideshare mission and was built with the satellite operator Planet. According to Google, the weeks since were spent on a commissioning checklist that included powering up the host and TPU trays and running early machine learning tests. After the haiku, the team sent space-themed limericks and telemetry prompts that asked the model to report the technical status of the payload.

TechCrunch reported additional hardware details that Google's post does not state: the spacecraft carries four TPUs and generates up to 1 kW of power, but it can run only one TPU at a time, for about 20 minutes, because of the chips' power draw. A Google spokesperson told TechCrunch the satellite was operating nominally.

What comes next

Google says the next phase is a series of tests on how the TPUs behave under thermal stress and radiation exposure in orbit. Those are the two questions that ground testing could only approximate:

• Radiation. Google said in September that its Trillium TPUs, tested in a proton beam at UC Davis's Crocker Nuclear Laboratory, survived a total ionizing dose greater than a five-year mission would deliver. Bit flips that silently corrupt data are the concern; TechCrunch notes that ground results suggested inference is less exposed to them than training.
• Cooling. In a vacuum there is no airflow, so heat must be moved out through heat pipes and radiators. Google has tested its design in a thermal vacuum chamber and says it will refine it based on what it measures in orbit.
• Interconnects. The longer-term design calls for clusters of satellites, each carrying dozens of TPUs, linked by short-range, high-bandwidth lasers. Google plans to test this with two interconnected satellites in 2027.

Why it matters, and what it does not mean

Suncatcher is Google's bet that some AI compute could eventually move to orbit, where it says satellites in low Earth orbit can get up to eight times more solar power than panels on the ground. The motivation is the same energy and siting pressure that is shaping data center construction on Earth. Google describes the project as a long-term research moonshot and has not given a date for anything beyond the 2027 two-satellite test.

For developers and teams choosing models, nothing changes today. A single satellite running one TPU in 20-minute bursts is an engineering experiment, not capacity anyone can rent, and there is no product, pricing or API attached to it. The useful signal is narrower: Google is now collecting real in-orbit data on how its chips handle radiation and heat, rather than relying on simulations and accelerator tests. Whether space-based compute becomes economical depends on results Google has not yet published and on launch costs that are outside its control.

One detail worth noting for those who follow open models: the first model Google chose to run in orbit was Gemma 4, its open-weight family, rather than a Gemini model.