One of the most frequently viewed posts here is one that I made back in August of last year, on “How Many Servers Does Google Have?”. (the estimate, at that time, was about 900,000.) Recently, Wired published an article to answer a similar question for Amazon, specifically for Amazon’s Elastic Compute Cloud [EC2] service. It reports that Huan Liu, a researcher for Accenture, estimates that Amazon uses about 445,000 physical servers to power EC2.
This estimate is not easy to compare directly with the estimate for Google, because the two are for rather different things. The Google estimate was for the servers used to deliver its array of end-user services, including not only search, but also GMail, YouTube, Blogger, and so on. The Amazon estimate is, as I mentioned, for the EC2 “computing utility” that provides pay-as-you-go computing capacity; it does not, as far as I know, include the servers that power Amazon’s retail operations.
Another difference is that the Google estimates were inferred from data on electricity consumption provided by Google to Stanford professor Jonathan Koomey. Mr. Liu was not lucky enough to have similar data, so he uses an ingenious technique to estimate how many server racks EC2 is using, based on their assigned IP addresses, both external (visible via DNS) and internal to the EC2 cloud. (His original blog post, which explains the technique, is well worth a read.) He says he is reasonably confident of the estimated number of racks, and that his assumption of 64 servers per rack is an educated guess. (Te blog post is also clear about the caveats.)
Nonetheless, this is an interesting analysis. Even those of us who have had some experience in operating data centers generally have never seen anything remotely close to the scale of Amazon or Google. Maybe there’s something to this whole new-fangled Internet thing, after all.