Self-hosted inference removes the per-token meter that cloud-managed inference exposes. The opex savings are real. The discipline cost is also real: an unmetered substrate can quietly burn power, accelerators, and operator attention at rates that erode the inflection-point advantage. This paper specifies four runtime guardrail mechanisms — auto-stop, budget alerts, runtime budget enforcement, and cost-attribution telemetry — that keep self-hosted inference on the favorable side of the inflection. Cost discipline is treated as a runtime governance surface that the agentic system must respect at execution time, in the same way it respects the governance envelope; not as a finance-team concern bolted on retrospectively. Without runtime cost enforcement, opex savings degrade into capex-by-accident.