I don't think you're misunderstanding anything at all My opinion of what constitutes "middle of nowhere" may differ from the average San Franciscan's, but I've

I don't think you're misunderstanding anything at all My opinion of what constitutes "middle of nowhere" may differ from the average San Franciscan's, but I've had the opposite estimation of costs that jedberg did. That is, EC2 was 20% more expensive than operating ones own hardware in a South Bay datacenter for a year. This would have been about a year ago, but, IIRC, including operating costs, one would break even at around 10 months of unreserved EC2. Pricing has certainly changed in the meantime, so 30% in the other direction is well within the realm of possibility. Back then, it cost about $3 per year per watt. Now it's pushing $4. The trick is, in the case of ones own hardware, one usually pays for the available capacity, not just what one uses. On a 500W server, that's $2k annually. That's the flexibility that AWS offers: one doesn't pay for the operating capacity, only the use. It would seem, however, that one pays for the hardware up front either way. The great increase in flexibility of operating ones own hardware is not suffering from only ordering off an abbreviated menu. Germane to the OP's question, I have never found any virtual or dedicated server offerings come close (say, a factor of 10) to the I/O performance of something custom configured/assembled (even from inexpensive, commodity blocks). Whether this constitutes a problem is subject to interpretation In general, I remain skeptical that there exists in The Cloud cheap-enough commodity servers suited to even a majority of applications. Remember, even Google had custom hardware very early on.