Corrp.com | Company Histories podcast - Page 2

March 16, 2026 Nvidia: From A Denny's Booth To A Four Trillion Dollar Empire

There is a Denny's restaurant on Berryessa Road in East San Jose, California. It looks like every other Denny's you have ever seen. Vinyl booths, laminated menus, bottomless coffee. The kind of place where truckers stop at two in the morning and college students nurse hangovers on Saturday. Nobody would look at it and think, this is where a four trillion dollar company was born. But in late 1992, three men slid into a booth at that Denny's, ordered coffee, and sketched out an idea that would eventually reshape the entire computing industry. Their names were Jensen Huang, Chris Malachowsky, and Curtis Priem. They were engineers. They were friends. And they had a vision for a future that, at the time, registered as a zero dollar market.

To understand how they got to that booth, you have to understand where each of them came from, because the story of Nvidia is inseparable from the story of Jensen Huang, and Jensen Huang's story begins on the other side of the Pacific Ocean.

Jen-Hsun Huang was born on February 17, 1963, in the city of Tainan, in southern Taiwan. His father, Huang Hsing-tai, was a chemical engineer. His mother, Lo Tsai-hsiu, was a grade school teacher. When Jensen was five years old, the family moved to Thailand, chasing economic opportunity. But the Vietnam War was raging across the region, and his parents grew anxious about the instability. They decided their sons needed to get out. So when Jensen was nine years old, he and his older brother were sent to live with an uncle in Tacoma, Washington. The plan was for the boys to attend a reputable school and get a proper American education. But there was a miscommunication. The uncle misunderstood the arrangements, and instead of enrolling the boys in a good school in Washington state, he sent them to a boarding school in Oneida, Kentucky. It was not the elite academy their parents had envisioned. It was closer to a reform school, a rural institution where many of the students had been sent by the juvenile courts. Jensen was nine years old, far from home, barely spoke English, and was surrounded by tough kids, many of them teenagers with knives. His roommate was covered in tattoos. The other boys bullied him. He cleaned toilets to earn his keep. It was, by any measure, a harsh introduction to America.

But Huang has said, many times, that this period forged something in him. It taught him self-reliance. It taught him that suffering is not something to avoid but something to survive and, eventually, to use. Decades later, standing on stage at Stanford University, he would tell a room full of elite business students, "I wish upon you ample doses of pain and suffering." He was not being cruel. He was being honest. As he explained to those students, "Greatness comes from character. And character isn't formed out of smart people. It's formed out of people who suffered."

Eventually Jensen's parents were able to join the boys in America, and the family settled in the suburbs of Portland, Oregon. Jensen enrolled at Aloha High School, where he became an excellent student and, improbably, a nationally ranked table tennis player. After high school, he chose Oregon State University because the tuition was affordable. He studied electrical engineering and graduated in 1984 with a bachelor's degree with highest honors. He was immediately recruited by several Silicon Valley companies, including Texas Instruments, Advanced Micro Devices, and LSI Logic. He chose AMD first, working as a microchip designer. But when he heard about cutting-edge chip design processes at LSI Logic, he jumped ship and joined LSI as a technical officer. And this is where the threads of destiny begin to weave together, because at LSI Logic, Huang found himself doing work for a client called Sun Microsystems. And at Sun Microsystems, two other engineers were already plotting the future.

While working at LSI Logic during the day, Huang was simultaneously pursuing graduate classes at night at Stanford University. He was determined to deepen his technical knowledge, and Stanford was the best place in the world to do it. In 1992, he earned his master's degree in electrical engineering from Stanford. He was twenty-nine years old, armed with a top-tier education, years of industry experience, and a growing conviction that the future of computing lay in graphics. But more importantly, through his work at LSI Logic and his connections at Sun Microsystems, he had met the two men who would become his co-founders.

Chris Malachowsky was born on May 2, 1959, in New Jersey, and grew up in the Oakhurst section of Ocean Township in Monmouth County. He studied electrical engineering at the University of Florida, earning his bachelor's degree in 1983, then continued his education with a master's degree from Santa Clara University in 1986. Before arriving at Sun Microsystems, Malachowsky had worked at Hewlett-Packard as a manufacturing design engineer. At Sun, he became a senior staff engineer, specializing in integrated circuit design. He was quiet, methodical, and deeply technical. Over the course of his career, he would accumulate close to forty patents, most of them related to graphics chip design.

Curtis Priem's path was different but converged in the same place. Born into a family that valued education, Priem attended Rensselaer Polytechnic Institute in Troy, New York, one of the oldest engineering schools in the country. He graduated in 1982 with a bachelor's degree in electrical engineering. Right out of school, Priem did something remarkable. He designed the IBM Professional Graphics Adapter, one of the first graphics processors for the personal computer. This was the early nineteen eighties, when personal computers were still novelties, and the idea that they might one day display rich, dynamic graphics was considered fringe thinking. From 1986 to 1993, Priem worked at Sun Microsystems as a senior staff engineer, where he developed the GX graphics chip. He was brilliant, inventive, and restless. By the early nineties, he held nearly two hundred patents.

So here were three engineers, all working in or around Sun Microsystems, all obsessed with graphics processing, all sensing that the industry was about to undergo a seismic shift. The personal computer was evolving. Microsoft was building Windows. Video games were getting more complex. And the three of them believed, passionately, that three-dimensional graphics would become the next great frontier of computing. The problem was, in 1992, the market for 3D graphics on personal computers was essentially nonexistent. As Nvidia's long-time board member Mark Stevens from Sequoia Capital would later describe it, it was "a zero billion dollar market." But Huang, Malachowsky, and Priem saw what others could not. They saw that this zero dollar market was about to become everything.

Curtis Priem was the first to make a move. He resigned from Sun Microsystems, effective December 31, 1992. According to Priem, this put pressure on the other two, because they had all been talking about starting a company together, and now Priem was out there alone. He did not want to "flail alone," as he put it, and neither did the others want to leave him hanging. Jensen Huang left LSI Logic and officially joined Priem on February 17, 1993, which also happened to be Huang's thirtieth birthday. Malachowsky left Sun in early March. The three of them began meeting regularly at that Denny's on Berryessa Road. Jensen chose Denny's because he had once worked there as a dishwasher and busboy during his college years, and he knew the place was "quieter than home and had cheaper coffee."

In those early meetings, the three men debated what kind of company to build. They knew they wanted to make graphics chips for personal computers, but the details were everything. Even though Huang was the youngest of the three, just thirty years old, both Priem and Malachowsky believed he should be the CEO. As Priem would later recall, "We basically deferred to Jensen on day one. We told him, you're in charge of running the company, all the stuff Chris and I don't know how to do."

The company needed a name. All their early files had been labeled "NV," shorthand for "next version." When they went to register the company, they looked for words containing those two letters. They considered "NVision," but that name was already taken by a toilet paper company. Then someone suggested "invidia," the Latin word for envy. In ancient Rome, invidia meant an intense gaze, associated with the evil eye, with desire and admiration. The founders liked the symbolism. They wanted to build products that others would envy. They dropped the "i" and became Nvidia. The green color scheme they would later adopt was a deliberate nod to the phrase "green with envy." Their logo incorporated an eye motif. Everything was intentional.

On April 5, 1993, Nvidia was officially incorporated. They had forty thousand dollars in the bank.

But forty thousand dollars does not get you very far in the semiconductor business. Designing a chip is extraordinarily expensive. You need engineers, software tools, fabrication partners, and months of development time before you have anything to sell. The founders knew they needed venture capital, and they needed it fast. This is where a legendary figure in Silicon Valley enters the story. Don Valentine was the founder of Sequoia Capital, one of the most successful venture capital firms in history. Sequoia had backed Apple, Cisco, Oracle, and Yahoo. Valentine was famously intimidating, a former National Semiconductor executive with a gruff manner and an unforgiving eye for bullshit. Wilfred Corrigan, the head of LSI Logic where Jensen had worked, apparently told Valentine to take a look at Huang's new company. Valentine agreed to invest. Sequoia Capital, along with Sutter Hill Ventures, put in two million dollars at a post-money valuation of six million dollars. Valentine's warning to Jensen was characteristically blunt. As the story goes, he told Huang, "If you lose my money, I'll kill you."

Mark Stevens, a partner at Sequoia, joined Nvidia's board as the firm's representative. He would remain on the board for decades, becoming one of the most important advisors in the company's history. Over the next several years, Nvidia would raise a total of approximately twenty million dollars in venture capital from Sequoia, Sutter Hill, and other investors. But money alone was not going to determine whether Nvidia lived or died. What mattered was the product, and the first product was about to become one of the most spectacular failures in the history of the semiconductor industry.

