A satellite carrying Google's custom AI processors has successfully performed machine learning inference above Earth. During a recent pass over ground infrastructure, the orbiting system ran the Gemma 4 language model and transmitted results back to the surface. According to Travis Beals, a senior director at Google, the experiment marks a milestone for Project Suncatcher, the company's initiative to explore computational workloads in space.
The satellite launched on 1 October as part of this research programme. When prompted during passage over a receiving station, the onboard model generated verse reflecting on its orbital vantage point. As the spacecraft passed above Svalbard, Norway, it transmitted the completed response. Google characterised the output as the first AI generation from a data centre-class TPU operating beyond Earth's atmosphere.
Blue marble below / Data streams in silent grace / Starlight feeds my core.
Gemma 4, on the Suncatcher satellite
Following this initial success, the team experimented with additional prompts, including space-themed verse forms and requests for the satellite to report on its own operational status. The model was instructed to compose its responses from the perspective of a TPU orbiting the planet.
Hardware and power constraints
The spacecraft carries four TPU units and produces up to 1 kilowatt of electrical power. This energy limitation means the satellite can activate only a single processor at any given moment, with each operational window lasting approximately 20 minutes. System validation required roughly three weeks, and Google confirmed the hardware was performing as expected.
Omar Sanseviero, a member of the Gemma team at Google DeepMind, noted via X that the model drew power from stored solar energy during execution.
Upcoming durability testing
The company's next phase involves evaluating how its TPU designs withstand the orbital environment. Space radiation poses particular challenges, as energetic particles can alter individual bits within chip memory, potentially corrupting stored information. Google's terrestrial testing indicated this phenomenon affects model inference less severely than the training process.
Should the chips demonstrate sufficient resilience, Google intends to deploy a pair of connected satellites in 2027. These follow-up missions will incorporate enhanced power generation and test optical communication links between spacecraft. The company's broader strategy depends on SpaceX achieving full reusability of its Starship vehicle.
Google is not alone in pursuing orbital computing infrastructure. Other companies are developing similar concepts, with Starcloud securing $250 million in funding during August. However, SpaceX has cautioned investors that its own plans for space-based data centres may prove commercially unviable.
Source: The Next Web
