The NRE (non recuring engineering - in this case likely only mask genereration) on a project that does not hav significant volume would be so stagering that you wouln't be able to do it. You might want to try "contracting" with a local university. Back around 10 years ago when I graduated there was a VLSI course atht he University of New Hampshire as a senior project I did a VLSI chip in 2 um technology (which was way below the cutting edge at the time). These chips were fabbed in a short run with Mosis. I think we got back 4 packaged dies. Perhaps you could work with a student doing a similar project (and possibly help fund it) by cramming on a secondary function which can be selected with a bondout.

As a proof-of-concept this would work great...but you would likely be limited to one die in the end.