The inference bill nobody budgeted for

**Executive summary:** The report compares public cloud and EU private on‑premises inference for a GPT‑4‑class model, concluding a migration reduced monthly spend from $85,000 to $35,000, cut compute cost per decision (from $0.071 to $0.029), slashed latency (340ms to 22ms), eliminated CLOUD Act exposure, and satisfied EU AI Act requirements while advocating that resilience be treated as a measurable cost.