One can imagine uses for sensors with ultra high sampling. If you can make a sensor that has sufffcient sampling to outresolve the lens, it is possible (IMO) to eliiminate the AA filter without penalty.
Yes, however... Good lenses (for general imaging) have a limiting resolution of 150 cycles/mm or better (tiny dedicated lenses are much better). That would need an R+G+B sampling density of approx. 3 micron or smaller. In case of a Bayer CFA that would mean something like a 1.5 micron sensel pitch. At a sensel pitch of 1.5 micron, the sensels will have approx. an effective surface of 1x1=1 square micron. If we assume a storage capacity of 1500 electrons per square micron, the photon shot noise is approx. sqrt(1500)=39 with maximum exposure. A signal to noise ratio of 39:1 (at best) will look quite poor (6-7 accurate shades per byte). Such small sensels will also suffer from diffraction at any aperture narrower than f/1.4 .
Assuming the packing fraction of the sensels can be close to one, there is little penalty in term of the total number of detected photons. The designer is then free to use post-porocessing as desired. For example, one could resample at a lower sampling frequency, recovering signal-to-noise ratio, at the expense of resolution without introducing Moire or alaising artifacts.
A 24x36mm sensor array with 1.5 micron pitch sensels will be approx. 16000x24000= 384 Megapixels. You'd need a lot of storage and a lot of processing power to process each image. Also, such small sensels are not very sensitive and will need a very low ISO to keep sensor noise limited.
That seems to be quite a penalty just to replace the AA-filter, if you ask me. And it'll require massive averaging to get anything resembling a decent dynamic range, and will result in a number of output pixels that we already achieve routinely today.
No, to me the most plausible solution is still many large (say 9x9 micron) sensels and 16-bit ADC processing.
Bart