If I read the blogpost it sounds they didn't care as much as for memory in general, but in the datasegment.
TP was notorious for only allowing one (static) datasegment, they never implemented the multi segmented options.
So one could argue it was a hack to not having to change compiler architecture too much in what they knew to be their last 16-bit compiler series.
And of course multi segment datasegment models carry an overhead (since any global variable must be loaded as far pointer, iow with a segment), so even introducing then would cause whining that programs were slower.
I experimented with such options in Topspeed compilers on a Cyrix 686 p166+, and it was generally a few percent.