The people who need Amharic transcription most are not holding a datacentre budget. They are holding a mid-range Android phone on metered data, which is why the subtitle generator transcribes locally and only sends the render job out.
What the constraint cost
Accuracy, a little. A larger model does better on hard audio, and the hosted API exists for exactly that. But a model that runs where the user already is removes an upload, a queue and a bill, and for most clips the smaller model is already good enough to edit rather than retype.
It also changes who can build on us. An open model at this size can be embedded in an application that has no server at all, which is a different market from the one an API serves.