NVidia does have all of Transmeta’s IP and from outside this looks like an Efficeon with some capacity to execute ARM instructions in hardware letting them only optimize the hot spots, and a few other changes. I don’t see that translating ARM instructions into Mill instructions would be any more difficult than what Transmeta initially did with VLIW, but I suppose that might be the reason NVidia seems to be doing better than Transmeta did.
Good luck licensing your load semantics to NVidia then? 