I'm really surprised that 300ms at startup would result in 25% fewer pods.... What % reduction in the total startup time is that?

Is it possible the prior measurement happened during a high traffic period and the post measurement happened in a low traffic period?