Dynamically scaling load is only really interesting if the load is extremely variable. You could just buy 5x the dedicated hardware, which will likely be able to serve 10x the traffic and may still end up paying less. Who cares if it sits idle most of the time.
Now Spotify, like Netflix probably has extreme variance in load, much more so than a normal company. So maybe the cloud works for them. But that's not a common situation at all.
Our customers like it because their existing ops staff can greatly reduce their workload and allow development staff to move much faster.
The thing with them is if they go down you have Google to worry just as much as you (if not more) for bringing them back up.
You don't have that with your own servers.
Googlers might be more skilled than your ops team, yeah, but I would never bet on huge companies to worry about the fate of something that is not their main revenue generator.
Dedicated servers create a different kind of lock-in: Dynamically scaling up dedicated servers is like laying down the tracks while the train is moving. Acceptable to some, but not to a company like Spotify.