Orbital AI Haiku Broadcasts From Svalbard
“Blue marble below / Data streams in silent grace / Starlight feeds my core.”
That verse represents Google’s first artificial intelligence response from a prototype satellite operating in orbit. Reported by yna.co.kr on October 31, 2026, the milestone marks a major step for Project Suncatcher, the company’s experimental space-based data center initiative. Operating over the Svalbard ground station in Norway, ground researchers queried the satellite’s onboard open-source AI model, Gemma 4, asking it to capture the feeling of being in space.
Tensor Processing Units in Space
The transmission marks the first time an AI response has been generated by a data center-class tensor processing unit running on a satellite in Earth orbit. Following the initial test, researchers executed additional prompts, including a five-line poem about space and technical telemetry status checks.
After years of research, Project Suncatcher is scheduled to embark on its first test in orbit, launching a prototype satellite to evaluate how Google Tensor Processing Units (TPUs) perform in space. Learn more about the science behind this moonshot. ⬇️ https://t.co/GGTH2tZOHZ
A SpaceX Falcon 9 rocket carried the prototype satellite into orbit on October 1. A three-week commissioning and subsystem validation phase preceded intensive machine learning compute trials.
Managing Extreme Thermal Loads
The spacecraft houses four TPU units powered by onboard solar generation capable of producing up to 1 kilowatt of electrical power. Yet high power consumption constraints limit operators to running only a single TPU at a time for 20-minute intervals.
Engineering teams plan to evaluate how the hardware withstands extreme orbital temperature swings and cosmic radiation without generating data errors. Ground and orbit validations will pave the way for Google’s next move: launching twin satellites equipped with high-performance TPUs and optical laser communication systems next year.
Commercial Race and Orbital Hurdles
The test places Google alongside competing initiatives in the commercial space compute sector. In December of last year, the Nvidia-backed startup Starcloud used its Starcloud-1 satellite—equipped with an Nvidia H100 GPU—to run Gemma and return an AI greeting.
Meanwhile, Elon Musk has outlined plans to launch Nvidia chip-driven AI satellites in the fourth quarter of next year, projecting significant commercial scaling by 2028. Despite these deployments, industry specialists note that space data centers face persistent hurdles. Cosmic radiation, maintenance limitations, data center cooling, and orbital debris hazards stand in the way of widespread operational commercialization.
