The D800 is halfway there; analog gain ends at ISO 1600. Now all Nikon engineers have to do is realize (doh!) that there is nothing to be gained by bit-shifting the data and throwing away all that highlight info, and put the ISO setting in metadata.
Thanks; that is nice to know. (As an aside, I recall many years ago that with the Olympus E-1, ISO settings beyond 800 had no effect on raw except through meta-data: is this common?)
And the rough figures I have seen suggest that full-well capacity on these sensors is in the range 30,000 to 60,000, or 2^15 to 2^16, so with 14-bit ADC output, there are 2 to 4e- per level at base ISO speed, which is 200 or less, so that by ISO 800, there are already at least as many levels as electrons.
But as you indicate in this comment
it's not the 'unity gain' which determines when pushing lower ISO competes with 'normal' exposure at higher ISO; rather it's when read noise in electrons stops decreasing.
the relevant maximum useful gain level is about "one level per photo-site dark noise electron count", rather than counting every electron, and that noise floor seems to be at least 2e- of dark noise from the well itself, so even one level per 2e- is enough, and that probably happens by ISO 400.
But speaking as one semi-physicist to a real physicist: it is not that increasing ISO speed decreases the number of electrons of read noise, since that converse of output noise levels back to an electron count is at best a convenient fiction, converting measured output noise levels back on the false assumption that all noise comes from noise present in the charge in the electron well, before ISO gain. I would prefer to say that the useful threshold for analog gain is when any additional increase in gain produces a linearly proportional increase in read noise, measured downstream of the amplification. Or more simply,
"further analog gain is only useful if it improves the S/N ratio in the output signal".