Starcloud trains LLM in orbit; orbital data center plans

Starcloud reports the first LLM trained in space and outlines orbital data center ambitions. - Starcloud trained nanoGPT in orbit on Starcloud 1 - An H100 GPU was used on the satellite - Models run in orbit included Gemma and a variant of Gemini - A test message returned: “Greetings, Earthlings” - Orbital compute taps near-continuous solar energy - Space avoids terrestrial cooling and water use constraints - Starcloud proposes a five-gigawatt orbital data center - Google, SpaceX, Blue Origin also pursuing similar plans - Orbital solar cited as 4–10x more efficient than ground systems
Key insights
First LLM training conducted in orbit
Starcloud used its Starcloud 1 satellite to run and train language models in orbit, including training nanoGPT on the complete works of Shakespeare and running models such as Gemma and a version of Gemini. The posts state an H100 GPU was operated aboard the satellite to enable these experiments. A test message from an in-orbit model to Earth read “Greetings, Earthlings,” which the posts present as an example of an operational in-orbit model communicating with ground systems.
Energy and cooling advantages cited for orbital compute
Posts emphasize orbital compute advantages: direct access to solar energy reduces reliance on terrestrial grids and water used for datacenter cooling, and the cold environment avoids many ground-based cooling constraints. The posts claim orbital solar power is between four and ten times more efficient on average than ground-based systems, and that satellites receive direct sunlight for roughly half of each 90-minute orbit despite eclipse periods.
Starcloud’s five-gigawatt orbital datacenter plan
Starcloud articulated ambitions for a five-gigawatt orbital data center, framed as a scale that could outpower the largest power plant in the United States. The posts state this plan as Starcloud’s target and note that large technology companies including Google, SpaceX, and Blue Origin are pursuing similar orbital compute and data-center concepts.