When I saw the DynamoDB pricing, I assumed the setup was something along the lines of "N=3, writes go to all replicas and wait for 2 responses, consistent reads go to 2 replicas, inconsistent reads go to 1 replica". Given that DynamoDB is using SSDs, I don't see why they would need to broadcast read requests to all replicas -- that's only useful if you have high variability in your read latency.

Someone who wanted an interesting research project might be able to pull back the covers on DynamoDB a bit by issuing a large number of requests, very carefully measuring the response latency curves for consistent and inconsistent reads, and looking at what model fits them best.