The Arithmetic of Idle Silicon: Self-Hosting GPU vs API Cost Simulator

Workload Parameters

Traffic Archetype Preset

Monthly Request Volume 22,000

Dedicated Cloud GPU Lease

Avg Prompt Tokens / Req 350

Avg Output Tokens / Req 450

API Input Price ($/1M tokens) $1.50

API Output Price ($/1M tokens) $3.00

DevOps/Driver Maintenance (hrs/mo) 12 hrs

Engineer Hourly Burden Rate $95/hr

Silicon Idle Rate 96.8% $815.20 burned idling

Self-Host Cloud Bill $842.40 $0.0383 / req

Managed API Spend $44.80 $0.00204 / req

True Crossover Volume 412,941 Reqs/mo needed for parity

24-Hour Diurnal Utilization & Active Silicon Duty Cycle Peak: 14:00 (8.2% daily load)

00:00 (Night) 06:00 (Morning ramp) 12:00 (Midday) 18:00 (Evening) 23:00 (Off-peak)

Cost Trajectory vs Monthly Volume (Self-Hosted GPU vs API) Emerald = Managed API | Purple = Dedicated GPU

Architecture Breakdown & Hidden Maintenance Tally

Component	Dedicated Self-Host Cloud	Managed API Routing	Delta / Penalty
Direct Compute / Token Invoices	$842.40 / mo	$44.80 / mo	+$797.60/mo
DevOps & CUDA Driver Triage (Labor)	$1,140.00 / mo	$0.00 / mo	+$1,140.00/mo
Fully Burdened Monthly Cost	$1,982.40	$44.80	+$1,937.60
Effective Cost Per 1,000 Inference Calls	$90.11	$2.04	44.2x higher on GPU

The Self-Hosting Delusion Verified: With 22,000 monthly queries, your dedicated GPU spends 96.8% of each hour burning electricity and cloud lease fees with completely zero active tensor execution. The author paid an $797.60 hardware surcharge plus 12 hours of CUDA/driver triage to save a $44.80 API invoice.