Nvidia's first chip was called the NV1. Development took two years, and the chip finally shipped in May 1995. To understand why the NV1 was a disaster, you have to understand the technical debate that was consuming the graphics industry at the time. Most companies were building their 3D graphics chips around a technique called triangle-based rendering, which breaks up three-dimensional objects into tiny triangles and then processes those triangles to create images on screen. But Curtis Priem, Nvidia's chief technical officer, had a different idea. He wanted to use something called quadratic texture mapping, which used curved surfaces instead of flat triangles. Mathematically, quadratics were more elegant. They could, in theory, produce smoother images with fewer polygons. But there was a problem. Microsoft was developing a new graphics standard called DirectX, and DirectX used triangles. Only triangles. Priem's quadratic approach was fundamentally incompatible with the direction the entire industry was heading.

The NV1 launched as the Diamond Edge 3D, manufactured by SGS-Thomson Microelectronics. It was marketed as a multimedia card because, in addition to 3D graphics, it integrated a thirty-two channel sound card and even had joystick and Sega Saturn controller ports. It was, as one observer later described it, a Swiss Army knife of a chip, trying to do everything at once. And it was expensive. The market wanted cheap, simple 3D acceleration. What Nvidia delivered was an over-engineered, overpriced product built on the wrong technical foundation.

Nvidia shipped approximately 250,000 NV1 units to Diamond Multimedia. Nearly 249,000 of them were returned. The failure cost Nvidia roughly ten million dollars, a devastating amount for a startup that had raised only twenty million total. Jensen Huang would later admit that the NV1 was built on a flawed architectural bet. But at the time, the company still had hope. They had been working with Sega on a follow-up chip, the NV2, which was intended to power Sega's next-generation console. Sega had invested in the relationship, and Nvidia hoped this partnership would save them.

But then Microsoft made its move. In late 1996, Microsoft formally committed DirectX to triangle-based rendering. This was not a suggestion. It was an industry mandate. Any chip that wanted to work with Windows games had to use triangles. Nvidia's quadratic approach was dead. And the Sega partnership was collapsing. Sega had been evaluating multiple vendors for its upcoming Dreamcast console, and ultimately chose PowerVR's triangle-based technology over Nvidia's quadratic design. The NV2 project was cancelled.

This was the darkest moment in Nvidia's history. The company had spent all of its money on the wrong architecture. Its only major partnership had just evaporated. And its first product had been rejected by the market. Nvidia was thirty days from running out of cash.

But here is where the story takes a turn that Jensen Huang would retell for decades. Sega's president at the time was a man named Shoichiro Irimajiri. When it became clear that Sega would not use Nvidia's chip for the Dreamcast, Huang went to Irimajiri with an extraordinary request. He asked Sega to terminate the contract but still pay Nvidia the remaining money owed under the deal, roughly five million dollars. Nvidia had not delivered what Sega wanted. By any rational measure, Sega owed Nvidia nothing. But Irimajiri agreed. According to Huang, Irimajiri believed in Nvidia's potential and persuaded Sega's management to give the company a lifeline. Huang later reflected that this act of generosity gave Nvidia six more months of life. As Mark Stevens, the Sequoia board member, would later say, "You can't discount the kindness of people when you're starting your company."

With that five million dollars, Huang did something brutal but necessary. He laid off more than half of Nvidia's workforce, reducing the headcount from around one hundred employees to approximately forty. He told the remaining team that they had one shot. One chance to build a chip that could save the company. There would be no second product cycle if this one failed. Every engineer in the building understood the stakes. They were designing a chip that would determine whether they had jobs in six months.

The chip was called the RIVA 128, internally designated NV3. RIVA stood for "Real-time Interactive Video and Animation." This time, Nvidia abandoned the quadratic approach entirely and built a chip based on triangle-based rendering, fully compatible with Microsoft's Direct3D standard. The team adopted an innovative development methodology. Instead of designing the hardware first and writing the software afterward, they ran the software on an emulator before the hardware was finalized. They purchased a used emulator from a bankrupt company called Icos, which allowed them to test and debug their chip design in software before committing it to silicon. This meant that when the RIVA 128 taped out, when the physical chip was fabricated for the first time, it worked on the first try. In the semiconductor industry, where first silicon almost never works perfectly, this was a remarkable achievement.

The RIVA 128 shipped in August 1997 and was a sensation. It offered both 2D and 3D acceleration on a single chip, with a 100 megahertz core and memory clock that essentially doubled the performance of 3dfx's Voodoo 1, the dominant 3D accelerator of the era. Within four months, Nvidia sold one million RIVA 128 units. The company had come back from the edge of bankruptcy.

This near-death experience left an indelible mark on Jensen Huang. To this day, he opens employee meetings with the phrase, "Our company is only thirty days away from going out of business." It is not a statement of fear. It is a philosophy. It is the idea that complacency kills, that the moment you believe you are safe is the moment you begin to die. And this philosophy would define Nvidia's culture for the next three decades.

But the RIVA 128 was only the beginning. Nvidia was now in a race, and the race was getting faster. In 1997 and 1998, the PC graphics market was one of the most competitive in all of technology. There were at that time roughly seventy companies trying to build graphics accelerators for personal computers. The big names were 3dfx, with its legendary Voodoo line; ATI Technologies, with its Rage series; S3 Graphics; Matrox; and numerous others. Each company was releasing new chips every twelve to eighteen months, and the performance improvements were breathtaking. But Nvidia, having nearly died, had learned something crucial: speed kills, and not just performance speed, but development speed. The company that could release new chips faster than its competitors would win the market, because gamers and PC builders always wanted the newest, fastest thing.

Jensen Huang established what he called a six-month product cycle. While most competitors were releasing new chips every twelve to eighteen months, Nvidia would release a new chip every six months. This was insanely aggressive. Designing a new graphics chip in six months meant overlapping product development, running multiple chip designs simultaneously, and pushing engineers to their limits. But it worked. In early 1998, Nvidia released the RIVA TNT, codenamed NV4. The "TNT" stood for "TwiN Texel," a reference to the chip's ability to process two texture elements simultaneously. The RIVA TNT added a second pixel pipeline, essentially doubling the rendering speed of its predecessor, and used significantly faster memory. It was announced in March 1998 and shipped in August, directly competing with 3dfx's Voodoo 2 and ATI's Rage 128. The RIVA TNT was not the fastest chip on the market in every benchmark, but it was close, and it was a single-chip solution that did not require the elaborate multi-card setups that 3dfx's SLI technology demanded.

Then came the RIVA TNT2 in early 1999, which pushed performance even further with 32-bit color rendering at a time when most competitors were still stuck at 16-bit. Nvidia was establishing a pattern: relentless iteration, each generation faster than the last, each release timed to keep competitors off balance.

And then, on August 31, 1999, Nvidia did something that would define the company forever. They released the GeForce 256 and, with it, coined a term that would enter the global vocabulary. They called the GeForce 256 "the world's first GPU," or Graphics Processing Unit. Nvidia defined a GPU as "a single-chip processor with integrated transform, lighting, triangle setup/clipping, and rendering engines that is capable of processing a minimum of ten million polygons per second." The key innovation was hardware transform and lighting, or T&L. Previously, the CPU handled the mathematical calculations that determined the position and lighting of 3D objects, and the graphics card only handled the final rendering. The GeForce 256 moved those calculations onto the graphics chip itself, freeing the CPU for other tasks and dramatically improving 3D performance. This was not just a new product. It was a new category. NVIDIA had invented the GPU, and the term stuck.

The GeForce 256 shipped on October 11, 1999. But even before that, Nvidia had achieved another milestone. On January 22, 1999, the company went public on the Nasdaq stock exchange, trading under the ticker symbol NVDA. The IPO price was twelve dollars per share, giving the company a market capitalization of approximately six hundred million dollars. For Sequoia Capital and Sutter Hill Ventures, which had invested at a six million dollar valuation just six years earlier, this represented a one-hundred-fold return. It was a spectacular outcome for everyone involved.

But the IPO was just the opening act. As the year 2000 began, Nvidia was on an absolute tear. The GeForce 2, launched in April 2000, pushed performance even further. And then came a deal that would thrust Nvidia into the living rooms of millions of homes worldwide. Microsoft was developing a new gaming console to compete with Sony's PlayStation 2, and they needed a graphics chip. After evaluating multiple vendors, Microsoft chose Nvidia's technology. The original Xbox, which launched in November 2001, contained a GPU derived from the GeForce 3 architecture, specifically a custom chip called the NV2A. Microsoft reportedly paid Nvidia a two hundred million dollar advance for the project. It was a massive validation of Nvidia's technology and a sign that the company's chips were now powering not just PC gaming but console gaming as well.

Meanwhile, Nvidia made a move that would eliminate its most dangerous rival. In December 2000, Nvidia acquired most of 3dfx's assets for seventy million dollars in cash and one million shares of Nvidia common stock. 3dfx, the company that had once defined the 3D gaming revolution with its Voodoo line of accelerators, had overextended itself. It had tried to become a full card manufacturer instead of just a chip designer, burning through cash and alienating its hardware partners. By the time Nvidia came knocking, 3dfx was essentially dead, and Nvidia picked up its intellectual property, its patents, and its brand for what amounted to pocket change. 3dfx ceased supporting its products on February 15, 2001, and formally filed for bankruptcy in October 2002. The company that had once been synonymous with PC gaming was gone, and Nvidia had absorbed its legacy.

