We actually worked out the math - we don't use any external APIs like OpenAI (use our own models) so those costs aren't prohibitive, and we have a bunch of experience as a team building scalable ML systems, so we have driven down the serving costs by a lot.