Because orders need to be processed in exactly the order they are received.  This is inherently a "single-threaded" process that can't really be parallelized very efficiently.

That said, they seem to have two monster servers that do the work of the trading engine, website, database & api.  I would imagine that work could be split up more efficiently by having a separate machine dedicated to the trading engine, other machines serving the web front end, another for the database server, etc.  They may be doing something like that internally with VMs already though.