I've spent a bunch of time thinking about ways to do something like this on a budget. The general idea would be to have a bunch of bare camera sensors directly connected to commodity PCs (maybe with Arduinos interfacing between them) with each PC driving multiple camera sensors.
The bandwidth requirements are quite ridiculous when you want to take a hundred photos simultaneously, so I'm not sure how many sensors I could really drive with one PC.
Not sure what you mean with "bare sensors", a sensor won't do you much good without shutter, lens, and a way to control the aperture?
To go cheap, I'd replace the DSLRs with webcams. It sacrifices resolution (obviously), but reduces bandwidth and computation requirements, and lets you play with the effects until you get a client who can foot the bill for a high-res version.
You are very right. I really meant "the parts of the camera minus the embedded system that drives the whole thing", but that wasn't very succinct. :-P The application I was thinking of was indie movie making, so I would want a decent resolution up front, and I wouldn't expect a high-paying client to be forthcoming.
Along this tangent, I talked today to a friend of mine who is convinced that I would be better off using some of the lower-end consumer point-and-shoot cameras, or even purchasing old film cameras.
That question is about time lapse with a single camera, but it includes some camera suggestions, including the GoPro action cam and Canon P&S with the CHDK hack:
http://chdk.wikia.com/wiki/CHDK
I agree with your suggestion on the webcams. Another approach is to use machine learning to generate the high-definition image from the mutual information contained in the multiple images by using probabilistic estimators. As an example, [1] has empirically shown decent results using MAP estimators.
The bandwidth requirements are quite ridiculous when you want to take a hundred photos simultaneously, so I'm not sure how many sensors I could really drive with one PC.