At what request volume does hosting Llama actually become cheaper than API keys?

Right now I’m spending around $400–$600/month on Claude and APIs. My traffic is super spiky heavy during daytime US hours almost dead at night. I know GPU platforms exist but cold starts sound like a UX nightmare for interactive apps. On the flip side paying for dedicated A100/H100 instances 24/7 just to handle afternoon spikes feels like burning cash while the GPU sits idle all night. For developers who made the leap at what monthly API spend did dedicated GPU hosting actually break even for you?