Nvidia, AMD and the AI boom
Almost every AI picture you have made ran on a graphics chip. Usually Nvidia's.
How a company that made cards for video games ended up underneath ChatGPT, every image generator and most AI video, why AMD and a crowd of others are chasing it, and whether the money behind it all is a bubble.
Why a graphics chip?
A video game redraws millions of pixels sixty times a second. Each pixel is a small, simple sum, and none of them waits for the others, so graphics chips were built to do thousands of small sums at the same time instead of a few big ones in a row.
That turned out to be exactly the shape of modern AI. An image model starts from noise and cleans it up over dozens of steps, and every step is billions of multiplications that can all happen at once. A regular processor works through them a handful at a time. A GPU does them in a flood. That is the whole trick, and it is why "AI chip" and "GPU" get used as if they meant the same thing.
The real lock is the software
Plenty of companies can build fast chips. What Nvidia has that they don't is CUDA, the programming kit it released in 2007, years before anyone was calling it AI. Researchers wrote their experiments for it because it was there and it worked. The tools that came out of those labs, the ones nearly every AI model is trained with today, were tuned for Nvidia first.
So switching away is not just buying a different card. It means making sure years of code runs just as well somewhere else. That is the moat people mean when they talk about Nvidia, and it is why rivals with good hardware still struggle to take customers.
From Nvidia's yearly results. Its fiscal 2026 ended in late January 2026, with $193.7 billion of its $215.9 billion coming from data centers.
Four years ago Nvidia was a big company. Now it is the company. In October 2025 it became the first business ever worth five trillion dollars on the stock market, and it has stayed at or near the top of the market since. All of that rests on one fact: the people building AI keep buying more chips than Nvidia can make.
AMD, and everyone else who wants a slice
AMD is the closest thing to a direct rival. Its Instinct accelerators, the MI300X in 2023, the MI350 in 2025 and the MI400 family with its Helios racks in 2026, go head to head with Nvidia's data center chips, and its ROCm software is its answer to CUDA. AMD's AI business is growing fast. It is also still a fraction of Nvidia's, for the software reasons above.
The bigger threat may be Nvidia's own customers. Google runs much of its AI on its own TPU chips. Amazon has Trainium, Microsoft has Maia and Meta has MTIA. Broadcom designs custom chips for several of them. Intel sells Gaudi accelerators. Each of them would rather not hand Nvidia most of its AI budget forever, and every chip they build themselves is one Nvidia doesn't sell.
On your desk, the picture is simpler. Running image models at home mostly means an Nvidia GeForce card, an AMD Radeon card that takes more setup, or an Apple Mac whose built-in chip shares memory with the graphics.
So is it a bubble?
Here is the honest version of the argument, both sides.
Why people say yes
The spending is far ahead of the earning. Tech giants are pouring hundreds of billions a year into chips and data centers, and the AI products on top of them don't bring in anything close to that yet. Some of the money also goes in circles: in September 2025 Nvidia announced plans to invest up to $100 billion in OpenAI, a company that then buys Nvidia chips. And the market is jumpy. In February 2026, within two days of both companies posting strong results, Nvidia and AMD together lost about $711 billion in stock market value.
Why people say no
The dot-com bubble was full of companies with no profits. This one isn't. Nvidia makes enormous profits on every chip, and the companies buying them, Microsoft, Google, Amazon and Meta, are some of the most profitable businesses that have ever existed, spending money they actually have. People are also really using the stuff, every day, for work and for fun.
Both can be true at once. The railways and the early internet were real, and their booms still ended in crashes that wiped out a lot of investors. Something being useful doesn't stop people from paying too much for it.
And the AI takeover?
The scary version, machines deciding they don't need us, makes good films. The version happening now is quieter and more ordinary. AI is being bolted onto every app you already use. It is changing what some jobs look like, especially in writing, design, support and code. And the data centers behind it need so much electricity that power companies and governments now plan around them.
If there is a takeover, it is the chips taking over the budgets. Whoever makes the hardware, and whoever can afford the most of it, decides a lot about what AI gets built next.
What it means if you just want to make pictures
You don't need any of this hardware. To run models on your own computer, a graphics card with a lot of memory matters more than anything else, and our GPU guide for AI art goes through which cards make sense. But most people would rather skip the card, the drivers and the setup.
That is what Wurt.app is for. You get the results right in your browser: pictures up to 4K in Picwurt, videos up to 30 seconds with sound in Vidwurt, and one-tap animation with Quick Animate, on a phone or an old laptop. It is free to start.
How an AI picture gets made
Tap a layer to see what it does.
You type a sentence: a fox in the snow, wide shot. That is the only part you see.
Words you will keep hearing
- GPU
- Graphics processing unit. A chip that does thousands of small sums at once.
- CUDA
- Nvidia's toolkit for programming its GPUs. Most AI software is written for it.
- ROCm
- AMD's answer to CUDA.
- VRAM
- The memory on a graphics card. Bigger models need more of it.
- HBM
- High bandwidth memory, the very fast memory stacked next to data center AI chips.
- Blackwell, Rubin
- Nvidia's data center chip generations: Blackwell now, Rubin next.
- Training
- Teaching a model, once, on huge amounts of data. The most expensive part.
- Inference
- Using a trained model, like every time you make a picture.
- TPU
- Tensor processing unit, Google's own AI chip.
Questions
Why does AI run on Nvidia GPUs?
AI models are mostly huge piles of multiplication, and a graphics chip can do thousands of those at once. Nvidia also built CUDA, the software most AI research was built on, so its chips became the default.
Is AMD a real rival to Nvidia in AI?
Yes, and a growing one. AMD sells Instinct accelerators for data centers and has its own software, ROCm. It is still far smaller in AI revenue, and most AI code is written with Nvidia in mind first.
Is AI a bubble?
Nobody knows yet. Companies are spending far more on AI chips and data centers than AI earns today, which looks like a bubble. But Nvidia and the biggest buyers make real profits, which the dot-com bubble mostly did not.
Do I need an Nvidia graphics card to make AI pictures?
No. To run models on your own computer, a card with plenty of memory helps a lot. To just make pictures and videos, a browser app like Wurt.app works on any phone or laptop, with no graphics card needed.
Is this investment advice?
No. This page explains how the chips and the AI business fit together. It does not tell you to buy or sell anything.