Also don't forget that frameworks like Phoenix, or alternatively aggressive optimization in other ways, can be significantly faster at that 'first load' than you describe.