Ask HN: What is currently the best way to host AI models?

AI infrastructure moves really quickly and best practices are constantly evolving so I was wondering what is HN's opinion on the best way to host custom AI models for inference in Q1 2024?