GPU compute, explained for investors

A GPU — graphics processing unit — is a chip built to do thousands of simple calculations at once. That's exactly the shape of work AI models need: training and running a model means multiplying enormous grids of numbers, millions of times per second. CPUs do a few things fast; GPUs do thousands of things at once. That's the whole trick.

Why compute became a commodity

An AI company doesn't fundamentally care which GPUs it uses — it cares about getting enough of them, reliably, at the best price per hour. That makes GPU time behave like a commodity: it's priced per GPU-hour, traded under contracts, and scarce when demand spikes. When you hear "compute," think rentable machine time, the way "power" means rentable electricity.

The three layers of the market

Chipmakers (NVIDIA and rivals) sell the hardware — the shovels. Data centers buy the shovels, house them, power them, cool them, and rent out their time. AI companies rent that time to build and run models. Each layer has different economics: chipmakers earn on scarcity of supply, data centers earn on utilization of their fleet, AI companies earn on what they build with it.

Infrastructure investing, Tervolt's lane, is the middle layer: you help finance the fleet, and the rental income — minus real costs — is where returns come from.

What to remember

• One modern AI server contains several GPUs and costs as much as a house.
• Its product is GPU-hours; its income depends on how many of those hours are actually rented (utilization).
• GPUs age fast — new generations arrive every 18–24 months and older chips rent for less. This depreciation is the biggest cost in the business.

If you understand those three bullets, you understand more about the AI economy than most headlines will ever tell you.

Educational content — not investment advice. Capital at risk.