The whole argument comes down to bursty vs. non-bursty workloads. What type of workloads make up the fat part of the distribution? If most use cases are bursty (which I would argue they are) then the author's argument only applies for specific applications. Therefore, most people do indeed see cost benefits from the cloud.