Thanks — this is a great question and I can hear your goal: keep operational cost as low as possible while getting reliable, low-latency behavior for a voice agent. Quick answer up front • For the lowest operational overhead and predictable low latency: use LiveKit Cloud (Build plan gives 1,000 free agent minutes/month; agent sessions are billed $0.01/minute; inference billed separately). This removes most infra work and routing/scale problems. • If you must self-host to reduce recurring inference or hosting bills, prefer a simple always-on container platform (fly.io or Render) or a small dedicated VPS (non-burstable CPU) over serverless Modal — Modal’s cold-starts / dispatch model can produce long (~20s) latency for real-time websockets and is not recommended for voice agents. What I’m responding to • Your Modal compute experience (needing min_container=1, and ~20s latency when doing manual dispatch). • Your list of candidate hosts (Hetzner, AWS ECS, fly.io, Render) and requirement to minimize running cost for a voice agent. Why Modal felt slow • Modal’s serverless nature and dispatch/cold-start behavior means you either keep containers always warm (min_containers=1) or you face reconnect/dispatch delays. That causes the long latency you hit and makes Modal a poor fit for websocket-based, low-latency voice agents. Options and trade-offs (short) • LiveKit Cloud (recommended if you want minimal ops) ◦ Pros: Managed scaling, global routing, predictable sub-second latency, minimal ops. ◦ Cons: Session + inference usage fees (but Build plan gives free minutes). ◦ Good when you prefer simplicity and predictable performance. • fly.io ◦ Pros: Global edge placement, good for low-latency websocket apps; easy to run always-on containers. ◦ Cons: You pay for always-on instances; cost higher than raw VPS but lower ops than full cloud infra. ◦ Good if you want low-latency self-hosting with moderate ops. • Render ◦ Pros: Simple always-on services, built-in CI and HTTPS, easier than ECS. ◦ Cons: Pricing higher than basic VPS, but simpler to run reliably. ◦ Good for small teams who want reliability without deep infra work. • Hetzner VPS (or similar low-cost VPS) ◦ Pros: Very low monthly cost (small VPS), good if budget is primary. ◦ Cons: You must manage scaling, monitoring, and restart logic. Avoid burstable instances — use dedicated CPU, otherwise CPU throttling can cause high latency (especially for VAD/turn-detection/inference work). ◦ Good for hobby/low-scale deployments where you can accept more ops work. • AWS ECS / EC2 ◦ Pros: Full control, enterprise features. ◦ Cons: Higher cost and much more operational complexity to get reliable websocket affinity, autoscaling, and non-throttled CPU. ◦ Only choose this if you need enterprise integrations/compliance. Concrete suggestions to minimize cost while keeping reasonable performance 1. If you want to move fast with minimal ops: start with LiveKit Cloud Build plan to validate the agent and user flows. You’ll avoid the routing/dispatch headaches and get sub-second latency for most cases. 2. If you want to self-host after validation: ◦ Deploy the agent as an always-on container (do not rely on serverless cold-starting). ◦ Prefer platforms that let you guarantee an instance is always running (fly.io, Render) or use a low-cost dedicated VPS with a non-burstable CPU (Hetzner dedicated). ◦ Run multiple agent workers for redundancy and to handle concurrent sessions. 1. Minimize inference cost: ◦ Use smaller models where acceptable (e.g., lighter LLMs and STT/TTS). ◦ Cache repeated responses where you can. 1. Test latency end-to-end early: ◦ Measure websocket connection time, STT/LLM/TTS turn times, and end-to-end response time. ◦ Watch for CPU throttling and cold-starts — either will spike latency. Questions to help narrow a final recommendation • Do you expect steady always-on traffic, or mostly occasional sessions? (If mostly occasional, LiveKit Cloud is especially attractive.) • Do you need data residency or full control of data (that would push you toward self-hosting)? • What inference providers/models do you plan to use (payg cloud models vs local/on-prem models)? That affects both latency and cost. If you share those answers I’ll give a tailored recommendation (specific instance sizes / fly.io plan / Render settings / basic autoscaling rules) to match your budget and latency targets. Sources: Deploy and scale agents on LiveKit Cloud | LiveKit | LiveKit’s Series B: Building the all-in-one platform for voice AI agents | ... | Introducing LiveKit Inference: A unified model interface for voice AI | Liv... | Introducing LiveKit Inference: A unified model interface for voice AI | Liv... | Agents on LiveKit Cloud | Deploy voice AI on global infrastructure | LiveKi..., with interactive elements