My recent project involved AWS with complicated autoscaling and database multimaster clustering which handled smoothly the ramp up of simultaneous 30 users to 3000 users in 2 hours. The portals pattern was pre known, and the layout could be done depending on the same. Even with all those, the client ended up paying about one third of what they estimated using the caclculator. The application was heavily using the SES for mailing, and at one point we had about 400K mails per 24 hour period.