I have little real-world Hadoop experience but...
...isn't this just telling us that a solution designed for a problem is better than a general purpose cloud?
My limited experience with Hadoop shows that it is very sensitive to RAM and I/O bottlenecks (both disk and network).
If the off-the-shelf servers aren't as well matched in RAM (16 threads for 8GB RAM verses 4 threads for 4GB of RAM), disk (I'm unsure how the SM10000-64 presents disk to the compute boards - is it as one large pool, or multiple smaller pools) and network I/O (my Hadoop tests easily hit 1Gbps).
i.e. how would the Xeon servers have faired if there were fewer of them and they used more RAM (say 2GB/thread), more local disk and bonded 1Gbps network adapters to reduce the bottle necks.
Management of the SeaMicro box will probably be easier.