You can now pay a per-request rate and skip the autoscaling / wcu & rcu math.
The more common use case is that you over-provision your provisioned capacity to be able to handle some spikes in the number of accesses well. While there was already DynamoDB Autoscaling, I suggest everybody using it checks out DynamoDB OnDemand, as that might (depending on the access patterns) still save a lot over DynamoDB Autoscaling and is overall way less hazzle, because you simply don't have to care about scaling the provisioned capacity at all anymore.
In OnDemand mode tables have practically no limits on throughput or storage (just like provisioned mode).
Disclaimer: I work for DynamoDB but comment is my own.