Time. Bugs and features coupled with (tight) deadlines will push aggressive optimizations "for later". This is especially true if the product is performing adequately and nobody really wants to mess with it lest they introduce unknown regression bugs.