How NVIDIA Accidentally Built the AI Revolution

0:00 On October 29th, 2025, NVIDIA became the first company in history worth five trillion dollars. Today, most of its money comes from machines like these: racks of chips, cables and cooling pipes, built for artificial intelligence. In the three months to July 2026, 89 of its 96 billion dollars in revenue came from data centers. But NVIDIA was not built for AI. It was built to make video games look better. So why were chips made for games so good at AI? Rivals built fast chips too, so why did NVIDIA's software matter so much? And what changed when NVIDIA stopped selling chips, and started selling entire data centers?

0:46 To answer that, we have to go back to a Denny's restaurant in 1993. This is the entire history of NVIDIA. In 1993, three engineers met at a Denny's in San Jose, California: Jensen Huang, Chris Malachowsky and Curtis Priem. Huang knew the place well. As a teenager, he had washed dishes at a Denny's. Their idea was a chip that could bring realistic 3D graphics to the personal computer. At the time, Huang later said, 3D graphics was impossible to make cheap. Video games would pay for it. Here is the problem they wanted to solve. A computer's main processor, the CPU, does one complicated task after another. But a 3D game has to color every pixel on the screen,

1:34 hundreds of thousands of them, dozens of times a second. Most of those calculations are small and almost identical. A chip built only for that job could do many of them at once. On April 5th, 1993, they founded a company. They wanted to call it NVision, but a toilet paper maker already had the name. So Huang borrowed from invidia, the Latin word for envy. NVIDIA's first product, the NV1, arrived in May 1995. It tried to do everything at once: 3D graphics, video and sound on a single card. It even had a port for Sega Saturn controllers. But the NV1 drew 3D worlds in its own way. Most of the industry built scenes out of flat polygons.

2:22 The NV1 used curved surfaces. Huang later admitted that this turned out to be incompatible with Direct3D, Microsoft's new standard for games on Windows. NVIDIA was also designing a second chip, the NV2, for Sega's next game console. By Huang's own account, he realized the design was wrong. He told Sega's president, Shoichiro Irimajiri, that the technology would not work, and asked Sega to invest in NVIDIA anyway. Sega agreed. NVIDIA stopped work on the NV2, and let go of more than half of its roughly one hundred employees. Everything now depended on one new chip. Huang later put the odds at fifty-fifty. That chip was the RIVA 128, and it shipped in August 1997.

3:12 This time, NVIDIA built what the market wanted: a fast 3D chip that worked with Direct3D. When it reached stores, NVIDIA had enough money for one month of payroll. Then it sold a million units in four months. For years afterward, Huang opened company presentations with the same line: "Our company is thirty days from going out of business." In January 1999, NVIDIA went public. Later that year, it released the GeForce 256, and marketed it as the world's first GPU, a graphics processing unit. Graphics chips already existed. What changed was that the math of placing and lighting 3D objects moved off the CPU, and onto the graphics chip itself.

3:58 In 2000, Microsoft chose NVIDIA to build the graphics chip for its first Xbox, and NVIDIA bought the graphics assets of its old rival 3dfx. The next year came the GeForce3, with programmable shaders. Before, a graphics chip had a fixed menu of effects. Now, developers could write small programs that ran on every pixel, to make metal shine or water ripple. Its rival ATI fought for the same customers, and in 2006, the chipmaker AMD bought ATI. But programmable shaders had planted an idea. If a graphics chip could run thousands of small programs at once, why should those programs only draw pictures? In November 2006, NVIDIA released the GeForce 8800,

4:48 with 128 processors that could take on any kind of calculation, not just graphics. Every one came with a new technology called CUDA. It let programmers write code for the GPU in C, one of the most common programming languages. Version 1.0 followed in 2007. Think of a CPU as a few brilliant workers who can handle any task, and a GPU as thousands of ordinary workers who can each do one simple job. If you need to add up a billion numbers, the crowd wins easily. But if every step depends on the one before it, the crowd just waits. A GPU is not faster at everything. It is faster at work that splits into many identical pieces.

5:35 When CUDA launched, Wall Street reacted with dismay. For years, it looked like a solution waiting for a problem. Not every bet paid off quickly. In 2008, NVIDIA launched Tegra, a chip for small devices that would one day power the Nintendo Switch. That same summer, it warned investors of a charge of 150 to 200 million dollars, because some of its older chips were failing inside a number of notebook computers. Then, in 2012, two researchers in Toronto bought two GeForce GTX 580 cards. Alex Krizhevsky and Ilya Sutskever worked with Geoffrey Hinton, and they set the cards up in Krizhevsky's bedroom. Krizhevsky wrote his own GPU code to train a neural network.

