I wonder how it scales beyond that.
I've found that graph queries only have acceptable performance if the graph fits in local memory.
Sparql can scale significantly beyond local memory size. As it is really a relational algebra. Especially for the pattern matching style queries that most Cypher and sparql queries really are. Traversal depend on locality and depend vastly on the implementation of they scale of of ram or not.
Does anyone know if it's leveraging any of the graph database?
Wikidata, uses Blazegraph and the shown services run against that using the SPARQL api/general query language.