Foundation models have transformed language, but can they conquer the final frontier of AI: the complex, interconnected world of graphs? Graph Foundation Models (GFMs) promise a universal AI for networks, capable of unlocking discoveries in everything from social dynamics to drug development and supply chain analysis. This talk cuts through the hype to offer a frank assessment of the current state of GFMs, outstanding challenges and lays out a realistic course for the future, structured around four key takeaways: → The GFM Paradigm: Understanding how large AI models are being adapted to learn universal patterns from network data. → Core Architectures: A critical review and comparison of the different foundation model architectures being developed for GFMs. → Key Challenges: Delving into the major obstacles facing GFMs, including scalability bottlenecks, data scarcity and diversity, the complexities of non-Euclidean structures, and fundamental architectural uncertainty. →Next Steps: Identifying the crucial research and engineering breakthroughs required to make GFMs a practical reality.