Graph AI now runs 500 times faster on huge datasets — without losing accuracy

Researchers built a faster attention mechanism for graph neural networks that reduces memory use from quadratic to linear, meaning the same GPU can now process graphs with 500,000+ nodes instead of a few thousand. This removes a hard ceiling on scale — the type of problem that's been bottlenecking real-world deployment of graph AI in logistics, molecular simulation, and network analysis.

Graph transformers have been stuck. All-to-all attention — the thing that lets them see long-range relationships in complex networks — costs exponentially more memory as graphs grow. Every workaround trades either speed or accuracy, leaving practitioners choosing between slow-and-right or fast-and-wrong. This mechanism (k-MIP attention) selects only the most relevant nodes per query instead of looking at everything, which cuts memory by orders of magnitude while proving mathematically that it doesn't lose expressive power. What changes: researchers and companies can now run graph AI on real-world scale without engineering around the constraint. That unlocks new categories of problems — supply chain routing, protein folding, drug discovery on massive molecular libraries — that were previously compute-prohibitive.