what I have in mind for a video-centric 16-18Mpix sensor is:
* it makes sense to build it for the nov3 camcorder, then put it also on a video-centric, cheaper, DSLR (shared R&D costs, not too many videocamera canibalization because of the lack of bells and whistles that the pro video guys need)
* you can do two very sensible things with such a sensor, depending on how good it is (and how expensive it is to make it):
- put it in a cheaper version of the 5D2 body (same size, same AF, same everything, but no magnesium and/or no weather sealing; also, swivel screen), call it 6D or 8D, and sell it as an entry-level full frame camera for <$2K; the video-centric sensor delivers 1080p60 4:2:2 without aliasing/moire, which is absolutely awesome, but there is no manual audio, XLR inputs, zebras, waveform, peaking, etc., so you can still sell the camcorder for 4x the price; resolution is lower, AF is old, body is weaker, so you can still sell the 5D2 for $200 more, and eventual 5D3 for $3K
- put it in a close-to-pro-but-not-quite body, with the 7D AF or better, and make it able to shoot short bursts of, say, 50 stills in a second, stored in full raw, with full electronic shutter (so it is rolling shutter, but a fast sensor designed for video could make this a very minor problem; the bigger problem, as discussed here, would be: no AF between shots unless -quite unlikely- pellicle mirror); call it 3D or 2D, and sell it for $5K, still under the videocamera price, which shoots the same video but has all the bells and whistles; it caters to a very specific niche so you also keep on selling the eventual 1D+1Ds reflesh blend (high speed, high MPix, low light, colossal price)
edit: also, for people thinking RAW video, or even 50 full raw still in a second, is too much bandwith direced to your poor CF card... yes, it is, but if you put a thunderbold port in the camera (as Canon is rumoured to be doing relatively soon) you can write directly to an external SSD drive, which brings all this closer to attainable
but what do you mean by easy 4-to-1 pixel binning?
I mean: read the full 5K sensor, then average every 4 neighbouring same-color photosites (R=r+r+r+r, etc) to create a 2.5K bayer pattern; store that (which would be awesome) or, in a second step, debayer this to get a sharp 1080p image without aliasing/moire, with lower computational power required (uninformed guess; ur may require digic V to have a dedicated pipeline for averaging in the first step, but I don't think you need a lot of transistors for that)