[AI Library] Chapter 5. GPU as a Weapon
The Jensen Huang Story
Part 1. Building an Empire
Chapter 5. GPU as a Weapon
Kim Kyung-jin
When a Roman legionnaire drew his sword on the battlefield, it was no mere piece of metal. The gladius, short but razor-sharp, was the weapon that let Rome conquer the Mediterranean. A good weapon changes how battles are fought, and a changed style of warfare redirects the course of history.
Silicon Valley in 1999 looked a lot like the Mediterranean coast in the third century BC.
Just as countless city-states had clashed over supremacy in that ancient era, graphics chip companies were crowding the field and strangling one another. 3dfx, ATI, Matrox, S3, and NVIDIA. Amid this tangle of names, Jensen Huang was quietly sharpening his blade.
His blade had a name. GPU. Those three letters did not exist anywhere in the world before 1999.
1999: Inventing a Word That Never Existed
Caesar was great not just because he conquered vast lands. He planted Roman law and language in the lands he took. To name a new concept is to become its owner. Jensen Huang understood this truth by instinct.
On August 31, 1999, NVIDIA announced a new chip and introduced an unfamiliar term. The chip, called GeForce 256, was presented as the world's first GPU. GPU: short for Graphics Processing Unit.
Why was this word, which seems obvious today, so revolutionary? To understand that, you need to know how the computer world worked in 1999.
In those days, the computer had only one king. The CPU, the Central Processing Unit. This chip, made by Intel, governed everything that happened inside a computer. Writing a document, running a calculation, playing music, running a game; everything had to pass through the CPU.
The CPU was the computer's brain and heart.
So what was a graphics card? It was little more than a servant that splashed the CPU's processed results onto the monitor. People called it a graphics accelerator. As the word "accelerator" suggests, it meant a helper device that sped up the CPU's work a little. The chip's name made its limitations plain from the start.
Jensen Huang questioned this arrangement.
Was rendering three-dimensional graphics on screen really a simple task that needed only a little assist from the CPU?
His answer was an emphatic no.
Drawing a three-dimensional object on screen requires complex math.
You have to calculate the position of each object, calculate where light is coming from, and calculate where shadows fall. These calculations must happen millions of times per second for the image to move smoothly. The CPU had to handle all of this on its own while also computing enemy movements inside a game, so the burden was enormous.
Jensen Huang thought: Isn't this a fundamentally different kind of computation from what the CPU does? If the CPU is a single genius mathematician, wouldn't graphics calculation be better served by thousands of ordinary clerks working at the same time?
He decided to give the graphics chip a new name. He declared it would no longer be called an accelerator. He defined it as an independent processing unit that served as a second brain for the computer, standing on equal footing with the CPU.
NVIDIA defined the GPU this way:
"A single-chip processor that integrates transform, lighting, triangle setup and clipping, and a rendering engine, capable of processing a minimum of 10 million polygons per second." One phrase in this definition deserves attention: Processing Unit.
Until that moment, "Processing Unit" had been a sacred title reserved for the CPU alone. Jensen Huang dared to attach that title to his own chip. This was not just a marketing term. It was a declaration of independence, a challenge thrown at the CPU's kingdom.
People in the industry scoffed.
A graphics card calling itself a processor? But language has power. The moment people started using the term GPU, the chip's standing began to shift inside their heads. It was no longer seen as a helper device but as an independent brain.
Whoever controls the name controls the concept, and whoever controls the concept controls the market. Jensen Huang grasped this truth exactly.
GeForce 256: The Opening Act of Parallel Computing
Having staked out a new name, he had to prove the substance behind it. On October 11, 1999, NVIDIA officially launched the GeForce 256. Carrying the title of the world's first GPU, the chip lived up to expectations.
To explain what set the GeForce 256 apart from earlier graphics chips, you first need a basic idea of how three-dimensional graphics are created.
The 3D characters and buildings you see on screen are made up of countless small triangles called polygons. Every time a character moves, the new position of each of those thousands of triangles must be calculated. This is called coordinate transformation, or in technical terms, "Transform."
Next comes calculating the effects of light.
You have to determine where the sun is shining from, where shadows fall, which parts of an object are bright and which are dark. This is called lighting processing, or simply Lighting. Combined with the first word, Transformation, the two are known as T&L.
Before 1999, T&L calculations fell entirely on the CPU. The graphics card only stepped in to paint colors after the CPU had finished its math.
Think of it this way. On a construction site, the architect draws the blueprints, calculates coordinates, and decides the direction of light before the painter even picks up a brush. That was how it worked.
