Hacker News Reader

The parent specifically said inference, which does not include training and staffing costs.

But those aren't things you can really separate for proprietary models. Keeping inference running also requires staff, not just for the R&D.