Very large numbers of Amazon servers are used for something else than cranking out HTTP pages. Expect a lot of them to be crunching numbers for bio-informatics problems, physics simulations and so on. That's why there is a CUDA enabled instance.
Rendering web pages is actually one of the worst use cases for Amazon from a bang-for-the-buck perspective, especially when you factor in bandwidth.
When people are building 30,000 core compute clusters [1] on EC2 - presumably with zero publicly available web servers, I'd be very interested in any methodology that provides reasonable estimates of revenue based on public web servers.
Rendering web pages is actually one of the worst use cases for Amazon from a bang-for-the-buck perspective, especially when you factor in bandwidth.