I would bet that they have inference setup for internal use on a separate system from the customer-facing production environment. The same way telemetry infrastructure needs to be run separate from normal production systems, so you aren't "blind" when you need it most.