Local hardware vs a cloud API

Whether buying hardware costs less than paying per token, using measured power draw and throughput from a published result rather than a nameplate rating.

Paying the API is cheaper over 3 years

Hardware pays for itself after about 303.1 months at this volume.

Cloud, 3y
$216
Local, 3y
$1,802.22
Of which electricity
$2.22
Difference
-$1,586.22

Local running cost uses 26.8 W and 360.9 tok/s measured on llamacpp-1787391945. Cloud pricing is yours to supply — this site bundles none, because provider prices change and a stale figure here would quietly mislead. It also ignores everything that is not electricity or purchase price: your time, cooling, failure rates, and the fact that a cloud API needs no capital up front.