My current thinking is ORMs are big hulking things that need to handle multiple use cases - more use cases than any individual project will need. So ideally, you would use raw SQL with some lightweight scaffolding around it specific to your project. But that scaffolding must be done right, and that can be a hard problem to solve that probably required significant domain knowledge. If the stakes are high, teams will avoid that risk of a poor custom solution and choose the ORM and the ongoing costs associated with it.