Graph neural networks (GNNs) trained on large graphs usually depend on in-memory graph storage and hand-built data processing pipelines for both training and inference. According to a new paper on arXiv, this common setup can become a bottleneck as graphs grow, because the data must be held in memory and managed with custom code.

The paper argues that graph databases offer a different foundation: scalable storage and efficient traversal for graphs that exceed memory limits. Rather than exporting graph data to a separate deep-learning framework, the authors propose moving GNN training and inference closer to the database itself, potentially simplifying the pipeline and improving scalability.

Because the abstract is brief, the exact methods and results are not yet available in this summary. But the direction is clear: instead of bolting GNNs onto ad hoc infrastructure, the paper suggests that graph databases themselves could serve as the execution environment for graph learning. This would align graph storage, traversal, and model computation within one system.