With 3dfx out of the picture, the GPU market consolidated into a two-horse race between Nvidia and ATI Technologies, a Canadian company headquartered in Markham, Ontario. ATI had been a formidable competitor since the mid-nineties, and in the early 2000s, the rivalry between the two companies became one of the fiercest in all of technology. ATI released the Radeon DDR in April 2000, and in a humorous bit of corporate theater, Nvidia announced the GeForce 2 GTS just twenty-four hours later. The two companies were in a relentless back-and-forth, each trying to one-up the other with every product cycle.

And for a while, ATI had the upper hand. In 2002, ATI released the Radeon 9700 Pro, a graphics card that was, by virtually every measure, the best GPU on the market. It was fast, it was cool, it was quiet, and it handled the latest games beautifully. Nvidia's response was the GeForce FX 5800, codenamed NV30, and it was one of the most embarrassing products in the company's history. The GeForce FX 5800 ran so hot that it required an enormous dual-slot cooling fan that reviewers immediately nicknamed "the Dustbuster" because of its resemblance to a handheld vacuum cleaner. At full speed, the fan generated seventy-seven decibels of noise, compared to sixty-four decibels for the Radeon 9700 Pro. Reviewers compared it to a hair dryer mixed with a dentist's drill. And it was not just loud. It was slow. The GeForce FX 5800 had only four pixel pipelines compared to the Radeon 9700's eight, and three vertex processors compared to ATI's four. The card could barely compete with a GPU that ATI had released nine months earlier, let alone ATI's newer Radeon 9800 Pro. It was a humiliating defeat.

Nvidia scrambled to recover, releasing the GeForce FX 5900 Ultra in May 2003 as a replacement for the disastrous 5800. But the damage was done. For a brief period, ATI was the king of the GPU market, and Nvidia was playing catch-up. It would not be until the GeForce 6 series in 2004 that Nvidia fully regained its competitive footing, introducing Shader Model 3.0 support and dramatically improved floating-point performance.

The Xbox deal, meanwhile, had turned sour. In 2002, Microsoft and Nvidia entered into arbitration over pricing disputes. Microsoft wanted discounts on the chips Nvidia was supplying for the Xbox. Nvidia's SEC filing indicated that Microsoft was seeking a thirteen million dollar discount. The relationship had soured to the point where Microsoft decided that for its next console, the Xbox 360, it would partner with ATI instead. The dispute was privately settled in February 2003, but the message was clear: even a massive deal with the world's largest software company could turn into a headache. Nvidia ceased producing the Xbox's GPU in August 2005.

Meanwhile, the GPU wars were producing astonishing technological progress. Each generation of chips from both Nvidia and ATI pushed the boundaries of what was possible in real-time graphics. The GeForce 7 series in 2005 introduced SLI technology, which allowed two Nvidia GPUs to work together in a single computer, doubling performance for gamers willing to pay the premium. ATI countered with CrossFire, its own multi-GPU solution. The two companies were locked in an arms race, and the primary beneficiaries were gamers, who were treated to ever-more-spectacular visual experiences. Frame rates went up. Resolutions went up. Effects like anti-aliasing, anisotropic filtering, high dynamic range lighting, and motion blur became standard. Video games went from looking like crude cartoons to approaching cinematic quality, and this transformation was driven almost entirely by the GPU competition between Nvidia and ATI.

In 2006, AMD acquired ATI Technologies for five point four billion dollars, one of the largest semiconductor acquisitions of its time. The deal was intended to allow AMD to offer integrated CPU-plus-GPU solutions, competing with Intel on multiple fronts. ATI's Radeon brand continued under AMD's ownership, and the GPU rivalry continued, but now it was Nvidia versus AMD, with Intel lurking on the periphery. The competitive dynamic remained intense, but Nvidia's six-month product cadence and its growing investments in software and developer tools began to give it a structural advantage that AMD struggled to match.

But while the consumer GPU wars raged on, Jensen Huang was thinking about something much bigger. The mid-2000s were a period of strategic transformation for Nvidia, and the decisions Huang made during this period would prove to be the most consequential in the company's history. Not because they paid off immediately. They did not. But because they positioned Nvidia for a revolution that was still a decade away.

The first pivotal decision was the launch of CUDA. The origins of CUDA trace back to the early 2000s, when a Stanford PhD student named Ian Buck began experimenting with using GPUs for purposes beyond rendering graphics. Buck realized that the massively parallel architecture of GPUs, designed to process millions of pixel calculations simultaneously, could be repurposed for general scientific computing. He developed a language called Brook that allowed programmers to run non-graphics calculations on GPU hardware. In 2004, Nvidia hired Buck and paired him with John Nickolls, the company's director of architecture for GPU computing. Together, they began transforming Brook into something far more ambitious.

In November 2006, alongside the launch of its G80 GPU, Nvidia released CUDA, which stood for Compute Unified Device Architecture. CUDA was a parallel computing platform and programming model that allowed developers to write code in C++, the industry-standard programming language, and run it on the GPU. For the first time, a scientist or researcher could write code that looked like standard C++, compile it, and execute it on the GPU without knowing anything about 3D graphics. The initial CUDA software development kit was made public on February 15, 2007, for both Windows and Linux.

This was a bet on a market that did not yet exist. In 2006, the idea of using graphics chips for scientific computing was considered eccentric at best. Nvidia's core business was selling graphics cards to gamers and PC enthusiasts. Investing in a programming platform for scientists and researchers made no obvious business sense. The data center GPU market, as the board member Mark Stevens would later note, was "a zero billion dollar market." Jensen Huang was betting on a future that he could see but that very few others could.

In June 2007, Nvidia took the next step, launching the Tesla line of GPU computing products specifically designed for scientific and technical applications. The "Tesla" name was a tribute to the inventor Nikola Tesla, and the products were marketed not to gamers but to researchers, scientists, and engineers who needed massive parallel processing power. The first Tesla GPU cluster was shipped to the Max Planck Institute in Germany in 2008, supporting research in molecular biology. Later that year, Tokyo Institute of Technology deployed 170 Tesla S1070 servers in its TSUBAME 1.2 supercomputer, which became the first GPU-equipped system to appear on the TOP500 list of the world's fastest supercomputers.

But CUDA and Tesla were slow burns. They would not generate significant revenue for years. In the meantime, Nvidia had other ambitions and challenges to confront. In February 2008, Nvidia made a strategic acquisition that would pay dividends for years. It acquired Ageia, the company that had developed the PhysX physics engine and the world's first dedicated physics processing unit. Ageia's technology allowed games to simulate realistic physical interactions, from explosions and particle effects to fluid dynamics and cloth movement. After the acquisition, Nvidia integrated PhysX directly into its GPU hardware, allowing its graphics cards to handle physics calculations alongside rendering. This gave Nvidia-powered games an edge in visual quality, because developers could add complex physics effects that would only run well on Nvidia hardware. It also added another layer to Nvidia's ecosystem moat, another reason for developers and gamers to choose Nvidia over AMD.

Also in 2008, Nvidia announced the Tegra system-on-chip, entering the mobile processor market. The Tegra APX 2500 was designed for smartphones and mobile internet devices, promising desktop-class graphics in a pocket-sized package. The first commercial product to use the Tegra was Microsoft's Zune HD media player in September 2009. Nvidia followed with the Tegra 2 in 2010, the world's first dual-core mobile processor, and the Tegra 3 in 2011, the world's first quad-core mobile processor. The company was trying to break into the smartphone and tablet market, competing against Qualcomm's Snapdragon chips and Apple's custom-designed processors.

The mobile push was only partially successful. While Tegra found its way into some Android tablets and smartphones, Nvidia never gained significant market share against Qualcomm, which had locked up relationships with most major handset makers. But the Tegra line would eventually find a different purpose, one that would prove enormously valuable: powering gaming devices and, eventually, autonomous vehicles.

The year 2008 also brought crisis. The global financial meltdown sent Nvidia's stock plummeting by eighty percent. And the company was hit by a different kind of crisis: a hardware defect. In July 2008, Nvidia revealed that certain mobile GPU chipsets, specifically the G84 and G86 chips used in laptops from Apple, Dell, and HP, had abnormally high failure rates due to manufacturing defects that caused the chips to overheat and fail. Nvidia took a charge of approximately one hundred and ninety-six million dollars to cover the costs of replacing the defective chips. Lawsuits followed. Apple MacBook Pro models from May 2007 through September 2008 were affected, along with numerous Dell and HP laptop models. The settlement eventually included a two million dollar reimbursement fund, but the reputational damage was significant. Nvidia maintained that it had done nothing wrong, with the company stating in court filings that it "denies all allegations of wrongdoing." But the episode was a painful reminder of the risks inherent in the semiconductor business.

