Lesson 6 of 7

Cost: hardware vs API

6 min read

Local models cost nothing per message — so is running your own always cheaper than paying an API?

Two ways the money works

The two options have opposite cost shapes. An [API](glossary://api) is pay-as-you-go: near zero to start, rising with every call — no upfront cost, but it never stops. Local is the reverse: a real upfront cost (a capable computer, plus the electricity to run it) and then almost nothing per use. Plotted against how much you use AI, those two lines cross.

API cost rises with every call; local is a big upfront cost then near-free per use — the lines cross.

Find your break-even

For light or occasional use, the API wins easily — you might spend pennies, while local's hardware never pays for itself. Local only comes out cheaper past a high, steady volume of use — and even then, people often choose it for privacy and offline more than for the money. Don't guess: estimate your real usage and see which side of the crossover you're on.

Local pays off only past a high, steady volume — below it, a pay-as-you-go API is cheaper.

Count the hardware honestly. A machine that can run big models comfortably isn't cheap, and it draws power — fold both into the comparison before calling local 'free'.