I was shat on multiple times by throbbing-brain architecture guys for this, but what I've done multiple times that worked in prod, was to start with a single monolith with dependency injection, and run all services in the same process.

Once the compute and robustness demands mandated we should scale the thing across multiple machines, both horizontally and vertically, I would replace most DI services with proxies that wrapped the original functionality and called out to a remote host.

So for this scaled up version, I would only need to change how the DI container got started, which could either be in 'Manager' mode (meaning a high level functionality, with the submodules injected as proxies calling out to services on client machines), 'Worker' mode (meaning it served the proxy requests on said worker machines) or 'Standalone' mode (meaning the DI container actually injected the actual fat versions of the services, that allowed the whole thing to run in a single process, very useful for local testing and debugging).

The was only a single executable which could be run in multiple 'modes' selected via a command line switch, which made versioning and deployment trivial.

In what programming language did you do this? Sounds like a nice mostly future-proof architecture.