Nvidia is an American semiconductor company that designs graphics processing units, or GPUs — the chips that do the bulk of the heavy math behind training and running artificial intelligence. It doesn’t manufacture the chips itself (that’s outsourced to foundries like TSMC), and it doesn’t build AI models either. What it supplies is the hardware and software layer nearly every AI lab, cloud provider, and startup builds on top of, which is why a single company’s product roadmap can move markets and reshape international trade policy.

From graphics cards to AI infrastructure

Nvidia was founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem to make chips that render 3D graphics for video games. Its GeForce line, launched in 1999, defined the modern GPU: a processor built to do thousands of simple calculations at once, rather than one complex calculation at a time like a CPU. That design turned out to be exactly what’s needed for AI — training a model or generating a response involves the same kind of massively repetitive arithmetic as rendering pixels, just applied to a neural network instead of a 3D scene.

The pivot from gaming to AI hardware didn’t happen by accident. In 2006, Nvidia released CUDA, a software platform that let programmers write general-purpose code for its GPUs rather than just graphics instructions. For years CUDA was a niche tool for researchers. When deep learning took off in the early 2010s, those same researchers were already writing their training code in CUDA, and Nvidia’s GPUs became the default hardware for AI research almost by inheritance.

Why the AI industry runs on Nvidia chips

Today, Nvidia’s data-center GPUs — sold under names like Hopper and Blackwell — are the chips that large AI labs use to train frontier models and that cloud providers rent out to run them. Two things keep Nvidia ahead of rivals like AMD and the custom chips that Google, Amazon, and Meta build for their own use:

  • The CUDA ecosystem. Nearly two decades of libraries, tools, and trained developers are built around CUDA. Switching to a competitor’s chip often means rewriting and re-optimizing software, a cost most AI teams would rather avoid.
  • Networking and memory, not just compute. Training modern models means linking thousands of GPUs together so they act as one machine. Nvidia’s NVLink interconnects and its GPUs’ high-bandwidth memory are built specifically for that job.

The result is a company that, by most industry estimates, supplies well over three-quarters of the GPUs used for AI training and inference worldwide — a concentration unusual for a critical piece of technology infrastructure.

The bottleneck this creates

Because so much of the AI industry depends on one supplier, Nvidia’s production capacity, pricing, and export eligibility function almost like a chokepoint for the whole field. When Nvidia’s supply is tight, AI labs delay training runs. When a government restricts which Nvidia chips can be sold to which countries, it reshapes who can build frontier AI at all — a dynamic covered in more detail in our explainers on semiconductor foundries and AI export controls. It’s also why Meta, Google, Amazon, and others are investing heavily in designing their own AI chips: not to replace Nvidia outright, but to reduce how exposed they are to a single vendor.

In the news

Meta’s push to build its own AI chip, code-named Iris, is a direct response to this dependency — our report on Meta’s in-house chip plans explains why even a company as large as Meta is trying to cut its reliance on Nvidia GPUs.

FAQ

Does Nvidia make AI models, like ChatGPT or Claude? No. Nvidia builds the chips and software that AI labs use to train and run their own models; it doesn’t compete directly in building consumer chatbots.

Is Nvidia the only company that makes AI chips? No, but it’s by far the largest. AMD sells competing GPUs, and Google, Amazon, and Meta design custom chips for their own data centers — though none currently match Nvidia’s combined share of AI training and inference hardware.

Why can’t AI labs just switch to a cheaper chip? They can, and some do for specific workloads, but most of their existing software is written and optimized for Nvidia’s CUDA platform, so switching has a real engineering cost on top of the price of the hardware itself.

Sources: Nvidia — Wikipedia; CUDA — Wikipedia; Nvidia Blackwell architecture product page (nvidia.com).