Does it matter that each token takes additional milliseconds on the network if the local inference isn't fast? I don't think it does.

The privacy argument makes some sense, if there's no telemetry leaking data.