This statement is false for people writing large scale distributed systems. It turns out tens or hundreds of thousands of computers are more expensive than engineering time. No amount of clean will fix gaping million dollar holes in the budget.
And given migration of large clusters from one piece of code to another is expensive you want to think about performance ahead of time.
Isn't the primary requirement for writing a highly scalable distributed system to have clean and strictly decoupled code?
Otherwise scaling becomes a series of rewrites.
And in that scenario engineering time may be less expensive than tons of iron, the passing of said time is a killer. In most software project, time is a bigger issue than money.
And given migration of large clusters from one piece of code to another is expensive you want to think about performance ahead of time.