>Actually, no it isn't and at scale it makes it even more expensive, because you need redundancy. We are talking about RAID and CDN, not to mention that at scale hardware breaks a lot more often.
That's related to reliability not scaling per see. Building a highly reliable system is far cheaper when you reach the scale of Google or Amazon because you can rationalize purchase. Actually, I bet price of hardware can totally be neglected next to admin and bandwidth cost.
> On the other hand I've got 50 GB worth of data uploaded on Flickr. You can't tell me that 50 GB of redundant on-demand storage is worth $25 / year, unless you live in some kind of fairy land.
I must live in fairy land. :-) You can buy 4 decent 1TB for 400$ that will allow you to cater for 20 users at 20$ per user. Conservatively you can expect them to live 3 years on average which brings the cost to 7$ per user per year. Add a 400$ server good enough to do software RAID. Let's say 5 years of life expectancy it adds 5$. It's 12$ per user plus network cost on a lan.
Flickr uses better material and pay bandwidth and admin. But they have economy of scale and can put far more than 20 50GB users on 1TB on average because a lot of them don't use them fully. I think they are above break even point at 25$/year.