One of the problems with having a 'dumb' DAC (worst case: a sound card) depending on a host system is performance. Especially when your host is running a poorly programmed operating system with a dodgy driver model (e.g. Windows), you may very well run into some timing issues, or even crashes.
When you have the Raspberry Pi on board the projector, the DAC *is* the host system. Think of it as having multiple computers with their own instance of your favorite laser show package running simultaneously, one in each projector. They're not dependent on each other or on a specific host to get their frame data, only on the network for synchronizing them together and for interfacing with the laserist.
I'm not saying you should get rid of the host system at the front of house altogether and have some dumb terminal at that end, on the contrary -- I only suggest moving the frame generation logic to where it is needed most (the projectors), freeing up more processing power on your host for things like new user interfaces and really processor intensive stuff that is not timing critical (as opposed to the frame generators, which need to be fairly real-time), like modeling or rendering in 3D, processing video and other complex operations. When that's done, a new bunch of vector data can be streamed to the projectors which can use them on demand.
The only thing the DACs have to do is crunch a bunch of points (sets of X, Y coordinates), which is an operation that can be done on that ARM 700MHz processor at a rate of millions per second. So I doubt we're going to run out of processing power to do complex frame transforms there in the future.
Just for the sake of comparison, how much processing power is there on high end boards like the QM2000.NET or Lasergraph? Judging by the specs, the Raspberry Pi could run rings around those boards.