where do people get these numbers? You should move to physical hardware well before $200k/month.
To be fair, that is part of the cost. However, unless you are really big, you will just pay remote hands which is like $100-200/hr. The number of incidents are probably way too few, even at 200k a month, for 24/7 on-site sysadmins/techs.
[1]http://www.webhostingtalk.com/showthread.php?t=920078 [2]http://redd.it/16m9pq
A decent internet connection from level3 or the colo house internet for another few hundred.
So even at $2k/month we've still got $180k/month to buy servers and storage (or $2M for the year).
$200k/month is a lot of money.
So you build your colo'd private cloud on openstack, and then use their's to handle surges. As your traffic grows, you add your own gear to handle the base line.
But ~200K per month, you're operating at a very different scale from 2k or even 20k. Having a couple of idle servers 23 hours a day is not wasteful at that scale, unless your CIO is Ebeneezer Scrooge.
The failure rates vs physical hardware are incredibly high (just personal experience but I doubt you'll find many who disagree).
Once your systems are up and running you don't need to do much. Remote hands will replace a hard drive for $100, for example.
1) you aren't including operations staff.
2) what are you doing for DR/HA?
3) how about CDN etc?
4) how about load balancing between datacenters? DNS failover?
5) how many units of computer would $200k be at amazon? I'd imagine a lot. I recently (nearly a year ago) specced out 16 servers on AWS with HA/DR on opsworks for $15k over 3 years ($5k/year). Back of the envelope calculations make that 640 cores for your $200k. With 400GB transfer a month.
6) what are you doing for backups?
7) what if you don't need to utilize a whole rack? Most of the time systems don't fully utilize racks. But you're paying for them anyway.
1- you still need staff to deal with aws. I would argue that you need more.
2- HA is the same; DR is different. Use AWS for your DR ;)
3- What about them? there is literally no difference
4- see 3
5- a lot, I'm sure. Enough to justify owned physical equipment that you aren't sharing and doesn't get "retired" every other week
6- whatever I want, including putting them on glacier or sending them to my data center's cheap tape storage
7- it's a trivial part of the cost, why worry about it? or rent half a rack if you don't mind taking on a little extra cord-pulling risk
I think EC2 makes total sense for variable load (I paid like $0.25 to run some Coursera homework on a huge computer for 15 minutes and that was amazing) but if you're running something that needs to be up constantly and reliably (e.g., a database) EC2 gets expensive pretty fast.