This is something I’m busy struggling with: Performance of a Production bot
It seems very dependant on your exact configuration (which tracker store you use), and how complex your bot is (how many NLU examples, how many stories).
In my case, I’m getting around 10req/s per CPU. So for 500 concurrent users, I would estimate that I need around 25 instances, so at least 25 CPUs total in my worker machines, plus somewhere to run a postgresql, redis, and the actions server.
For the actions server, it depends on what your actions server is doing, but in my case, usage was very low.
The way to know for sure is to benchmark your specific bot, with the same settings and services that you will be using in production.