GPU

A CPU is designed to do one thing after another, very fast, and to handle whatever that thing turns out to be. A GPU is designed to do the same simple operation on thousands of numbers simultaneously. It is worse at almost everything a CPU is good at, and vastly better at the narrow case.

That narrow case happens to be exactly what neural networks are made of. Training a model is, underneath the terminology, an enormous number of multiplications and additions arranged in grids — the same operation repeated across millions of values with no dependency between them. That is the one shape of problem a GPU was built for, and it had been built for it years earlier, for drawing 3D graphics.

Why this became a strategic issue

The coincidence turned a components market into a geopolitical one. When the hardware that trains frontier models comes from essentially one company, and the chips are manufactured in one region, the supply chain becomes a lever — which is why export controls on chips have become foreign policy rather than trade policy.

It also shapes the economics of AI companies more than software costs do. Training a large model is a capital expense measured in tens of thousands of GPU-months. Running it afterwards — inference — is a smaller cost per request but never stops, and at scale that is the number that decides whether a product makes money.

Terms you will meet nearby

VRAM is the memory attached to the GPU. It is usually the binding limit for running a model locally: a model that does not fit in VRAM either runs very slowly or does not run at all.

TPU is Google’s own chip, built for the same job. NPU is the small accelerator now appearing in laptops and phones for running modest models on the device.

The practical upshot for anyone running AI locally is simple: the amount of memory on the card matters more than the speed of the card, and that is the opposite of what the marketing emphasises.