Self-Hosted LiveKit – Voice Agent Concurrency Benchmarks and Infrastructure Sizing

Self Hosting
1

Hi everyone,

We’re exploring self-hosting LiveKit instead of using LiveKit Cloud and would like to understand its real-world scalability.

Has anyone performed concurrency or load testing on a self-hosted LiveKit deployment, particularly for AI voice agents?

I’d love to know:

  • How many concurrent voice sessions can a server reliably handle?

  • What CPU, RAM, and network specifications are you using?

  • Are you using a single instance or a distributed setup?

  • At what point did you encounter performance bottlenecks?

Any real-world benchmark numbers or production experiences would be helpful in estimating the infrastructure requirements for our deployment.

Thanks in advance!

2

Its hard to guess without all the information. As the main things that take resources are LiveKit, Egress, and Agent Worker services.

So it depends on which ones you are planning to use, as if you plan to record or use agents or not. Plus depends if you want voice or video.

As a guide, LiveKit suggests 4 vCPUs and 8GB RAM as a starting point. You should get around 10-16 simultaneous calls for voice with egress off.
With all on including egress, expect 100% CPU on around 8-10 sim calls.

At this point though you would need to optimise things.

You then increase the server specs but then you start adding multi-cluster or more servers to the pool https://www.youtube.com/watch?v=AW2jJ_L93tM