Despite these challenges, Jensen Huang never wavered from his long-term vision. Throughout the late 2000s and early 2010s, Nvidia continued investing in CUDA, expanding its developer tools, hosting GPU computing conferences, and building relationships with the academic research community. In 2010, Nvidia released the Fermi architecture, which brought double-precision floating-point performance critical for scientific computing. The Kepler architecture followed in 2012, offering dramatically improved power efficiency. Each generation of GPU was better not just at gaming but at the general-purpose computing workloads that CUDA enabled. Nvidia was patiently building a platform, laying track for a train that most people could not yet see coming.

Most Wall Street analysts did not understand why. Nvidia was a gaming company, they said. Why was it spending money on scientific computing? Why was it building products for researchers who could not afford expensive hardware?

The answer came in 2012, and it changed everything.

In September 2012, a team of researchers at the University of Toronto entered the annual ImageNet Large Scale Visual Recognition Challenge. ImageNet was a massive database of labeled images, millions of photographs classified into thousands of categories. Every year, computer vision researchers competed to build software that could correctly identify objects in these images. For decades, the best systems had relied on hand-crafted algorithms designed by human experts, and progress had been incremental. Each year's winner was slightly better than the last, but the improvements were small.

The Toronto team was led by Geoffrey Hinton, a computer science professor who had spent decades working on neural networks, a branch of artificial intelligence that most of the academic establishment had written off as a dead end. Hinton's graduate student, Alex Krizhevsky, had built a deep convolutional neural network, essentially a program inspired by the structure of the human visual cortex, that could learn to recognize objects by being shown millions of examples. The catch was that training this network required enormous computational power. Running the training on a CPU would have taken weeks or months. But Krizhevsky realized that the same parallel processing capabilities that made GPUs great at rendering video games also made them ideal for training neural networks. Neural network training involves repeated matrix multiplications, thousands of them happening simultaneously, and this is exactly what GPUs are designed to do.

Krizhevsky trained his network, which would become known as AlexNet, on two Nvidia GeForce GTX 580 graphics cards. He ran the training on a computer in his bedroom at his parents' house. And when AlexNet competed in the ImageNet challenge, it obliterated the competition. AlexNet achieved a top-five error rate of 15.3 percent, meaning it correctly identified the object in an image among its top five guesses 84.7 percent of the time. The second-place system, built with traditional hand-crafted algorithms, had an error rate of 26.2 percent. It was not even close. AlexNet did not just win. It demolished every other approach.

This was the big bang of the deep learning revolution. Almost overnight, researchers around the world realized that deep neural networks, trained on GPUs, could achieve results that no hand-crafted algorithm could match. And the GPU of choice was Nvidia's. Not because Nvidia had planned it this way, necessarily, but because CUDA existed. Because Nvidia had spent years building a software ecosystem that made it easy for researchers to write and run neural network code on GPUs. No other GPU manufacturer had done this. AMD had not done it. Intel had not done it. Only Nvidia had invested in making its hardware programmable for general-purpose computing, and now that investment was paying off in ways that nobody, not even Jensen Huang, had fully anticipated.

The connection between Nvidia and the AI research community deepened rapidly. Andrew Ng, a Stanford professor who would co-found Google Brain, had been experimenting with GPU-accelerated deep learning since around 2008 and 2009, when his students demonstrated that GPUs could speed up neural network training by ten to a hundred times compared to CPUs. Ng became an evangelist for GPU computing, and his work attracted attention from Google, Facebook, and other technology giants. Deep learning went from an academic curiosity to the hottest area in technology in the space of just a few years.

Jensen Huang recognized the opportunity and committed to it with the same intensity he brought to everything. In 2012, the AI chip market was essentially nonexistent. As Mark Stevens recalled, committing significant research and development resources to an unproven market was the opposite of the risk-averse behavior you would expect from a mature, publicly traded company worth billions of dollars. But Huang operated from first principles. As he put it, "If first principles say you're right about a non-consensus view of how the future will work, bet on it." Nvidia had always been a company that invested in zero billion dollar markets, betting on emerging categories before they existed. The NV1 had been a bet on 3D graphics for PCs when that market was zero. The GeForce 256 had been a bet on dedicated GPU hardware when most rendering was done on the CPU. CUDA had been a bet on general-purpose GPU computing when nobody was asking for it. Each time, the company was early. And each time, when the market finally materialized, Nvidia was in position to dominate it.

In 2015, Nvidia entered the autonomous vehicle market with the launch of its Drive platform. Announced at the Consumer Electronics Show in Las Vegas in January 2015, the Drive CX and Drive PX platforms were designed to bring AI-powered computing to the automotive industry. Drive CX was a digital cockpit computer, handling the screens and displays inside the car. Drive PX was the real prize, an image-processing computer designed to help vehicles perceive and navigate the world around them, the computational brain of a self-driving car. A year later, at CES 2016, Nvidia unveiled the Drive PX 2, which it described as the world's first in-car AI supercomputer. In mid-October 2016, Tesla Motors announced that every new vehicle rolling off its production line would include a Drive PX 2, which would power the neural network processing for Tesla's Enhanced Autopilot and eventual full self-driving capabilities. At CES 2017, Nvidia revealed the Drive PX Xavier, based on its new Volta architecture, and in October 2017, the even more powerful Drive PX Pegasus system, designed for fully autonomous Level 5 driving. The automotive business was small relative to Nvidia's gaming and data center segments, but Huang saw it as another zero billion dollar market that was about to become enormous. By fiscal year 2025, automotive revenue had grown to one point seven billion dollars, up fifty-five percent year over year, vindicating the long-term bet.

Meanwhile, Nvidia's consumer gaming business continued to thrive. The company's GeForce GTX 1080, launched in May 2016 on the Pascal architecture, was widely considered the best gaming GPU ever made at the time of its release. The Pascal architecture represented a major leap in both performance and power efficiency, delivering significantly more performance per watt than the previous Maxwell generation. Gamers loved it. Reviews were rapturous. And Nvidia's market share in the discrete desktop GPU market, which had been locked in a roughly sixty-forty split with AMD for years, began to tilt even further in Nvidia's favor.

In August 2016, Jensen Huang did something that symbolized his commitment to the AI future. Nvidia had developed the DGX-1, what they called the world's first deep learning supercomputer. It packed eight Tesla P100 GPUs into a single box, delivering 170 teraflops of half-precision performance. It was equivalent, Nvidia claimed, to 250 conventional servers, and it cost one hundred and twenty-nine thousand dollars. The first DGX-1 off the production line was not sold. It was donated. Jensen Huang personally carried it to the San Francisco offices of OpenAI, the nonprofit artificial intelligence research lab that had been co-founded by Elon Musk and Sam Altman in December 2015. Huang hand-delivered the supercomputer to the OpenAI team. The inscription on the box read, "To unlock the powers of superhuman capabilities." According to Ilya Sutskever, OpenAI's chief scientist at the time, the DGX-1 accelerated their research experiments by weeks. Much of the work that OpenAI would do on that machine, and on the many Nvidia GPUs that followed, would eventually lay the foundations for ChatGPT and the generative AI revolution.

Then, in May 2017, Jensen Huang took the stage at GTC and unveiled the Volta architecture, specifically the Tesla V100 GPU. Volta was a watershed moment. The V100 contained twenty-one point one billion transistors on a die size of eight hundred and fifteen square millimeters, fabricated on a custom twelve-nanometer process by TSMC. It was, at the time, the largest and most complex chip ever built. But the truly revolutionary feature was not its size. It was a new type of processing unit called the Tensor Core. Tensor Cores were specialized hardware units designed specifically for the kind of matrix multiplication operations that are fundamental to deep learning. Each Tensor Core could perform mixed-precision matrix multiplies, dramatically accelerating the training and inference of neural networks. The V100 packed six hundred and forty Tensor Cores and delivered one hundred and twenty teraflops of deep learning performance, which Nvidia claimed was equivalent to the performance of one hundred CPUs. It was twelve times faster for deep learning training and six times faster for inference than its predecessor, the Pascal-based P100. The Volta architecture was a declaration of intent. Nvidia was no longer just making GPUs that happened to be good for AI. It was designing chips from the ground up with AI as a first-class workload.

The V100 became the workhorse of the AI research community. Universities, national labs, and tech companies snapped them up. Google, Facebook, and Amazon built massive clusters of V100 GPUs in their data centers. The chip powered breakthroughs in natural language processing, computer vision, drug discovery, and climate modeling. Revenue from Nvidia's data center segment, which had been a modest contributor to the company's bottom line, began growing rapidly. The age of GPU computing was no longer a theoretical bet. It was real, and it was accelerating.

But then something unexpected happened on the consumer side. Cryptocurrency mining. In 2017, the price of Bitcoin and Ethereum began skyrocketing. Ethereum, in particular, used a proof-of-work mining algorithm that was well-suited to GPU computation. Miners discovered that Nvidia's GeForce GTX 1060 and GTX 1070 were excellent mining cards, and they began buying them in enormous quantities. The popular GeForce GTX 1060, which had retailed for as low as two hundred dollars, suddenly cost between four hundred and eight hundred dollars, when you could find it at all. Graphics cards disappeared from store shelves. Gamers, the people who had been Nvidia's core customers for two decades, were furious. They could not buy the products they wanted because cryptocurrency miners were hoarding them.

