Why Everyone's Obsessed with NVIDIA (And Why You Should Care)

You've probably noticed NVIDIA in the news. A lot. The company that used to make graphics cards for video games is now worth more than most countries' GDP. Their CEO, Jensen Huang, shows up at conferences in a leather jacket like a rock star. And world governments are treating their chips like strategic weapons.

What's going on?

Here's the short version: NVIDIA makes the chips that power AI. And right now, they're the only ones who can make them well enough, fast enough, at scale.

Let me unpack that.

If you read last week's issue on how large language models work, you'll remember the key insight: AI doesn't think step-by-step like a calculator. It does billions of calculations all at once, in parallel.

Regular computer chips—the CPUs in your laptop—aren't built for that. They're designed to do one thing at a time, very fast. Like reading a book word by word.

This is where NVIDIA comes in.

NVIDIA sells GPUs—Graphics Processing Units. Originally, these were designed to render video game graphics, which means calculating the color of millions of pixels simultaneously. Turns out, that same ability—doing tons of simple math all at once—is exactly what AI needs.

But here's something most people don't realize: NVIDIA doesn't actually build the chips.

Think of NVIDIA like an architect. They draw up the blueprints for incredibly sophisticated chips. Then they hand those blueprints to TSMC (Taiwan Semiconductor Manufacturing Company), which runs the world's most advanced chip factories and actually builds the physical chips. Those chips then get bought by companies like OpenAI, Google, and Anthropic to train and run AI models like ChatGPT, Gemini, and Claude.

So the flow is: NVIDIA designs → TSMC manufactures → AI companies buy → you get chatbots.

Here's the wild part: each step in that chain is essentially a monopoly. NVIDIA dominates chip design. TSMC dominates manufacturing. And TSMC depends on machines made by a single Dutch company called ASML—the only company on earth that can build the equipment needed to make cutting-edge chips. Three bottlenecks, three companies, entire AI industry.

They're trying. AMD, Intel, Google, Amazon, and dozens of startups are all racing to compete. Some of their chips are actually pretty good on paper.

But here's NVIDIA's real advantage: it's not the hardware. It's the software.

In 2006, NVIDIA released something called CUDA—a programming platform that lets developers write code for their chips. Over nearly two decades, millions of developers have learned CUDA. Universities teach it. Every major AI framework is optimized for it. The entire ecosystem of AI tools assumes you're using NVIDIA.

Think of it like this: imagine trying to compete with the iPhone, except everyone in the world has already spent 18 years building apps exclusively for iOS, and switching to your new phone would require rewriting all of them from scratch.

That's NVIDIA's position.

The numbers are staggering. NVIDIA controls somewhere between 80% and 92% of the AI chip market, depending on how you count. Their flagship chips sell for $30,000+ each. Their gross profit margin is around 75%—unheard of the traditional manufacturing industry.

Here's another way to think about it: Google, Microsoft, Amazon, and Meta are all trying to reduce their dependence on NVIDIA. These are some of the richest, most technically sophisticated companies in history. They've spent billions developing alternatives. And they still can't stop buying NVIDIA chips. Google's official statement captures the tension: "We are experiencing accelerating demand for both our custom TPUs and NVIDIA GPUs."

This is where it gets interesting—and a bit scary.

The U.S. government has decided that advanced AI chips are too strategically important to let China have them freely. Starting in 2022, the U.S. implemented export controls that essentially ban the sale of top-tier NVIDIA chips to China.

The logic: whoever controls the chips controls the AI—and whoever controls the AI holds a massive strategic advantage.

China, predictably, isn't taking this lying down. They've responded with restrictions on rare earth minerals that are essential for making chips. They're pouring billions into domestic chip production. Huawei, the tech giant, has been racing to build homegrown alternatives. They've got a chip called the Ascend 910C and plans for four more by 2028. They're even building their own memory components to avoid relying on Western suppliers.

Meanwhile, Taiwan—which manufactures most of the world's advanced chips through TSMC—sits nervously in the middle of all this. If anything disrupted Taiwan's chip production, the global AI boom would grind to a halt.

A few things to watch:

The software moat could crack. Google and Meta are reportedly working together on a project that would make it easier to run AI code on non-NVIDIA chips. If developers can write code once and run it anywhere, NVIDIA's lock-in weakens.

New chip architectures could find niches. As AI moves from "training" (teaching models) to "inference" (running models), different types of chips might work better. NVIDIA just spent $20 billion licensing technology from a startup called Groq that specializes in this.

China could catch up—eventually. Chinese companies have reportedly built a prototype of the advanced machinery needed to make cutting-edge chips. It's nowhere near ready for mass production, but five years is a long time in technology.

For now, though, NVIDIA remains the gatekeeper of the AI revolution. Every company racing to build AI products is ultimately dependent on Jensen Huang's supply chain.

You'll hear these two words constantly in AI discussions. Here's the difference:

Training is teaching an AI model. It's like going to school—expensive, time-consuming, and you only do it once (or occasionally). Training ChatGPT required thousands of chips running for months.

Inference is using a trained model. It's like taking a test—quick, repeated millions of times a day, every time someone asks an AI a question.

The chips optimized for each task are different. Training needs raw power. Inference needs efficiency and speed. As AI becomes more about using models than building them, the chip market is shifting—which is why companies like Groq are suddenly getting attention.

Drop this at your next meeting: "The real margin opportunity in AI chips is moving from training to inference."

Noah Smith's newsletter "Noahpinion" has been tracking the U.S.-China chip war with clarity and nuance. His recent piece on whether export controls are working is a great deep dive for anyone who wants to understand the geopolitical stakes.

That's Gist for this week. See you next week.