For example, Amazon could estimate 99.95% downtime, because of physical and geographical redundancy, etc. But this analysis would be faulty, as their outage earlier this year showed.
There are a litany of long-tail black swan events that could bring down entire datacenters, that people just can't anticipate. Not even including earthquakes, terrorist attacks, etc, but even simple upgrades or misconfigurations like the one that took down their datacenter in the East Coast. Yet, they still advertise a SLA of 99.95% availability, etc. Is the risk to downtime really only 0.05%? Was the event that occurred really a 3 standard deviation event? I highly doubt it.
This complete lack of true ability to estimate risk means that customers also have essentially an inaccurate view on what their risks are. Like the one commenter who said that a small business ran their POS device over the cloud, if you told them that they would be down 2 days out of the year, would they really be interested in that? Probably not.
In a similar vein, the authors were likely promised great uptime, but no guarantees on I/O or CPU performance, which is something you don't think of. The cloud provider doesn't have to be down for your web service to be drastically affected. I suppose since this is all new, the customers are learning which questions to ask, and the cloud providers are learning which things to guarantee, so hopefully this is worked out in the next year or so.