Is self-hosting *actually* cheaper?

Change infrastructure, energy and test-pass assumptions. No model-quality score is pre-certified. Unlike a token-only chart, this compares money spent per verified successful task, including failed-output repair.

The hardware cost is a straight-line economic allocation, not a cash invoice. Sequential prefill+decode approximation ignores batching, overlap, concurrent requests, downtime and rate limits. Wall power and throughput are hypothetical default inputs; measure your own hardware. At zero pass rate, cost per success is undefined, not zero. This assumes one attempt per task, comparable evaluations, and only one recovery event per failed task.

NVIDIA GPU specs · Eurostat tariffs · vLLM throughput testing · UsageBox cost per task