Seems like a better approach would be to add a "gain pixel" to traditional sensors.
So imagine you've got your Bayer grid (or a layout that functions similarly), one of the pixels in the grid gets read out first. That one sets the gain for the ADC for the surrounding pixels. The downside is your iteration couldn't be linear and the gain control would be a lot more complex, but you could do it with technology that actually exists, and you wouldn't lose half of your vertical resolution by doing it a whole line at a time like Magic Lantern's dual-ISO mode.
You'd have to do fancy processing around high-contrast edges, but what else is new.
The underlying idea is sound -- storing bits is a linear problem but storing photons is an exponential problem. Of course you're still going to have to dump exponential amounts of energy... (So forget the "unbounded" part.)
No system is ever unbounded, period, unless you live in a world of frictionless spherical cows. At some point you run into issues caused by accuracy, energy, capacity, etc.
Oh the 1/10,000,000th comes from assuming they need to reset 128 * 8000 times per second to merely equal current sensor performance before losing sensitivity to misc losses.
If they want one more bit of DR they need to cycle in half as much time, and so on. Similarly, if we generously assume they're only losing 50% efficiency to cycle time and real estate then double again.