The only thing which seems to inflate more drastically than hashpower are hashpower expectations. It seems that people look at growth slopes and then extrapolate half a year into the future. Beware that slope changes can have dramatic effect on timing, it's easy to get the expected hashpower wrong by a factor of 2-5 over a period of 6 months just based on slope analysis.

Certainly there will be a need to bring petahashes to the network. That's why one of the most important aspect of suppliers is to optimize production and deployment scalability. Failure at that stage can lead to loosing competitiveness either from a fiscal and/or a reputation point. Just having a fancy chip won't cut it.