I've found the "developer time is more important than machine time in 99% of cases" argument to be particular effective when preaching python to the C++ masses.
It's effective, but less-so when the audience hasn't programmed anything Big before. If you've only done programming in school, and all of your C++ assignments take anywhere from 1 to 12 hours to complete, it's hard to see the dev time benefits.
Well, if we're talking about scale, there is a significant difference between needing 100 servers and needing a 10,000. As in, you suddenly have far fewer choices as to where you can physically put them! At that point, suddenly, developers start to look very, very cheap next to hosting costs.
If you just need to buy 2 servers instead of 1 then sure, yeah, developer time is more expensive than hardware.
The one I work on, for a start :-) A good deal more in fact...
But it's not just servers. You are up against physical availability of stuff, and stuff gets exponentially more expensive as you approach the limit. A programmer who is wasteful of memory on a single box will eventually reach a point where you simply can't buy more, you will have to rewrite the app to run on more machines, for example. At the small scale, sure, programmer time is expensive. But on an industrial scale, programmers are pretty cheap compared to say "building and operating an entire new datacentre".
If you are in the 10+ server league, you already are in that 1% where code efficiency and performance become more important than iteration speed or ease of development. In fact, it would be interesting to map where this transition occurs ;-)
And computers are cheap, regardless of how many you have. You say a datacenter with 50K servers is expensive, but imagine how bad would it be to maintain a building with 50K programmers...
Even if 90% of startups fail, surely more than 10% of the survivors will have Twitter-esque scaling problems if all their prototypes were on platforms with severe performance drawbacks.