Fun little thought experiment:
If we go by sensorgen, the 1D-X's pixels have a full-well capacity (FWC) of 90,000. Since the pixels on this sensor are 7.5x larger, we can extrapolate that given similar sensor capabilities, the pixels on this sensor can hold ~675,000 photoelectrons.
Now, since each doubling of ISO halves the FWC, ISO 4,000,000 will yield a FWC of roughly 675,000/40,000 = 16.875. Let's be generous and round that to 20. That means white is made from 20 photons.
If we generously place middle grey at 3 EV below clipping, that'd mean midtones are made from 20/8 = 2.5 photons, which itself yields a signal with SNR of 2.5/sqrt(2.5) = 1.6, which is already below most reasonable DR cutoffs. In other words, you'll have ~3 EV dynamic range at best, assuming no read noise whatsoever (bad assumption).
So, either my calculations are way off, or there's a limit to these insane ISOs.
Thoughts?