The GeForce 256 flipped this structure on its head. It built a T&L engine right into the chip. Now the CPU only needed to issue rough commands. "Put a pillar over there."
The GPU would then calculate the pillar's coordinates on its own, analyze the direction of light, and render the shadows. The CPU was freed from this grueling work.
So what happened when the CPU had less to do?
The CPU could now focus on computing enemy AI and physics inside the game. Characters got smarter. Explosions and falling objects looked more realistic. The entire gaming experience jumped up a level.
The GeForce 256 proved itself in raw numbers too.
The chip could process over 10 million polygons per second. That was more than 50 percent faster than competing products at the time, such as 3dfx's Voodoo3 or NVIDIA's own previous chip, the RIVA TNT2.
Gamers cheered. The monsters on screen no longer looked like boxy crates. They felt like living creatures with bulging muscles and glistening skin.
But Jensen Huang was looking somewhere far beyond the game screen. He was focused on the essence of the T&L engine: the concept of parallel computing.
What is parallel computing? Here's a simple way to picture it.
Imagine a classroom where thousands of exam papers, each containing 20 questions, need to be graded. The CPU approach is like having one brilliant teacher pick up each paper and grade all 20 questions in order, one sheet at a time. It's not slow, but it has to go through them one by one.
The GPU approach is like assigning each teacher to grade only one specific question across all the papers. If there are 20 questions, 20 teachers each handle their own. The overall speed is far faster this way.
3D graphics is exactly this kind of situation.
When millions of pixels on the screen need to be calculated at the same time, thousands of workers beat one genius. The GeForce 256 was the first product to implement this philosophy of parallel processing in hardware.
Jensen Huang was quietly looking ahead at this moment.
What if this parallel processing power could be used for something other than games? What if it could be applied to scientific calculations or data analysis? The idea hadn't yet taken concrete shape.
But it would later bloom under the name CUDA.
And that CUDA would ride the massive wave of artificial intelligence and change the world.
The GeForce 256 of 1999 was a dazzling toy for gamers.
But in the history of computing, this chip carried a far greater meaning. It marked the moment when parallel computing appeared as a new way of thinking in a world that had been dominated by serial processing.
Just as the steam engine appeared in a world ruled by horse-drawn carts, the very method of computing itself began to change.
The Fall of 3dfx and the Rise of a New King
History has no permanent champion. The Roman Empire couldn't last a thousand years. Genghis Khan's empire fractured within a generation. The history of technology is no different.
In the mid-1990s, the undisputed ruler of the 3D graphics market was a company called 3dfx. Their graphics card, the Voodoo, was both legend and religion among gamers. Owning a Voodoo card was every teenager's dream in those days. At its peak, 3dfx commanded a market share of roughly 85 percent.
When NVIDIA's first chip, the NV1, failed miserably, it was 3dfx that ruled the market. In most people's eyes, NVIDIA was just another chaser in a crowded pack. A dwarf standing before a giant. But the throne of number one is a poisoned chalice. Those drunk on victory grow arrogant, and the arrogant fail to read change.
3dfx's first mistake was misjudging technical standards.
They insisted on Glide, their own proprietary programming language.
Glide was a closed technology that only worked on 3dfx cards. If a game developer built a game in Glide, that game only ran properly on 3dfx hardware.
At first, this looked like an advantage. 3dfx cards dominated the market, after all. But when Microsoft started pushing DirectX, an open standard, the picture changed. Games built on DirectX ran on any graphics card. Developers began tilting toward DirectX. They could sell their games to a far larger audience.
NVIDIA read the shift quickly.
Instead of clinging to proprietary technology, Jensen Huang chose to fully support open standards like DirectX and OpenGL. He called it an "ecosystem" strategy.
The idea was simple: don't try to do everything alone. Work with others and grow together.
3dfx's second mistake was restructuring its business model.
Originally, 3dfx only designed chips. Multiple manufacturers built and sold the actual graphics cards. Then in 1998, 3dfx acquired STB, a graphics card manufacturer. The ambition was to control everything from chip design to card manufacturing to retail sales. On paper, it looked like a smart way to capture all the profit.
Reality told a different story.
The moment 3dfx started selling cards directly, the other manufacturers who had been building cards with 3dfx chips turned hostile. They lined up behind NVIDIA instead. 3dfx found itself fighting alone, while NVIDIA commanded an alliance.
Jensen Huang took the opposite path.
He focused solely on chip design. The actual fabrication went to a Taiwanese company called TSMC. Building finished cards from those chips was left to dozens of partners: ASUS, MSI, Gigabyte, and many others. When NVIDIA made a good chip, scores of companies sold cards carrying that chip under their own brands, all around the world.
