Cache the request that comes in that triggers a tenant's server, then replay the request when the tenant's server has booted up.
It should be possible to run a minimum of infrastructure to scale from 0