This was disappointingly dishonest: they shave the AWS numbers down to a few grand under the self-hosting cost and end on a note saying these analyses should be done apples-to-apples, but neglect to point out the far more expensively flawed methodology in the self-hosting cost: buying and colocating 131 dedicated servers based on the peak requirement of 131 leased EC2 virtual servers.
I don't understand why the methodology is flawed: the objective isn't to price N servers. The objective is to serve all the load at the best price. If you need 131 during peak and 20 during off-peak and can not pay for the delta- yay!
My company works in the media space and I love EC2- during the week our loads are low but when it's Superbowl Sunday or the GRAMMYs, the ability to spin up capacity is huge. But the fact that I'm _not_ paying for that capacity 24/7 is even bigger.
It's dishonest because they adjusted their price to be under the self-hosting price, then cautioned that these studies should be done apples to apples. As if they'd just corrected it. What they neglected to do was point out that the self-hosting analysis focused on buying hardware instead of leasing as you do with AWS.
Here's why it sucks:
131 x amazon extra large: $56k after Amazon fixed the pricing
- 524 amazon cores
- tons of crap io
- 2tb of ram
131 x bought servers: $61k
- this includes colo and bandwidth and amortized cost of buying
- excludes labor which Amazon makes mention of
131 x $289 leased servers [1]: $37k
- 524 real cores, each with hyperthreading
- tons of dedicated disk io (raid 1 SSD and raid 1 SATA!!)
- 2tb of ram
- server breaks someone fixes it
These dedicated servers are obviously going to shit all over EC2 virtual servers if you're doing anything that involves disk or cpu so you could in all likeliness shave another $10k - $15k off the hosting bill to match the AWS capacity.
For those interested it's Amazon's rebuttal to the story discussed here: http://news.ycombinator.com/item?id=3580273