Ask HN: Looking for empirical advices on "serverless" ML hosting

I'm exploring serverless platforms for hosting custom machine learning models and could use some advice from those who have experience with these services. I get that "serverless" still involves servers, but I'm looking at options that market themselves this way.

I'm considering:

modal.com runpod.io replicate.com beam.cloud

I'm curious about two main things: Pricing: How do the costs compare for running custom ML models on these platforms? Any detailed breakdowns would be great. Cold Start Time: How quickly can each platform start up a custom model from a cold start? Are there big differences in startup times?

If you've used any of these, I'd love to hear your thoughts and experiences.