6:21 The network was too big for one card, so they split it across both, and trained it for five to six days. In the 2012 ImageNet competition, its top-five error rate was 15.3 percent. The next best was 26.2. We told the AI side of this story in our last video. For NVIDIA, the point was this: the breakthrough ran on gaming hardware, using tools NVIDIA had been shipping for six years. The paper even said the results could improve simply by waiting for faster GPUs. One card was no longer enough. AI needed many GPUs working as one. In April 2016, NVIDIA introduced the DGX-1: eight GPUs in a single box, with a list price of 129,000 dollars.

7:11 That August, Jensen Huang delivered the first one himself, to a small research lab in San Francisco called OpenAI. In 2017, NVIDIA's Volta chips added Tensor Cores. Neural networks spend much of their time multiplying huge grids of numbers. A Tensor Core multiplies small blocks of those grids in a single step. Then AI came back to games. In 2018, GeForce RTX cards added dedicated circuits for ray tracing, which simulates how light bounces around a scene, and a feature called DLSS, where a neural network rebuilds a sharper image from a lower-resolution frame. Gaming cards also attracted buyers who never played games.

7:57 Crypto miners bought them in bulk in 2017 and again in 2021, and prices soared. In 2022, the U.S. Securities and Exchange Commission fined NVIDIA 5.5 million dollars for failing to tell investors that mining was driving part of its gaming growth. NVIDIA settled without admitting or denying the charges. In AI, the next bottleneck was the wiring. Training a large model means thousands of GPUs passing results back and forth. If the network between them is slow, the fastest chips sit idle. So in 2019, NVIDIA agreed to buy Mellanox, a maker of high-speed data center networking, for about 6.9 billion dollars. Then it reached for something much bigger:

8:43 the chip designer Arm, for 40 billion dollars. Regulators pushed back, the U.S. Federal Trade Commission sued to block the deal, and in 2022, NVIDIA gave up. Meanwhile, the business was changing shape. In 2020, data center revenue passed gaming for the first time in a quarter. In 2022, NVIDIA announced the H100, and the U.S. began requiring licenses to sell top AI chips to China. Then, on November 30th, 2022, ChatGPT launched, and demand for AI chips exploded. By mid-2023, NVIDIA's data center revenue was up 171 percent from a year earlier, and in May, its value briefly topped one trillion dollars.

9:32 In 2024, NVIDIA announced Blackwell. Its flagship was not a single chip, but a full rack: 72 GPUs and 36 CPUs, wired together and cooled by liquid, built to work as one system. It was delayed by a design flaw that Huang called one hundred percent NVIDIA's fault. In June 2024, NVIDIA briefly became the world's most valuable public company. Then, in January 2025, a new model from the Chinese AI company DeepSeek rattled investors. NVIDIA lost nearly 600 billion dollars in a single day, the largest one-day loss in stock market history. China itself was closing. New export rules hit the H20, a chip NVIDIA had been selling there,

10:22 and NVIDIA took a 4.5 billion dollar charge. Today, its forecasts assume no data center revenue from China at all. Rivals were moving too. AMD sells its own AI accelerators. Google builds its own AI chips, and the AI company Anthropic plans to use up to one million of them. None of it stopped the climb, which brings us back to that five trillion dollar day. NVIDIA says its newest platform, Vera Rubin, is in full production, with the first racks already running at cloud partners. So, why were chips made for games so good at AI? Because the work has the same shape. Drawing a frame means running the same small calculation on hundreds of thousands of

11:08 pixels at once. Training a neural network means running the same multiplications on millions of numbers at once. Not every job works that way, but this one does. Why did the software matter so much? Because a chip is only as useful as the code people can write for it. CUDA shipped on NVIDIA's cards from 2006, years before most people cared. When the AI boom came, researchers already knew the tools. Rivals could build fast chips. Catching up on years of software and habits was much harder. And what changed when NVIDIA started selling entire data centers? The bottleneck moved from the chip to everything around it: the network, the cooling, the power. That made NVIDIA far bigger, and more exposed:

11:55 to shortages, to customers building their own chips, and to governments deciding who is allowed to buy. Thirty-three years ago, three engineers met at a Denny's to make games look better. They nearly failed, more than once. And Huang's old line still says it best: thirty days from going out of business. What matters more in the next era of computing: the chip, or the ecosystem around it? Tell us in the comments. And if you missed it, our last video tells the entire history of AI. See you in the next one.