Thing is, if the primary source of truth is in the database, often it's beneficial to do the computation as close to the data as possible, as the bottleneck is

Thing is, if the primary source of truth is in the database, often it's beneficial to do the computation as close to the data as possible, as the bottleneck is likely to be I/O, not computation. "just load the data into memory" requires the database to provide to the client all the data that might be needed for that algorithm, which is more load on the DB server than scanning the (much smaller) subset of the data that actually gets traversed during the algorithm directly on the database itself and providing just the results, as that requires less I/O. That's the reason why it's often more efficient (even if we care only about the server-side load) to do filtering, aggregation and summarization for tabular data in the DB instead of loading all the data into a client and calculating it there, and the same applies for graph data.