Why GPUs cost so much

A graphics processor was built to draw pixels: millions of simple sums, all at once. Training a neural network turns out to need exactly that kind of work, mostly multiplying large grids of numbers. That is why a chip designed for games became the most sought-after part in computing.

Four reasons the price stays high

1. Very few factories can make them

The most advanced AI chips are made on the newest manufacturing processes, and almost all of that capacity sits with one contract manufacturer, TSMC, in Taiwan. Building a new leading-edge factory takes years and tens of billions of dollars, so supply cannot jump when demand does.

2. Packaging is its own bottleneck

An AI accelerator is not one chip. The processor sits next to stacks of high-bandwidth memory (HBM) on a shared base, joined by thousands of tiny connections. This "advanced packaging" step has repeatedly been the tightest part of the supply chain, and HBM itself comes from only a handful of memory makers.

3. The software keeps buyers in place

Nvidia's CUDA platform has been the default way to program these chips for more than fifteen years. Most AI frameworks, libraries and tutorials assume it. Rivals make capable hardware, but moving a large training setup to new software costs engineers' time and adds risk, so many buyers pay more to stay.

4. Everyone wants them at the same time

AI labs, cloud providers, governments and large companies are all building capacity at once. When demand is this concentrated, sellers set prices, and buyers sign multi-year orders to secure supply.

What it means

What could change it

• More factories and packaging lines coming online.
• Custom chips from cloud providers designed for their own workloads.
• Better software for other chips, which lowers the cost of switching.
• Smaller, more efficient models that need fewer chips for the same job.