@arun more than the absolute number of devices, what is important is the number of API calls the backend needs to answer, which is strictly related to your polling interval. The Mender server scales up linearly, and you can run multiple instances of the different microservices to load-balance the traffic across multiple servers if needed. We run Hosted Mender using Kubernetes and auto-scaling, and serve several millions of API calls per hour.
If you have 100 or 3k devices, the server requirements are more or less the same because you need the basic infrastructure up and running in any case. Scaling up to 100k devices with a 10 minutes poll interval will require you approximately double the resources I mentioned above (2GB of RAM and 2 vcpus).