Nvidia’s been in the headlines all weekend for reasons that have nothing to do with AI. Its CEO, Jensen Huang, admitted to an affair, the board opened an investigation into reports that China targeted him, and the stock dropped. Whatever comes of that, it raises a fair question. What actually keeps Nvidia on top?
The short answer isn’t a chip. It’s four letters most people have never looked up.
What it means
CUDA is Nvidia’s free software toolkit for programming its GPU chips. Nvidia released it in 2006. Before that, graphics chips mostly drew video game graphics. CUDA let regular programmers use them for any math that can be split into thousands of small pieces and done at once, which turns out to be exactly what training and running AI models needs. A regular processor works through a to-do list one item at a time, really fast. A GPU has thousands of smaller workers that each grab an item at the same time.
The catch is that CUDA only runs on Nvidia hardware. So every program, tool and shortcut written for it quietly assumes you bought Nvidia.
Why it matters
Twenty years of head start is hard to copy. There are hundreds of CUDA libraries (ready-made building blocks for common jobs), and millions of developers learned on it. PyTorch, the tool a huge share of AI models get built with, was built to run on CUDA first. A rival chip can be just as fast and still lose, because nobody wants to rewrite their software to use it.
Competitors keep trying. AMD has its own version called ROCm, and it hasn’t caught on with developers the way CUDA has. In China, where the best Nvidia chips are banned, DeepSeek and Huawei built a free alternative called TileLang for Huawei’s chips. Analysts are, honestly, skeptical it pulls many people away.
It’s the reason one company matters so much. When people say Nvidia has a “moat,” this is mostly what they mean. It’s also why export rules, chip smuggling cases and, yes, CEO drama all land so hard. Nearly the whole industry is built on one company’s software.
Simple example
Think about your iPhone. Maybe an Android phone comes out tomorrow with a better camera and a cheaper price. Switching still means rebuying apps, moving your photos, relearning everything and losing your iMessage group chat. Most people just buy another iPhone.
That’s CUDA for AI companies. The chip is the phone. CUDA is all the stuff on it, and moving all that stuff is the expensive part.

