NVIDIA

NVIDIA makes chips, small silicon wafers that carry out calculations. Its main product used to be chips for graphics cards, the devices that make the picture in computer games look smooth. These chips are exceptionally good at one thing: doing thousands of small calculations at once, instead of one big calculation in sequence. It turned out that's exactly what artificial intelligence needs too. Training a large language model, the program behind chatbots like ChatGPT, is piles of small calculations at once. NVIDIA happened to be holding the exact right tool right when the whole world needed it.

Picture a gold rush. Hundreds of prospectors run along the streams, each sure they'll find the nugget. But none of them dig with bare hands. They all buy shovels and sifting pans from the same store at the start of the trail. It doesn't matter who finds the gold: the store profits from every prospector, no matter who gets rich. NVIDIA is exactly that kind of store, but for artificial intelligence. OpenAI (the company behind ChatGPT), Google and Meta are clawing at each other fiercely for first place, but they buy their chips from the same place.

All the prospectors are clawing to be first to reach the river. But every one of them, before driving a shovel into the dirt, paid for it at the same store. It profits, no matter who finds the gold.

Today, nine out of every ten dollars NVIDIA earns come from chips for servers built solely to train and power artificial intelligence. It's no longer just a player in this market. It is the market itself. If it stops selling tomorrow, the whole industry stops, not just one of its competitors.

Economists have a term for exactly this: supplier concentration risk, when the survival of an entire market hangs on the goodwill of a single seller. That's exactly the kind of supplier NVIDIA is for artificial intelligence. Except nobody calls it a risk. They call it leadership.

The store now sells for the desk too. On October 7, 2026 preorders opened for laptops with RTX Spark, a chip NVIDIA and Microsoft introduced back in June; sales start on October 16. By NVIDIA's numbers it combines a graphics chip and a processor with up to 128 GB of shared memory, and in Microsoft's words a laptop like that runs models that don't fit on a regular computer. The same day Microsoft made Microsoft Execution Containers (MXC) generally available on Windows 11: a closed space in which the organization decides which files and networks an AI agent can reach. The hardware is NVIDIA's, the cell for the agent is Windows'. Neither of the two pages gives a price.

But NVIDIA does not sell only chips, it also puts out models and data. On October 28, 2026 it published Nemotron 3.5 Transcribe, an open model for turning speech into text: of 30 billion parameters, about 4 billion work on each word, and it covers 40 languages. Granary v2 came out with it, the data a model like this learns from: by NVIDIA's figures 463,062 hours in 48 languages. The license is not in the source I read.

There's one detail

When I read the next story about a breakthrough from one artificial intelligence company or another, I no longer ask which one will win the race. I ask who profits, no matter which one wins. The answer is almost always the same. That's why I watch NVIDIA more closely than most of the companies the story is actually about.

The visual is generated code-art, no third-party images.