Why data-center AI chips became the hottest product in tech

A few years ago, graphics processors were mostly about gaming. Today the most expensive ones live in data centers, training large language models.

Why GPUs

Training a neural network means billions of similar calculations. Graphics accelerators run thousands of them in parallel, so they handle the job far faster than regular processors.

How many are needed

To train a large model, companies link thousands of accelerators into clusters connected by a fast network. That’s a huge investment — and the reason chipmakers have become some of the world’s most valuable companies.

The flip side

• Energy. Large data centers use as much electricity as a small town.
• Cooling. Liquid cooling is becoming common.
• Shortages. Demand still outstrips supply, so chips are ordered months in advance.

What’s next

Competitors are building their own specialised chips, and so are the big cloud companies. The market is likely to become more diverse in the coming years.