Sunk Cost - how many years until local AI hardware pays for itself

The case for running models locally is that tokens stop costing money once the machine is bought. This calculator puts a number on the word "once".

- the default scenario is unkind: a $3,499 Mac Studio running Qwen3 27B at a moderate coding-assistant load saves $0.22 a day, which works out to 43.6 years before break-even
- every assumption is editable - electricity price, tokens per day, context window, generation speed, or simply the monthly API bill being replaced
- the key-value cache math is shown alongside, so a context window that feels free turns out to cost gigabytes of memory and generation speed

The value here is the reframe rather than the verdict: it forces anyone arguing for local inference to state the workload they are assuming out loud, and 99 comments against 46 points on Show HN suggests that argument needed having.

https://sunkcost.ai/

📎 Read also:
→ Computable GPU Index - an open reference price for one GPU-hour
→ Larridin priced the AI-native engineer - $920 a month in tokens
→ PwC - where the $31.6tn AI buildout leaves compute prices