According to @AdiOltean on Twitter, nanoGPT has become the first large language model (LLM) to be trained and used for inference entirely in space, leveraging an Nvidia H100 GPU aboard the Starcloud-1 satellite (source: https://x.com/AdiOltean/status/1998769997431058927). The Starcloud team successfully trained nanoGPT, based on Andrej Karpathy's architecture, on the complete works of Shakespeare and demonstrated inference capabilities on both nanoGPT and a preloaded Gemma model. This milestone highlights the potential to shift high-performance AI workloads off Earth, utilizing space-based resources and abundant solar energy. The success sets the stage for new business opportunities in AI-powered satellite computing, distributed cloud infrastructure, and green AI innovation (source: @karpathy).

According to Andrej Karpathy (@karpathy), nanoGPT has evolved from a simple educational repository into a benchmark for recursive self-improvement in AI model training. Initially created to help users understand the basics of training GPT models, nanoGPT now serves as a baseline and target for performance enhancements, including direct C/CUDA implementations. This progression highlights nanoGPT’s practical utility for AI developers seeking efficient, lightweight frameworks for rapid experimentation and optimization in natural language processing. The project’s transformation demonstrates clear business opportunities for organizations aiming to build custom, high-performance AI solutions with minimal overhead (source: @karpathy, June 30, 2025).