The way Rome folded conquered nobles into alliances rather than making enemies of them, Jensen Huang never turned his partners into adversaries. This ecosystem strategy became the backbone of the NVIDIA empire.
3dfx's third mistake was misreading the technology trend.
They dismissed 32-bit color. 32-bit color could render far richer, more natural visuals than 16-bit. But 3dfx argued that 32-bit color would drag down frame rates, and stuck with 16-bit. "Gamers care about speed, not color depth," was their verdict.
NVIDIA saw it differently.
Jensen Huang decided that combining 32-bit color with hardware T&L could deliver both speed and image quality. The result was the GeForce 256. The moment people saw a GeForce screen, the Voodoo started to feel like yesterday's hardware.
The most decisive difference was speed, specifically the speed of product launches.
Jensen Huang set a fearsome target: a new chip every six months. Moore's Law, the prediction by Intel co-founder Gordon Moore, held that chip performance doubled every 18 to 24 months. Jensen Huang was chasing a pace far faster than that.
For NVIDIA's employees, it was a death march.
But that punishing tempo let NVIDIA release the RIVA 128, RIVA TNT, RIVA TNT2, and GeForce 256 in rapid succession. While competitors paused to catch their breath, NVIDIA already had the next product on shelves.
While 3dfx struggled to develop its next-generation chip, the GeForce 256 hit the market. Armed with hardware T&L, it made 3dfx's latest products look like relics. Gamers were ruthless. They said goodbye to yesterday's hero and cheered for the new king.
On November 16, 2000, 3dfx announced it would stop manufacturing cards. It was a declaration to go back to selling chips only. But it was already too late. A month later, on December 15, 3dfx's creditors initiated bankruptcy proceedings, and the company agreed to sell its assets to NVIDIA.
The acquisition price was $70 million in cash and $1 million worth of NVIDIA stock. A shabby sum for an empire that once held 85 percent market share. Jensen Huang picked up the fallen rival's sword and strapped it to his belt. 3dfx's engineers and intellectual property were absorbed into NVIDIA.
With that, the long first act of the graphics war ended in NVIDIA's victory. A formidable competitor named ATI still remained, but NVIDIA had become the undisputed ruler of the market.
The fall of 3dfx left an important lesson in the history of technology.
Technical prowess alone isn't enough to win. You need the vision to read where standards are heading, a strategy for building an ecosystem, and the speed to adapt when things change. Without all three, survival is not guaranteed.
3dfx was no slouch in technology, but the company stumbled on the other three fronts.
Jensen Huang did not let victory go to his head. Watching 3dfx collapse, he must have felt the old terror of being "30 days from bankruptcy" all over again. Today's winner can become tomorrow's loser. The moment you get comfortable, you fall apart. That was the lesson 3dfx taught him.
The success of the GeForce 256 brought NVIDIA a massive influx of cash.
The question was where to spend it. A typical executive would have paid dividends to shareholders or poured the money into expanding existing operations. Jensen Huang was not a typical executive.
He decided to invest it in a place nobody else was willing to look at.
The power of parallel computing he had witnessed through the GeForce 256 felt too valuable to confine to gaming alone.
What if thousands of tiny calculators working simultaneously could be applied to scientific research or data analysis? What if they could replace supercomputers?
This idea would eventually reach the world under the name CUDA. But it would prove to be a lonely investment, generating zero revenue for more than a decade.
Wall Street laughed. Analysts called it a waste. Jensen Huang did not stop.
In the next chapter, we will talk about that lonely decade of investment, and how a gamble that everyone dismissed as insanity rode a massive wave called artificial intelligence and changed the world.
Three letters Jensen Huang coined in 1999: GPU. It was not just a marketing term. It was the invention of a concept that altered the history of computing. It opened a door from the serial era, where a single CPU handled everything alone, to a parallel era, where thousands of small brains worked together. Beyond that door, a new world called artificial intelligence was waiting.
Jensen Huang did not yet know the specific shape of that world.
But he sensed that was the direction to go. "We started this company to solve problems that general-purpose computers cannot solve." Those words carry a consistent philosophy spanning from the 1993 founding to the 1999 invention of the GPU to the arrival of the AI era that followed.
Just as Roman legionaries armed with the gladius, a short sword, conquered the Mediterranean, NVIDIA, armed with the GPU, began building a new empire in the digital world.
But a true empire is not built overnight. It is completed only at the end of a solitary journey, staying up through countless nights, walking a road that no one else believes in.
Kim Kyung-jin
Attorney · Former Member of the National Assembly · AI Policy Researcher
© 2026 Kim Kyung-jin. All rights reserved.