Nvidia was in an awkward position. The crypto miners were driving enormous demand, which meant huge revenues and soaring stock prices. But Nvidia's brand was built on gaming, and gamers were its most loyal constituents. In January 2018, Nvidia publicly stated that it was not happy about the rising GPU prices and advised retailers to prioritize gamers. But the statement rang hollow. The company was clearly benefiting from the crypto boom, even if it was not deliberately courting miners.

In August 2018, even as the crypto boom was beginning to wane, Nvidia made another bold move on the consumer side, launching the Turing architecture and with it the GeForce RTX 20 series. Turing introduced two groundbreaking technologies to consumer graphics. The first was dedicated hardware ray tracing cores, called RT Cores. Ray tracing is a rendering technique that simulates the way light actually behaves in the real world, bouncing off surfaces, refracting through glass, casting accurate shadows and reflections. For decades, ray tracing had been used in Hollywood visual effects and architectural visualization, but it was far too computationally expensive for real-time video games. The RT Cores in Turing changed that, providing dedicated hardware acceleration that made real-time ray tracing possible for the first time in a consumer GPU.

The second innovation was DLSS, which stood for Deep Learning Super Sampling. DLSS used the Tensor Cores embedded in Turing GPUs to run a neural network that could intelligently upscale lower-resolution images to higher resolutions in real time. In simple terms, a game could render at a lower internal resolution, which was faster, and then DLSS would use AI to reconstruct the image at a higher resolution, producing visuals that looked as good as or better than native rendering at a fraction of the computational cost. It was an astonishingly clever application of AI to a consumer product, and it demonstrated how Nvidia's investments in deep learning were feeding back into its core gaming business.

The GeForce RTX 2080 Ti, the flagship of the Turing lineup, launched on September 20, 2018, with a retail price of eleven hundred and ninety-nine dollars, the most expensive consumer GPU Nvidia had ever released. Gamers and reviewers were initially skeptical. The ray tracing features were impressive in demos, but few games supported them at launch. The initial DLSS implementation was limited to just a handful of titles, and the results were sometimes blurry or artifact-laden because the algorithm had to be trained specifically on each game. Critics called it a solution in search of a problem, a technology that was ahead of its time. But Nvidia knew that software would catch up to the hardware. And it did. Over the next two years, dozens of games added ray tracing and DLSS support. DLSS itself improved dramatically through software updates, evolving from a first-generation technology with notable compromises to a genuinely impressive AI upscaling system that could double frame rates with minimal visual quality loss.

And then, as quickly as it had arrived, the first crypto boom collapsed. In the final months of 2018, cryptocurrency prices crashed. Miners stopped buying new GPUs. Worse, they began dumping their used cards on the secondary market, flooding eBay and Reddit forums with cheap, secondhand GeForce cards. Gamers who had been unable to buy new cards now had their pick of cheap used ones, and many chose the secondhand route instead of buying Nvidia's newest RTX products. Nvidia's stock price tanked. Shares fell by more than fifty percent from their peak, and the extent of the company's dependence on crypto-driven demand became painfully clear.

The crypto crash of 2018 was a painful episode, but Nvidia recovered. The Turing architecture, despite its rocky launch, proved the viability of ray tracing and AI-enhanced rendering. And the RTX brand became one of the most powerful product names in gaming, signaling not just speed but intelligence, a GPU that did not just compute faster but computed smarter.

In September 2020, Nvidia launched the Ampere architecture with the GeForce RTX 3080, and this time the reception was ecstatic. The RTX 3080 offered roughly double the performance of the RTX 2080 at the same price point of six hundred and ninety-nine dollars. It was one of the largest generational leaps in GPU history, and reviewers hailed it as the card that finally made ray tracing viable for mainstream gamers. The second-generation RT Cores were nearly twice as fast as the first generation, and DLSS had been significantly refined. The demand was overwhelming. Cards sold out within seconds of going on sale, and for months afterward, they were nearly impossible to buy at retail price. Part of the shortage was due to the pandemic-disrupted supply chain. Part of it was due to a second wave of cryptocurrency mining that hit in 2020 and 2021, as Bitcoin and Ethereum prices surged again.

Nvidia had learned from the first crypto boom. In 2021, it deliberately reduced the mining performance of its new GeForce RTX 3060 cards by fifty percent when used for Ethereum mining, and simultaneously launched a line of dedicated crypto mining processors, or CMPs, specifically designed for miners. The message was: if you want to mine, buy these cards. Leave the GeForce for the gamers. It was a pragmatic solution, though critics noted that Nvidia was still happy to profit from both sides.

But the Ampere era was not just about gaming. On the data center side, Nvidia launched the A100 GPU in May 2020, based on the Ampere architecture. The A100 was the first GPU built on TSMC's seven-nanometer process and featured third-generation Tensor Cores with support for new data formats that dramatically accelerated AI training and inference. The A100 would become the chip on which the generative AI revolution was built. OpenAI trained GPT-3 and GPT-4 on clusters of thousands of A100 GPUs. Google used them. Anthropic used them. Meta used them. Every major AI breakthrough of the early 2020s ran on the A100. It was, perhaps, the single most consequential piece of semiconductor hardware of the decade.

Let us now talk about one of the most strategically important acquisitions in Nvidia's history, one that would prove to be worth many times its purchase price. In March 2019, Nvidia announced that it would acquire Mellanox Technologies for approximately six point nine billion dollars in cash. Mellanox was an Israeli company, founded in 1999, that had pioneered the InfiniBand interconnect technology used to link servers together in data centers and supercomputers. InfiniBand was the high-speed fabric that allowed thousands of GPUs to communicate with each other, and it was essential for the massive AI training clusters that were becoming the backbone of the technology industry. At the time of the acquisition, Mellanox had about two thousand eight hundred employees, mostly in Israel, and was generating approximately one point three billion dollars in annual revenue with sixty-five percent gross margins.

The deal closed on April 27, 2020, after receiving approval from the European Union, the United States, and China's State Administration for Market Regulation. Jensen Huang called it a "homerun deal," and he was not exaggerating. The Mellanox acquisition gave Nvidia something it had never had before: ownership of the networking layer that connects GPUs in data centers. It was not enough to make the best GPUs. You also needed the fastest way to connect them. With Mellanox, Nvidia now owned the entire stack, from the compute chips to the networking fabric. This would prove incredibly valuable as AI training clusters grew from hundreds of GPUs to thousands and then tens of thousands.

China's regulatory approval of the Mellanox deal came with conditions. China's State Administration for Market Regulation required that Nvidia maintain fair competition in markets related to data transmission and networking equipment. Years later, this conditional approval would become a flashpoint, as Chinese regulators accused Nvidia of failing to comply with the conditions, but that controversy was still in the future.

The period from 2019 to 2022 was also marked by one of the most audacious acquisition attempts in the history of the technology industry. On September 13, 2020, Nvidia and SoftBank Group announced that Nvidia would acquire Arm Limited for forty billion dollars. Arm was the British semiconductor design company whose chip architectures powered virtually every smartphone on Earth, along with billions of other devices. Arm did not manufacture chips itself. Instead, it licensed its designs to companies like Apple, Qualcomm, Samsung, and hundreds of others. If Nvidia owned Arm, it would control the fundamental building blocks of mobile computing while simultaneously dominating the GPU market and the data center AI market. It would have been, by far, the largest semiconductor acquisition in history.

But the deal faced immediate and fierce opposition. Companies that depended on Arm's designs, including Google, Microsoft, and Qualcomm, voiced concerns that Nvidia would use its ownership of Arm to favor its own products and disadvantage competitors. Regulatory scrutiny came from every direction. The United States Federal Trade Commission, the European Union, the United Kingdom's Competition and Markets Authority, and China's regulators all launched investigations. On December 2, 2021, the FTC voted to file an administrative complaint to block the deal, arguing that the acquisition would harm competition in three specific markets: processors for advanced driver-assistance systems, data processing unit SmartNICs, and Arm-based CPUs for cloud computing. The regulatory pressure was overwhelming. On February 8, 2022, Nvidia and SoftBank announced that they were terminating the deal. SoftBank kept the one point two five billion dollars that Nvidia had already prepaid, and Nvidia retained its twenty-year Arm license. Jensen Huang responded with characteristic equanimity, telling analysts, "We gave it our best shot."

The failed Arm deal was a rare public defeat for Nvidia, but it did not slow the company down. If anything, it clarified Nvidia's strategic focus. The company would not try to become a broad semiconductor conglomerate. It would continue to do what it did best: build the fastest, most programmable chips in the world and surround them with an unassailable software ecosystem.

And then, in the final weeks of 2022, everything changed. On November 30, 2022, OpenAI released ChatGPT to the public. It was, technically, a chatbot, a text-based interface to a large language model called GPT-3.5. But in practice, it was something far more profound. ChatGPT could write essays, answer questions, compose poetry, debug code, explain scientific concepts, and carry on conversations that felt eerily human. It was not perfect. It made mistakes. It sometimes hallucinated facts. But it was astonishing, and the world reacted accordingly. ChatGPT gained one million users within five days of launch, making it the fastest-growing consumer application in history. Within two months, it had over one hundred million users.

