XNews

Orbital compute / Starcloud trains LLM in space

Orbital compute / Starcloud trains LLM in space

Starcloud ran an H100 GPU on its Starcloud 1 satellite to train and run LLM experiments in orbit. - Demonstration: trained nanoGPT on Shakespeare and ran Gemma/Gemini variants in orbit; sent back “Greetings, Earthlings.” - Advantage: orbital compute taps near-constant solar energy and avoids terrestrial water-intensive cooling. - Efficiency: orbital solar power cited as 4–10x more efficient on average than ground systems despite eclipse cycles. - Ambition: Starcloud plans a five-gigawatt orbital data center concept. - Industry context: Google, SpaceX and Blue Origin also exploring space-based compute strategies.

Key insights

Starcloud demonstrated LLM training and inference in space

Starcloud reported using an H100 GPU aboard its Starcloud 1 satellite to train nanoGPT on Shakespeare and to run models such as Gemma and a Gemini variant. The satellite returned a test response (“Greetings, Earthlings”) indicating successful orbit-based inference and training experiments.

Orbital compute cited for energy and cooling advantages

Starcloud argued orbital data centers can leverage near-constant solar energy and avoid terrestrial cooling/water-use constraints, potentially achieving 4–10x efficiency compared to ground-based solar on average. The company envisions scaling toward multi-gigawatt orbital power systems.

Space is emerging as a new cloud region pursued by major players

The posts noted that Starcloud’s activities are part of broader industry interest in orbital compute, listing Google, SpaceX and Blue Origin as entities pursuing similar plans for space-based data centers and satellite compute.

Source posts