Self-hosted voice agents hit different concurrency ceilings at each layer
mahimairaja · reddit · 2026-07-22
A self-hosted LiveKit voice-agent stack does not scale as one system: each layer has its own concurrency ceiling.
- The real bottleneck is often not compute, but the lowest-capacity layer: media/SFU, worker pool, STT/TTS inference, telephony, or app-side calls.
- Adding more workers alone may not help if STT/TTS concurrency or SIP channel limits are the first thing to saturate.
- The write-up walks through LiveKit load-test sizing, an autoscaling-threshold pitfall, a 500-concurrent-call capacity table, and a rough cost-per-call-hour model.
The post is focused on self-hosted / Kubernetes deployments and asks which layer usually caps production systems first.
More from Infra
- Primis explores hedging infrastructure to make AI compute pricing predictable — dolos_diary · 2026-07-22
- PyTorch Foundation rolls out quarterly updates for six hosted projects — PyTorch · 2026-07-22
- Puri.li opens a free web search API with a 185M-page index — skillplayed · 2026-07-22
- Open-Source Coding Agent Octomind Adds Hosted Machines and 21-Model Access — donk8r · 2026-07-22
- Atome LM beats or matches TFLite Micro on 18 MCU tasks while staying 5× to 70× smaller — themoroccanship · 2026-07-22
- Compute shortage could let Amazon, Microsoft and Google lift margins — RihardJarc · 2026-07-22