GPU

A GPU is a processor designed to perform many simple calculations simultaneously. Built originally for graphics, it turned out to match the mathematics of machine learning exactly, which is why it underpins both training and inference for modern models.

For most projects a GPU is something you rent by the hour through an API rather than own, and the decision only becomes relevant when you are training your own model or running an open-weight model on your own infrastructure for reasons of cost, latency or data residency.

Owning that infrastructure is a serious commitment: hardware is expensive and scarce, utilisation is hard to keep high, and the operational burden is real. It repays at sustained high volume and rarely before.