Pay per call, or rent the machine?

An API bill grows with every request. A rented GPU costs by the hour, plus the minutes it spends loading before it does anything. Put in your own numbers and see where the two lines cross.

API

Own GPU

How it's worked out

Both bills are monthly. The GPU bill counts the hours of real work, the hours spent starting up and loading the model each time you switch it on, and whatever you pay for storage while it's off.

api = units × price per unit
gpu = (units ÷ units per hour
       + starts × minutes per start ÷ 60)
      × price per hour + fixed

The number that matters most is units per GPU-hour. Measure it on your own model and hardware; the presets are rough examples.

What it doesn't count

Your time to set it up and keep it running.
Quality. A small open model and a frontier API are different products. Compare models in the same class.
Privacy and compliance. Keeping data on your own machine is often the real reason to self-host, and it has no price here.
Discounts, reserved pricing and price changes. Check current rates with your providers.