Oh here's the thing though: the company was already using RabbitMQ, as well as Kafka (for other things, mostly data science), and it was all deployed in Kubernetes and replicated across three data centers, including one in Amsterdam. The complexity was already there, but only for a very restricted set of applications.

All of these problems sound like they come from developers who don't know how to develop distributed systems. If anything this should be simplifying an otherwise intractible problem. I have seen people overuse messaging and had to fix what I could. I otherwise reach for RMQ as a simple solution to horizontal scaling and I write the software to easily switch brokers (Aws SQS, on prem RMQ, cloud RMQ). It's not something that you choose to adopt and then force developers to make every interaction a distributed message.