A 60-kilogram satellite trained NanoGPT in orbit on Shakespeare's works

A 60-kilogram satellite called Starcloud-1 has trained a language model in orbit, Starcloud announced on 10 December 2025. It is the first publicly disclosed in-orbit training run.

The model was NanoGPT, Andrej Karpathy's compact character-level demo trained on a Shakespeare dataset, not a frontier LLM. A separate Google Gemma model ran on the same satellite for inference, but with Earth-trained weights, not orbital training. The two demonstrations are routinely conflated in coverage, and the distinction is the point: this is a proof that a data-centre-class GPU can boot, survive launch, and run for weeks in low Earth orbit. It is not the arrival of "ChatGPT in space."

Starcloud-1 launched 2 November 2025 aboard SpaceX's Bandwagon-4 rideshare from Cape Canaveral. The bus is roughly refrigerator-sized and orbits at about 500 km with a 95-minute period, catalogued as Starcloud-1 / Lumen 1 (NORAD 66303, designator 2025-248L). The chip is the same Nvidia H100 used in terrestrial AI data centres; NVIDIA said the mission carries 100x more GPU compute than any prior space-based operation.

The open question is why anyone would want orbital AI compute at all. Latency, downlink savings, sovereign compute, and training-versus-inference economics remain unsettled. Starcloud has filed for an 88,000-satellite constellation to scale the thesis, and closed a $170 million Series A at a $1.1 billion valuation led by Benchmark.