Why AI Runs on GPUs

Why graphics chips became the engine of artificial intelligence, how Nvidia came to dominate, and the race to build alternatives.

Modern AI depends on performing enormous numbers of simple calculations, especially multiplying large tables of numbers. Graphics processing units, or GPUs, originally designed for video games, turned out to be ideal for this.

Why GPUs

Traditional computer processors, CPUs, handle tasks one after another very quickly. GPUs contain thousands of smaller cores that perform many calculations at the same time, called parallel processing. This suits both rendering game graphics and training AI models.

Nvidia’s dominance

Nvidia, an American company, became the dominant supplier of AI chips. Its advantages include:

Powerful chips designed for AI.
Software: its CUDA platform, launched in 2006, lets developers program GPUs easily. Years of tools and developer familiarity created a strong ecosystem that competitors struggle to match.
Networking technology to link many chips.

As demand for AI exploded after 2022, Nvidia’s revenue and profits soared. It became one of the world’s most valuable companies, at times the most valuable, with a market value above 4 trillion dollars in 2025.

Supply constraints

Nvidia designs its chips but relies on TSMC in Taiwan to manufacture the most advanced ones. Shortages of chips and advanced packaging capacity limited supply during the AI boom, raising prices and making access to chips a strategic advantage.

Competition

Competitors are racing to build alternatives:

AMD and Intel make competing chips.
Large cloud companies design their own chips, such as Google’s TPUs and Amazon’s Trainium.
Chinese companies, facing U.S. export controls on advanced chips, are developing domestic alternatives.

Export controls

Since 2022, the United States has restricted exports of the most advanced AI chips and chipmaking equipment to China, citing national security. These controls have shaped where AI development happens and pushed China to invest in its own chips.

The software moat

A company considering switching from Nvidia to a cheaper chip finds that its AI software, and its engineers' skills, are built around Nvidia's CUDA tools. Switching would require rewriting code and retraining staff. This cost of switching, not just chip performance, protects Nvidia's position, a pattern economists call a moat.

Thinking Nvidia manufactures its own chips

Nvidia designs chips but relies on TSMC to manufacture the most advanced ones. This shows how specialised and interdependent the chip supply chain is.

Key takeaways

GPUs' parallel processing makes them ideal for AI.
Nvidia dominates AI chips thanks to powerful hardware and its CUDA software ecosystem.
Supply depends on TSMC's manufacturing, creating bottlenecks.
Competitors and U.S. export controls are reshaping the AI chip market.