Google's Suncatcher Haiku Beamed Back From Orbit

Google said on October 30 that its Project Suncatcher prototype satellite has produced the first AI response from a data-centre-class TPU in orbit. Engineers uplinked a prompt during a ground-station pass, the onboard chip ran Google's open Gemma 4 model, and the answer came down as the satellite passed over Svalbard, Norway. The Suncatcher haiku read, in full: "Blue marble below / Data streams in silent grace / Starlight feeds my core."

The prompt asked the model to write a haiku about being in space. According to Google, the team followed it with space-themed limericks and telemetry prompts that asked the payload to report on its own technical status. Travis Beals, Google's senior director for Paradigms of Intelligence, announced the result in a post on the company's blog, describing it as the first-ever AI response from a data-centre-class TPU in orbit. Google chief executive Sundar Pichai described the test as a hello-world milestone for the project.

The satellite, built with Earth-imaging company Planet, launched on October 1 aboard SpaceX's Transporter-18 rideshare mission. Planet, which operates a fleet of Earth-imaging spacecraft, built the craft to carry Google's payload. Weeks of commissioning followed: engineers powered up the host computer and the TPU trays and ran early machine-learning checks before attempting a full prompt. A Google spokesperson told TechCrunch that the satellite is operating nominally.

The hardware on board shows how limited the experiment still is. The prototype carries four Trillium-generation TPUs, the chips Google designs to run its AI models, fed by solar arrays rated at roughly a kilowatt. Reported by TechCrunch, power constraints mean a single chip can run at a time, for about 20 minutes per session, which also complicates scheduling contact with ground stations. Nothing in the test changes what Gemini users see today; the satellite ran inference, using an existing model to generate an output, rather than training a model in orbit.

What the test proves, and what it does not

The result proves a narrow but real point: a chip built for a liquid-cooled data-centre rack survived launch, powered up in orbit, and completed a prompt-to-answer loop with the reply reaching a ground station. It says nothing yet about speed, cost, or how long such chips last in space, and Google's announcement gave no figures for any of those measures. Other companies have run AI models in space before, so Google is not claiming the first model in orbit in general, only the first response from a chip of this class.

Context matters here because the idea behind Suncatcher is much larger than a poem. Google announced the project in November 2025 alongside a research paper on space-based AI infrastructure, sketching clusters of solar-powered satellites carrying TPUs and linked by lasers. The argument is that orbit offers near-constant sunlight and natural cooling, two of the biggest costs for AI data centres on Earth. A rival effort points the same way: Starcloud, a startup backed by Nvidia, flew an Nvidia H100 graphics processor on its Starcloud satellite in the same month and later trained a small model in space. For readers following artificial intelligence, the pattern is clear: the industry is testing whether compute can move off the ground at all.

The harder tests are still ahead, and Google says so itself. The next phase will examine how the TPUs handle thermal stress and radiation exposure in orbit. In space, heat has nowhere to go except through radiators, and a particle strike can flip a stored bit, corrupting data without setting off the usual checks. Lab work published by Google researchers earlier in October tested hardware against a proton beam and found high-bandwidth memory the most sensitive part, a result that sets up the in-orbit radiation tests now planned.

What comes next for Suncatcher

If the TPUs keep working reliably, Google plans to launch twin interconnected satellites in 2027 to test high-bandwidth optical links between spacecraft. Those links are central to the wider design, since a useful orbital data centre would need satellites to talk to each other, not just to the ground. The prototype now in orbit sits in a standard orbit and depends on batteries, so it does not enjoy the near-constant sunlight the full Suncatcher design relies on, a gap the next satellites are meant to close.

For now, Suncatcher remains an experiment, not a product, and Google has published no commercial deployment date. The economics are unproven: launch costs, the short working sessions, and the unanswered radiation questions all stand between a haiku and a working orbital data centre. The test is still worth noting for what it settles. A data-centre chip has now answered a prompt from orbit, and the next year of testing will show whether it can keep doing so. Further science coverage will follow the twin-satellite launch when Google sets a date. Sources: TechBooky and Hardware Busters.