And what was ChatGPT built on? Nvidia GPUs. The model had reportedly been trained using approximately thirty thousand Nvidia A100 data center GPUs. Every major AI lab, OpenAI, Google DeepMind, Anthropic, Meta AI, Mistral, every single one of them was training their models on Nvidia hardware, using Nvidia's CUDA software stack, and running their inference workloads on Nvidia's data center GPUs. The company's decade-long bet on GPU computing, a bet that Wall Street had dismissed, that analysts had questioned, that competitors had ignored, had paid off on a scale that defied comprehension.

The financial results were staggering. In fiscal year 2024, which ended in January 2024, Nvidia reported revenue of sixty point nine billion dollars, roughly double the previous year. But that was just the beginning. In fiscal year 2025, which ended in January 2025, revenue exploded to one hundred and thirty point five billion dollars, an increase of one hundred and fourteen percent year over year. To put this in perspective: Nvidia's revenue more than doubled in a single year. The company's data center segment, powered by sales of AI chips like the A100 and H100, generated one hundred and fifteen point two billion dollars in revenue, an increase of one hundred and forty-two percent from the prior year. Data center revenue now represented approximately eighty-eight percent of Nvidia's total business. Gaming revenue, once the company's core, was eleven point four billion, just eight point seven percent of the total. Professional visualization contributed one point nine billion, and automotive added one point seven billion.

The H100 GPU, based on Nvidia's Hopper architecture, became the most sought-after piece of hardware in the technology industry. The Hopper architecture, named after pioneering computer scientist Grace Hopper, had been announced in March 2022 and represented another generational leap in AI performance. The H100 featured fourth-generation Tensor Cores with a new Transformer Engine specifically optimized for the transformer neural network architecture that powered large language models. It was the chip that trained the large language models powering ChatGPT, Claude, Gemini, Llama, and every other major AI system. The H100 retailed for somewhere between twenty-five thousand and forty thousand dollars per unit, depending on the configuration. At the peak of the shortage in 2023, some H100 cards were selling for more than forty thousand dollars on the secondary market. Estimates suggest that Nvidia sold approximately one point five million H100 units in 2023 alone, generating potentially thirty-seven point five billion dollars in revenue from a single product. Demand was so intense that companies were waiting months for delivery. Cloud providers like Amazon Web Services, Microsoft Azure, and Google Cloud were placing orders of tens of thousands of units at a time, building massive data centers filled with nothing but Nvidia GPUs.

Nvidia's profits were equally extraordinary. In the fourth quarter of fiscal year 2025, the company reported net income of twelve point two nine billion dollars in a single quarter, representing growth of seven hundred and sixty-nine percent from the same quarter a year earlier. The numbers were so large that they seemed unreal, like Monopoly money. But they were very, very real.

The stock market responded accordingly. Nvidia's market capitalization crossed one trillion dollars in May 2023, making it one of the most valuable companies in the world. It crossed two trillion dollars in February 2024, just nine months later. By June 2024, it had crossed three trillion dollars. On June 7, 2024, Nvidia announced a ten-for-one stock split, its sixth split in history, bringing the share price down from over twelve hundred dollars to around one hundred and twenty dollars per share, making it more accessible to retail investors. And the rise continued. In July 2025, Nvidia briefly touched four trillion dollars in market capitalization, becoming the first company in history to reach that milestone. By early 2026, the company's market cap fluctuated around four point three trillion dollars, making it arguably the most valuable company on Earth, trading places with Apple and Microsoft for the top spot depending on the day. And in fiscal year 2026, which ended January 2026, revenue climbed further to two hundred and fifteen point nine billion dollars, up sixty-five percent from the prior year. Fourth-quarter revenue alone was sixty-eight point one billion dollars, a staggering sum for any company, let alone one that had been thirty days from bankruptcy three decades earlier.

Jensen Huang, who had started with forty thousand dollars in a booth at Denny's, was now one of the wealthiest people on the planet. As of early 2026, Forbes estimated his net worth at approximately one hundred and sixty-four billion dollars, making him the seventh-richest person in the world. Nearly all of his wealth came from his approximately three point five percent ownership stake in Nvidia. When asked about a proposed billionaire tax that would cost him roughly eight billion dollars, Huang said he was "perfectly fine" with it.

But the story of Nvidia's ascent is not just a story of financial success. It is a story of ecosystem dominance, a concept that Jensen Huang understood better than almost anyone in the technology industry. NVIDIA's moat is not just hardware. It is software. It is CUDA. It is the fact that millions of developers, researchers, and engineers have learned to program using CUDA, have built their applications on CUDA, have trained their models on CUDA, and would face enormous switching costs if they tried to move to a different platform. As of 2025, CUDA had an installed base of over four million developers worldwide. Every major AI framework, TensorFlow, PyTorch, JAX, was optimized for CUDA first and foremost. The machine learning libraries, the debugging tools, the optimization frameworks, the pre-trained model repositories, the entire software infrastructure of modern artificial intelligence was built on Nvidia's platform. AMD could build a faster chip. Intel could build a cheaper chip. Google could design a custom TPU. But none of them could replicate the ecosystem that Nvidia had spent nearly two decades building.

Jensen Huang understood this deeply. In his characteristically direct way, he explained the strategy to audiences at Nvidia's annual GPU Technology Conference, or GTC. GTC had started as a modest gathering of GPU computing enthusiasts and had grown into one of the most important technology conferences in the world, a kind of Woodstock for AI engineers. Jensen's keynotes, delivered in his signature black leather jacket, black T-shirt, and black jeans, became legendary events. He would stand on stage for two hours or more, no teleprompter, no notes, walking the audience through NVIDIA's latest products and his vision for the future. The leather jacket had become his trademark. He had been wearing it for over two decades, and it had become as associated with him as Steve Jobs' black turtleneck or Mark Zuckerberg's gray T-shirt.

The GTC keynote in March 2024 was particularly significant. Huang took the stage at the San Jose Convention Center and announced the Blackwell computing platform, Nvidia's next-generation GPU architecture. Named after the mathematician David Blackwell, the Blackwell B200 chip contained two hundred and eight billion transistors and delivered five times the AI performance of its predecessor, the H100. Nvidia also announced the GB200 Grace Blackwell Superchip, which paired a Blackwell GPU with an Arm-based Grace CPU, and the DGX B200 system, which delivered three times the training performance and fifteen times the inference performance of the DGX H100. Huang announced advanced networking innovations, including the Quantum-X800 InfiniBand and Spectrum-X800 Ethernet platforms, capable of eight hundred gigabits per second. He unveiled Project GR00T, a foundation model for humanoid robotics, and the Jetson Thor computer designed to power humanoid robots. He talked about computational lithography, digital twins, and climate simulation. The keynote was a masterclass in technological ambition, and it made clear that Nvidia saw itself not just as a chip company but as the computing platform for the entire AI era.

At GTC 2025, Huang went even further, declaring that the world had reached a one trillion dollar computing inflection point. He introduced Blackwell Ultra, an enhanced version of the Blackwell architecture arriving later that year, and previewed the next-generation Vera Rubin platform. He described the concept of "AI factories," massive data centers optimized for producing intelligence rather than just storing and serving data. He spoke about physical AI, the idea of AI systems that understand and interact with the physical world, and announced collaborations with Disney Research and Google DeepMind on Newton, a next-generation physics engine for robotics. He also unveiled the DGX Spark, a desktop-class AI supercomputer powered by Grace Blackwell hardware, designed to bring serious AI development capability to individual researchers and developers. The message was clear: AI was not just a data center phenomenon. It was going to be everywhere.

By November 2024, Morgan Stanley was reporting that the entire 2025 production run of Blackwell chips had already been sold out. As of February 2026, NVIDIA had moved its Blackwell architecture, specifically the B200 GPU and the liquid-cooled GB200 NVL72 rack system, into full-scale volume production.

But not everything was smooth sailing. Nvidia faced growing geopolitical headwinds, particularly around its business in China. Beginning in October 2022, the United States government began imposing export controls on advanced AI chips, seeking to prevent China from acquiring the hardware it needed to develop cutting-edge artificial intelligence and potentially military applications. The initial restrictions targeted Nvidia's most powerful data center GPUs, including the A100 and H100. Nvidia responded by designing a modified chip, the A800, with reduced performance that complied with the export limits. But in October 2023, the Bureau of Industry and Security tightened the rules further, closing the loophole. Nvidia then created the H20, a further de-tuned chip designed to comply with the new, stricter regulations. But even the H20 was eventually banned from export to China.

