That 1000 levels are lost to read noise.
The chart clearly shows the relative contributions of read and shot noise for the exposure zones, and shot noise predominates in the green area of the chart, which comprises 4080 levels, whereas read noise dominates only in the very deep shadows, comprising 15 levels at most. How can you lose 1000 levels to read noise, when it predominates in only 15?
OK, now your objection is clearer.
First of all, the issue I addressed has nothing to do with how much of the recorded range has more read noise than shot noise (and read noise is significant far above this range, especially if there is 1-dimensional or "banding" noise, which is much more visually potent than 2-D noise). The issue is *what* the level of read noise is in ADUs, as this will be the minimum noise found anywhere in the image, in a linear sense, relative to the steps of the digitized levels. If quantization doesn't happen where there is only read noise, it shouldn't happen at any other range, where shot noise increases the absolute noise level.
The premise is that with a read noise of 2.0 ADU at ISO 100 in the 5D (not the same as the 1Dmk2), the number of levels used can be divided by 2.0/1.4, or 1.43, without incurring any practical quantization in the RAW data. The other way to look at it is that the steps can become 1.43x their current size. In this case, that would mean that the number of levels between black and RAW saturation in the 5D at ISO 100 could just as well have been recorded with about 1000 less levels (about 2500 instead of about 3500). It is not about clipping levels away from the shadows (or highlights, or anywhere).