Perhaps you guys are talking about different things? There are certain architectural decisions (schema design, etc.) that can dramatically impact future scalability. Sometimes if you implement it the "simple" way the design can very quickly reach its limits and have to be replaced.
And, honestly, sometimes you've done variants of the same problem enough that you know where some of the bottlenecks are inevitably going to be - so why not design them out right at the start?
And, honestly, sometimes you've done variants of the same problem enough that you know where some of the bottlenecks are inevitably going to be - so why not design them out right at the start?