The impact was significant. China had once accounted for at least one-fifth of Nvidia's data center revenue. Now that business was under threat. Nvidia estimated that the export bans could cost the company approximately five point five billion dollars in revenue. Jensen Huang publicly criticized the policy, arguing that export controls would not slow China's AI development but would instead spur domestic innovation. Chinese chipmakers, including Huawei, were indeed investing heavily in developing their own AI accelerators. China's Cyberspace Administration eventually directed all domestic tech firms to stop purchasing Nvidia chips and rely on local suppliers instead. Nvidia's CEO told CNBC in early 2026 that he was worried about Chinese AI rivals who were innovating rapidly to fill the void left by US export restrictions. The geopolitical dimensions of the AI chip market had become as important as the technical ones.

The geopolitical tensions were compounded by antitrust scrutiny on multiple fronts. In France, antitrust regulators raided Nvidia's offices on September 28, 2023, investigating suspected anticompetitive practices in the graphics card sector. The French Autorité de la Concurrence concluded a market study in June 2024, finding that Nvidia was likely abusing its dominance through price fixing, production restrictions, unfair contractual conditions, and discriminatory behavior. In the United States, the Department of Justice escalated its own antitrust investigation in September 2024, issuing subpoenas to Nvidia. Regulators were concerned that Nvidia was making it difficult for customers to switch to competitors and was penalizing buyers who did not exclusively use Nvidia products. In China, the State Administration for Market Regulation announced in September 2025 that Nvidia had violated antitrust law by failing to comply with the conditions imposed on its 2020 acquisition of Mellanox. Nvidia responded by saying that it "wins on merit" and that customers are free to "choose whatever solution is best for them."

The antitrust pressure reflected a broader concern in the technology industry: Nvidia's dominance of the AI chip market was so complete that it raised systemic risks. As of late 2025, Nvidia held approximately eighty-six percent of the AI accelerator market by revenue. Its share of AI training GPUs exceeded ninety percent. AMD, its largest merchant competitor, held only about five to eight percent of the market. Intel's share of discrete AI accelerators had fallen below one percent, though Intel retained a larger share of the broader data center market when CPU-based inference was included. Google, Amazon, and Microsoft were all developing custom AI chips, Google's TPUs, Amazon's Trainium and Inferentia, and Microsoft's Maia, but none of these posed an immediate threat to Nvidia's dominance in training workloads. The concern was not that any one competitor would overtake Nvidia. The concern was that no one could.

This concentration of market power was both Nvidia's greatest strength and its greatest vulnerability. The software ecosystem lock-in that CUDA provided was a nearly impenetrable moat, but it also made Nvidia a target for regulators who worried about single points of failure in the AI supply chain. If Nvidia's chips had a defect, if TSMC's fabrication facilities were disrupted, if export controls were tightened further, the entire AI industry could be affected.

And that brings us to the relationship with TSMC. Nvidia is a fabless semiconductor company, meaning it designs chips but does not manufacture them. The actual fabrication of Nvidia's GPUs is done by Taiwan Semiconductor Manufacturing Company, or TSMC, the world's most advanced chip foundry. TSMC has been the sole manufacturer of Nvidia's AI GPUs, and the relationship between the two companies is one of the most important in the global technology supply chain. Jensen Huang has said that he promised TSMC's founder, Morris Chang, that Nvidia would become TSMC's largest customer. As of 2025, Nvidia was indeed one of TSMC's most important clients, consuming a significant share of the foundry's most advanced manufacturing capacity.

The dependence on TSMC created a geopolitical vulnerability. TSMC's most advanced fabrication facilities were located in Taiwan, an island that China claims as its own territory. Any military conflict in the Taiwan Strait could disrupt the production of Nvidia's most important products. To mitigate this risk, TSMC began building advanced fabrication facilities in the United States, including a major complex in Phoenix, Arizona. In 2025, Jensen Huang visited TSMC's Phoenix facility to celebrate the production of the first NVIDIA Blackwell wafer manufactured on American soil. It was a symbolic moment, representing the beginning of a long and expensive effort to reduce the semiconductor industry's dependence on a single geographic location.

Now let us talk about Nvidia's culture, because to understand the company, you have to understand the man who runs it, and Jensen Huang is one of the most unusual CEOs in the technology industry. His management style is unlike anything you will find in a business school textbook. As of 2024, he had approximately sixty direct reports. For most CEOs, managing even ten direct reports would be considered a lot. Sixty is unheard of. Huang does not hold regular one-on-one meetings with his reports. Instead, he communicates through a combination of group meetings, email, and what employees describe as spontaneous deep-dives into specific projects. He is deeply involved in the day-to-day operations of the company, often reaching down several layers into the organization to ask questions or provide direction. Some employees call this micromanagement. Huang calls it staying close to the work.

He is famously allergic to hierarchy. Nvidia's organizational structure is deliberately flat, with as few layers as possible between frontline engineers and the CEO. Huang believes that hierarchy creates silos, and silos slow down innovation. He wants information to flow freely, without being filtered through layers of middle management. This makes Nvidia a demanding place to work. The expectations are high, the pace is relentless, and there is no place to hide. But it also creates an organization that can move faster than competitors that are many times its size.

Nvidia's physical environment reflects its ambitions. The company's headquarters in Santa Clara, California, consists of two buildings: Endeavor, a five hundred thousand square foot structure completed in 2017, and Voyager, a seven hundred and fifty thousand square foot addition completed in 2022. Together, the two buildings encompass one point two five million square feet and accommodate nearly six thousand employees. Designed by architecture firm Gensler, the campus was conceived to answer a fundamental Silicon Valley question: how do you get thousands of people to collaborate as if they were still working at a startup? Endeavor features a distinctive triangular floor plate inspired by Nvidia's logo. Voyager has a sloped glass curtain wall that floods the interior with natural light, and the two buildings are connected by a tree-lined walkway shaded by a trellis of photovoltaic panels. The solar trellis generates 390 kilowatts of electricity. Both buildings were built to LEED Gold certification standards, incorporating biophilic design elements that blur the boundary between inside and outside. It is a campus designed for intensity and creativity, a physical manifestation of Jensen Huang's belief that great work requires a great environment.

Huang's work ethic is legendary, even by Silicon Valley standards. He told Stripe CEO Patrick Collison that he is either working or thinking about work every waking moment, and that he works seven days a week, including holidays. In a December 2025 interview with Fortune, Huang admitted that he operates in a constant "state of anxiety" driven by a fear of going bankrupt, the same fear that has haunted him since 1996, when the company nearly went under. "I'm afraid of going out of business every single day," he said. This anxiety is not performative. It is the fuel that drives his intensity.

His personal style has become iconic. The black leather jacket, worn at every public appearance for more than twenty years, is now as recognizable as Nvidia's green logo. When Nvidia's stock price hit one hundred dollars per share, Huang made good on a semi-joking promise to his employees and got the Nvidia logo tattooed on his upper left arm. He went to the tattoo parlor with his family. His kids watched as their father went through the process. "I was crying like a baby," Huang later recalled. "My kids were with me, and they were like, 'Dad, you've gotta control yourself.'" The tattoo became one of the most well-known pieces of body art in corporate America. In 2023, when asked if he would get another tattoo to celebrate a new milestone, Huang demurred. "Aging gracefully," he said. "No more tattoos. It hurts way more than anybody tells you."

Huang's philosophy on leadership and success has been remarkably consistent over the decades. He believes that great companies are built through suffering and struggle, not through easy victories. He has said, "If you want to build something great, it's not easy. You have to suffer, you have to struggle, you have to endeavor. And there are no such things that are great that are easy to do." He has told employees and students that "resilience matters in success" and that suffering is "part of the journey." He encourages employees to maintain the same sense of urgency that the company felt when it was thirty days from bankruptcy, even as its market capitalization approaches five trillion dollars. To this day, he uses the phrase "pain and suffering" inside the company "with great glee," as he once told an interviewer.

Perhaps most revealing is what Huang said during an interview with the Acquired podcast in October 2023. When asked what kind of company he would start if he could go back to being thirty years old, Huang's response was immediate: "I wouldn't do it." He continued, "Building Nvidia turned out to have been a million times harder than I expected it to be, than any of us expected it to be." And then he added, "If I take all of my knowledge now, and I go back, and I say, I'm going to endure that whole journey again, I think it's too much. It is just too much." This was not false modesty from a man who had built one of the most valuable companies in history. It was a genuine expression of how difficult the journey had been. And yet Huang remained at the helm, driving the company forward, unable to stop because the opportunities ahead were too vast to ignore.

The other co-founders took different paths. Curtis Priem served as Nvidia's chief technical officer from the company's founding in 1993 until his retirement in 2003. After a decade of building the company's technological foundations, Priem stepped back from the day-to-day operations. And then he made a decision that, in hindsight, looks like one of the most expensive choices in business history. Between 2003 and 2006, Priem sold all of his Nvidia shares. As of November 2023, Forbes estimated Priem's net worth at approximately thirty million dollars. If he had retained his shares, Forbes estimated he would have been worth approximately seventy billion dollars. A February 2026 Fortune article calculated that Priem's original stake could have been worth as much as six hundred billion dollars at Nvidia's peak. But Priem has never expressed regret. He took the money and dedicated his life to philanthropy, focusing especially on his alma mater, Rensselaer Polytechnic Institute. Since 2001, Priem has donated more than two hundred and seventy-five million dollars to RPI. In 2004, he gave forty million dollars, the largest unrestricted gift in the university's history, which funded the construction of the Curtis R. Priem Experimental Media and Performing Arts Center. In 2023, his foundation made its largest gift ever, seventy-five million dollars, to deploy the first IBM Quantum System One computer on a university campus. As of 2024, Priem serves as vice chair of RPI's board of trustees. He lives largely off the grid, far from Silicon Valley, content with a life dedicated to science, education, and quiet generosity.

