I'm thinking this is going to be more for "batch compute" type operations where you'd probably have data replicated across multiple sites and the latency won't impact the job significantly.