[jules]

People generally seriously underestimate the power of modern hardware. Suppose you have 100 million requests in a month. That's a small country of people accessing your website, so that will require a large bunch of big iron, right? No. A single commodity PC box can easily handle that. 100 million requests per month is just under 40 requests per second. Say your peak is around 100 requests per second. If your site is amenable to caching you can probably handle that load on a single core. Even if your site can't be effectively cached, for $5000 you can buy a terabyte of RAM these days, which can probably hold the hot parts of your database (for comparison: compressed current-revision-only of Wikipedia is under 8GB). Of course you need expensive hardware to be able to put that much RAM in, but then again you probably don't need anything close to a terabyte.

[romaniv]

People generally seriously underestimate the power of modern hardware.

They also often underestimate the bloat of modern software. The difference in page generation times between WordPress and some minimalistic CMS can be in the order of 50 or 100. And while it's compelling to believe that WordPress "does more", I don't believe that's the case.

[jules]

Absolutely. A modern quad core $100 processor doing 10 requests per second has about 1 billion clock cycles per request. You can easily lose a factor of 100 by using an interpreted language, another factor of 10 by using a bloated framework, and another factor of 10 by using a DB optimized for disk instead of one optimized for RAM.

Due to network overhead (in performance, but more importantly in complexity when programming and sysadmining) and the speed and size of modern hardware, for the vast majority of sites it makes much more sense to run on 1 powerful box than on 20 small boxes. And even if you do need to scale to multiple boxes eventually (which >99.9% of websites won't have to), managing 10 boxes instead of 200 has advantages too.