If your usual working set (code+data), or better still your entire working set, fits into RAM easily and doesn't see a lot of write activity (where in RAM or not it needs to be copied to disk at some point), cheap VMs are often a no-brainer. The I/O hit will only seriously affect you if you need to reset he VM so the working set needs priming in RAM again.
As soon as you need I/O bandwidth (much writing, or a data-set that doesn't fit easily in RAM) you need a physical machine as the I/O contention between you and other VMs will completely kill performance.
So as you say: real machines for significant database work and cheap VMs for your web servers and other worker tasks is often the most cost effective arrangement. Don't go too cheap on the VMs though for production work. I generally avoid "burstable RAM" options as I find them less reliable (I'd rather have a fixed amount that can all reliably be allocated as needed, than unpredictable reliability forcing me to put procedures in place to deal with unexpected OOM situations). And make sure you properly test the bandwidth carrying capacity: I've seen cheap hosts get good reviews where the reviewer tested incoming bandwidth but filed to test outgoing - this is not all that useful as a hosts outgoing bandwidth usually sees much more contention than the incoming.
I'm now paying $405 / month for 6 servers, and they range from 2 dedicated db machines with 12GB and 16GB of memory to 1GB virtualized app and task servers that cost $10 to $20 per month.
What I've learned is that where it really matters, in terms of I/O performance, the dedicated servers make great machines, naturally. But the insanely cheap VPSs, which I use for distributed feed fetchers (task servers) and web servers, are wonderful because they are so inexpensive. I have redundant load-balanced machines, so when I need to take down an app server or it fails for some reason, nginx immediately routes around it. And it's costing me only $10 / month per machine.
And task servers have served me especially well as VPS. It depends on the intensity of the task, but for such a small expense, you get 4 cores and enough memory to go nuts. And if you're pulling off a task queue, you can programmatically spin up more to deal with a higher queue load. When you move entirely to dedicated machines, your spin up time is on the order of hours to get it prepped and connected.
I'm just thrilled that many decent hosting providers (with <table>s for salespeople) offer both. I wouldn't last without being able to have a mixed setup.