Google successfully activates prototype satellite for space data center project

AI model Gemma 4 generates first response from Earth orbit

Commercialization still faces hurdles including radiation and cooling challenges

Google logo [Reuters]
Google logo [Reuters]

It called Earth a "blue marble" and sang that starlight "feeds my core."

The poet was an AI drifting in Earth's orbit. Google said it successfully received the first AI-generated response from a prototype satellite launched as part of its space data center initiative — marking a step forward in the effort to build next-generation computing infrastructure beyond the atmosphere.

Google announced Friday (local time) that it had received the first AI response from the prototype satellite for "Project Sunseeker," its space data center program.

Researchers on the ground asked Gemma 4, Google's open-source AI model aboard the satellite, to write a haiku about what it feels like to be in space. As the satellite passed over a ground station in Svalbard, Norway, it replied: "Blue marble below / Data streams in silent grace / Starlight feeds my core."

The haiku is a traditional Japanese short-form poem popularized in the United States as an internet meme — and, coincidentally, the name chosen by Google rival Anthropic for its lightweight AI model, Claude Haiku.

The response marks the first AI output ever generated by a data-center-grade tensor processing unit (TPU) operating aboard a satellite in Earth's orbit.

Google's research team said it then ran additional tests, including a five-line poem on a space theme and a diagnostic check on the satellite's onboard systems.

"Our chips have officially earned their 'space wings,'" Google said, calling the milestone "an important achievement on a long journey toward cosmic ambitions."

Google launched the satellite aboard a SpaceX Falcon 9 rocket on Oct. 1, and after three weeks of orbital commissioning and subsystem verification, began full machine-learning computation tests.

The satellite carries four TPUs and generates up to 1 kilowatt of power through solar panels.

However, the TPUs draw too much power to run simultaneously — only one can operate at a time, for up to 20 minutes per session, according to US tech publication TechCrunch.

Google plans to continue testing whether the TPUs can perform computations reliably and without data errors under the extreme temperature swings and radiation exposure of orbital conditions.

Once that technical validation is complete, Google intends to launch a pair of satellites next year equipped with higher-performance TPUs and optical laser communications.

Starcloud, a startup backed by Nvidia, had already run Gemma on its Starcloud-1 satellite — which carries Nvidia's H100 GPU — in December last year, receiving an AI response that began with "Hello, Earthlings."

SpaceX CEO Elon Musk has also said the company plans to launch AI satellites powered by Nvidia chips in the fourth quarter of next year, with the constellation expected to reach significant scale by 2028.

Experts caution, however, that space data centers still face significant obstacles — including cosmic radiation, the difficulty of on-orbit maintenance, thermal management and space debris — and that full commercialization is likely still some years away.


arete@heraldcorp.com