All valid points if compared against a single-instance RDBMS. Except that modern RDBMS used in any large scale web service will also be guilty of these points. When you shard/cluster your data across various master/slave DB instances you essentially lose the ability to store your data in a non-atomic normal form, just like SimpleDB. In both a sharded RDBMS and SimpleDB your data is denormalized.

With a DB setup like this (and all major web sites are) you cannot feasibly join, group by (or do associated aggregations), or order by anymore anyway. Constraints are not really a good idea either. It would be nice if SQL was still useful in a situation like this but it just is not.

The days of hugely massive single machine databases is coming to and end if not already dead. Even if you are reporting on massive amounts of data and need to do aggregations across large datasets, then OLAP is increasingly becoming the way to go (again, with fact tables and cube partitions sharded across multiple instances to increase query performance).