We develop a Graph Foundation Model (GFM) using HydraGNN, an open-source framework for large-scale graph neural network training. Additionally, we introduce a multi-task parallelism approach that distributes individual output heads across GPU-accelerated computing resources, enabling efficient training on multi-source, multi-fidelity datasets. Our model was trained on over 24 million atomistic structures aggregated from five datasets and evaluated on the Perlmutter, Aurora, and Frontier supercomputers, demonstrating efficient scaling across all three heterogeneous HPC architectures.