It's a great question

March 11, 2025

It's a great question.

There's definitely work to do to support things like UDP fully in many of the higher-level cloud offerings. People ask questions in the @pipecat_ai discord regularly about how to run voice AI agents on AWS Lambda, Google Cloud Run, and Cloudflare Workers. Which is not possible today.

On the other hand, the global network footprint and performance of all of the major clouds is excellent, so the building blocks are there.

We've spent the last ~8 years building out the infrastructure and developer tooling for this at @trydaily, so we have strong opinions about it! Global mesh routing across 70+ availability zones. We have a 13ms median first-hop latency to our multi-cloud edge. That means we can route packets to our customers' Pipecat servers and GPU clusters really, really well.

I think you need three things:
- Very good global UDP mesh routing designed for these realtime use cases.
- Developer tooling that makes it easy to use HTTP, WebRTC, WebSockets, and SIP (telephony) transports together in the same "agent container"
- Orchestration tooling that allows developers to write realtime apps and deploy them to clusters (Kubernetes or a higher-level abstraction) that handles all the infrastructure components in a standard way (autoscaling, rolling deploys that don't terminate long-running processes, etc.)

Anoop Kumar@AnoopKumarBhat

@kwindla Do you think global edge infrastructure is ready for this change including being able to deliver at lower latency and lower price?