We have IIRC ~1k VMs there. It's rather unreliable - VMs are migrated often, so you have to expect failure. But as you have to factor failure in in any case, our company's position has been that constant failure avoids complacent design, and so we are very happy with them (and they keep our mitigation systems well fit).

Working with them is nice, the APIs are a breeze, and are really dirt cheap. And they take security very seriously.