You could do a proof of concept where you replicate a single table to 2500 databases on a different server.

Just write a program that generates the scripts and off you go.

Still the real issue is the the performance of the distributor.

Are you planning to use pull subscriptions to move some of the load off the distributor?