Tool · Run a model

Self-host vs API — at what volume does renting a GPU win?

Open-weight models can run on rented GPUs (~$1–$4/hr for an A100 / H100) or via per-token APIs. Plug in your numbers; we'll show the breakeven curve.

Monthly input tokens

30,000,000

@ 1,000,000/day

Monthly output tokens

15,000,000

@ 500,000/day

Self-host GPU

AMD AMD MI300X

needs 185 GB VRAM

Self-host monthly

$1,404

24/7 on cheapest provider

API pricing — cheapest first

Monthly bill at 30,000,000 input + 15,000,000 output tokens.

Provider	$/M in	$/M out	Monthly	vs self-host
Alibaba DashScope	$0.25	$0.75	$18.75	−$1,385.25	Open →
OpenRouter	$0.25	$0.75	$18.75	−$1,385.25	Open →
Together AI	$1.95	$8.0	$178.5	−$1,225.5	Open →

Cheapest fitting GPU for Qwen: Qwen2.5 VL 72B Instruct at FP16. Live at TensorWave at $1.95/hr → $1,404/month at 24/7 utilisation.