We present GPSocio (Reproducible code is available at: https://github.com/tracy3057/GPSocio .), a general-purpose social network representation system designed to support diverse downstream analytics and enable knowledge transfer to data-scarce domains. While prior methods optimize embeddings for specific tasks, they often lack generalization. GPSocio leverages the emerging Graph Foundation Model (GFM) paradigm by aligning social graph structures with Large Language Models (LLMs) through post propagation sequences encoded in natural language. Extensive evaluations across four downstream tasks show that GPSocio consistently outperforms strong baselines, achieving 6.74% gains in User macro-F1 for Sentiment Analysis, 21.29% gains in User ARI for Ideology Classification, 14.75% AUC improvement for Static Link Prediction, and 62.87% User N@10 improvement for Temporal Link Prediction, demonstrating robust modeling of social semantics and diffusion dynamics.