Frank Grießhammer. He did a talk on exactly this topic at ATypI Amsterdam.
My understanding is this. There are four groups of pairs if you kern between Latin and Cyrillic:
Typically the groups 2 and 3 are never accessed, yet the font renderer actively looks at them even if there is no pair occurring in the current text. It gets even worse when you add Greek to this.
I am just about to start kerning a Latin-Cyrillic-Greek font and I am terrified about the idea of having 9 groups of pairs!!! Does anyone know whether Glyphs finally manages to break the kerning groups by script, thus avoiding this subsequent processing issue?
When I tested this, splitting the classes by script automatically produced bigger GPOS tables. And if you keep the number of pairs at a sane number you should be save. All my fonts are done like this and I never had any problem.