That's reasonable. Also consider going the other way: keeping per-tenant logical databases, and splitting up some or all of the compute layer to have single tenancy or bounded tenancy. For example, if your compute layer is a web server, making multiple sets of webservers with something in front of them routing requests to a given set of servers based on a tenant identifier can chunk up your multiple-noisy-neighbors problem into at least multiple noisy "neighborhoods", with the (expensive) extreme of server-per-tenant. If your compute layer is e.g. a service bus/queue/whatnot worker, the same principles apply: multiple sets of workers deciding what to work on based on a tenant ID or per-tenant/group topics or queues. You can put the cross-cutting/weird workloads onto their own areas of hardware, as well.
I propose this because I think having database instances split up by tenant (even if multiple DBs share the same physical server) is actually a pretty good place to be, especially if you can shuffle per-tenant databases around onto new hardware and play "tetris" with the noisiest tenants' DBs. Moving back to multitenant-everything seems like a regression, and using (message|web|request) routing to break the compute layer up into per-tenant or per-domain clusters of hardware can often unlock some of the main benefits of microservices without a massive engineering effort.