tbh, that's a good point about the voice agents that I hadn't considered. I guess there are some latency-sensitive inference workloads. Thanks for pointing that out.