Workload
Real production utilization is usually 30–60%. A rented GPU costs the same idle or busy, so low utilization pushes cost up 2–3x versus a naive estimate.
GPU $/hr and throughput defaults are editable — real numbers drift by provider, model, quantization, and batch size.
Cost Comparison
Rough model. Actual cost depends on model size, quantization, batch size, context length, KV-cache pressure, and provider pricing. Output tokens dominate; input (prefill) is cheaper per token. Verify with a load test before committing to hardware.