What about special-purpose implementations that need many code points?

For a particular implementation, if someone really, really wanted a representation that encoded more characters in a series of 16-bit code units then a series of private-use characters would work. For example, suppose you use a representation that consisted one BMP private-use character followed by one private-use surrogate pair (e.g. three 16-bit units). With such a representation, you can encode 6400 x 131,072 ( = 838,860,800) private use code points.