Chris Malachowsky, the third co-founder, took a different path still. He stayed at Nvidia. For more than three decades, Malachowsky has served as a fellow and senior vice president of engineering and operations. He remains at the company, deeply involved in its technical direction, a steadying presence in an organization that moves at breakneck speed. He is less visible than Jensen Huang, by design. He does not give keynote speeches or pose for magazine covers. But his contributions to Nvidia's chip architecture and engineering methodology have been immense, and his longevity at the company is a testament to his belief in its mission.

The leadership team around Jensen Huang has been remarkably stable by Silicon Valley standards. Colette Kress joined as chief financial officer in September 2013, bringing nearly twenty-five years of finance experience from Cisco and Microsoft, where she had served as CFO of the Server and Tools division. Debora Shoquist joined as senior vice president of operations in 2007 and was promoted to executive vice president two years later, overseeing the company's IT, supply chain, manufacturing, logistics, and quality management systems. Jay Puri serves as executive vice president of worldwide field operations, managing global sales, and Tim Teter serves as executive vice president, general counsel, and secretary. This stability at the top has given Nvidia a continuity of vision and execution that many technology companies lack.

Now let me take you through the broader landscape that Nvidia is operating in as of 2026, because the company's dominance exists within a complex web of competitors, customers, and geopolitical forces.

On the competitive front, AMD remains Nvidia's most direct rival. Under the leadership of CEO Lisa Su, who took over in 2014, AMD has undergone its own remarkable transformation. AMD's MI300X GPU, launched in late 2023, offered more high-bandwidth memory than the Nvidia H100, with 192 gigabytes of HBM3 memory compared to the H100's 80 gigabytes. Some AI workloads that are memory-limited run better on AMD's hardware. But AMD has a fundamental disadvantage: software. AMD's ROCm platform, its answer to CUDA, has historically been less mature, less well-documented, and less widely supported by the AI research community. AMD has been working furiously to close this gap, but decades of accumulated software investment are not easy to replicate.

Intel, once the undisputed king of the semiconductor industry, has struggled mightily in the AI era. Despite spending billions on its Gaudi and Ponte Vecchio accelerators, Intel holds less than one percent of the discrete AI accelerator market. The company's foundry business has been plagued by manufacturing delays and quality issues, and its data center CPU business, once a reliable cash machine, has come under pressure from ARM-based alternatives. Intel's story is a cautionary tale about what happens when a dominant company fails to see and respond to a paradigm shift.

The hyperscale cloud providers, Amazon, Google, Microsoft, and Meta, are both Nvidia's biggest customers and its potential future competitors. Each of them is developing custom silicon designed to reduce their dependence on Nvidia. Google's Tensor Processing Units, or TPUs, have been used to train some of the company's most important AI models, including Gemini. Amazon's Trainium chips are designed for AI training workloads in AWS. Microsoft is developing Maia, a custom AI chip for its Azure cloud. Meta is building its own Training and Inference Accelerator, or MTIA. But none of these custom chips have made a significant dent in Nvidia's market share, at least not yet. The switching costs are enormous. Porting an AI training pipeline from CUDA to a custom chip framework requires months of engineering work, and the performance may not be equivalent. For most AI companies, especially smaller ones, NVIDIA remains the default choice, and default choices in technology tend to be very sticky.

Let us also discuss Nvidia's forays into gaming hardware beyond traditional GPUs. In 2013, Nvidia launched the Shield portable gaming device, powered by its Tegra 4 processor. It was followed by the Shield Tablet in 2014 and the Shield TV in 2015. These products were commercially modest but technologically interesting, showcasing Nvidia's ability to build complete gaming systems rather than just components. The Shield TV, in particular, found a niche as a high-quality streaming device and Android gaming console.

But the real payoff for Nvidia's Tegra line came in 2016, when Nintendo chose a customized Tegra X1 chip to power its hybrid gaming console, the Switch. On October 20, 2016, Nvidia announced the partnership, revealing that the two companies had invested over five hundred engineer-years of collaborative work to develop the custom processor. The Nintendo Switch launched in March 2017 and became one of the most successful gaming consoles in history, eventually selling over one hundred and forty million units worldwide. For Nvidia, which had been shut out of the console market since the Xbox, the Switch partnership was a validation of its mobile chip technology. And the relationship continued. When Nintendo developed the Switch 2, it turned to Nvidia again, this time using a custom chip based on the more powerful Tegra T239 processor with AI and ray tracing capabilities.

Nvidia also pushed into the enterprise software market with Omniverse, a platform for building and simulating digital twins and 3D collaborative environments. Launched for enterprise use in 2021, Omniverse uses Universal Scene Description, or USD, a 3D file format originally developed by Pixar, to allow designers, engineers, and architects to collaborate on complex 3D projects in real time. Companies like BMW used Omniverse to simulate factory operations. Ericsson used it to model 5G network coverage in digital replicas of real cities. The platform was part of Nvidia's broader strategy to expand beyond chip sales and into higher-margin software and platform businesses. At GTC 2025, Nvidia introduced Cosmos, a world foundation model for physical AI that uses Omniverse-based digital twins to create optimized workflows for robotics and industrial automation. As Huang put it, "Just as large language models revolutionized generative and agentic AI, Cosmos world foundation models are a breakthrough for physical AI."

The AI boom also created enormous demand for Nvidia's networking products, the legacy of the Mellanox acquisition. As AI training clusters grew to include tens of thousands of GPUs, the networking fabric connecting those GPUs became a critical bottleneck. Nvidia's InfiniBand technology, inherited from Mellanox, was the fastest interconnect available, and cloud providers were willing to pay premium prices for it. Nvidia's networking business grew to generate over ten billion dollars in quarterly revenue by late 2025, a staggering figure for a product line that had cost just seven billion dollars to acquire.

Looking ahead, Nvidia's technology roadmap continues to be aggressive. At CES 2026, Jensen Huang announced the Vera Rubin platform, Nvidia's next-generation architecture succeeding Blackwell. Named after the astronomer Vera Rubin, who provided key evidence for the existence of dark matter, the platform combines a new Vera CPU with two Rubin GPUs on a single chip. The Rubin NVL144 system promises three point six exaflops of dense FP4 compute, compared to one point one exaflops for the Blackwell B300 NVL72, representing a three point three times improvement. Rubin introduces HBM4 memory, upgrading from HBM3e, with bandwidth increasing from eight terabytes per second to thirteen terabytes per second. Nvidia says Rubin will enable a four times reduction in the number of GPUs needed to train the same model compared to Blackwell, and a ten times reduction in inference token costs. The Rubin R200 GPU began shipping in the second quarter of 2026, well ahead of schedule, with the more powerful Rubin Ultra expected by the second quarter of 2027. Beyond Rubin, Nvidia has added "Feynman" to its roadmap, signaling at least two more generations of architecture improvements. The pace is relentless. Each generation arrives roughly a year after the last, and each generation offers a multiple of the previous generation's performance. It is Moore's Law on steroids, driven not by the physics of transistor scaling but by the sheer architectural innovation of Nvidia's engineering teams.

The relentless cadence of innovation reflects a company that shows no signs of slowing down. But it also raises questions about sustainability. Can Nvidia maintain its dominance as customers develop alternatives? Can it navigate the geopolitical minefield of US-China relations? Can it avoid the antitrust hammer that is swinging with increasing force in both the United States and Europe? Can it continue to grow at this pace when it is already generating over two hundred billion dollars in annual revenue?

Jensen Huang, predictably, believes the answer is yes. He has argued that the transition from traditional computing to accelerated computing, from CPUs to GPUs, is a one trillion dollar opportunity that is still in its early stages. He believes that every data center in the world will eventually need to be rebuilt around GPU-accelerated architecture. He believes that AI will transform every industry, from healthcare to transportation to manufacturing to entertainment, and that Nvidia's hardware and software platform will be the foundation for all of it. He has described Nvidia's vision as building "AI factories," massive computational facilities that consume data and produce intelligence, the way power plants consume fuel and produce electricity.

There is something almost evangelical about Huang's conviction, the way he talks about the future with absolute certainty. But his track record supports his confidence. Again and again, Jensen Huang has bet on markets that did not yet exist and been proven right. 3D graphics for PCs was a zero dollar market when Nvidia was founded. GPU computing was a zero dollar market when CUDA launched. AI chips were a zero dollar market when Nvidia committed to deep learning. In each case, Huang saw the future before anyone else, invested accordingly, and was in position to dominate when the market finally materialized.

The Nvidia story also has a profoundly human dimension. Behind the market caps and the revenue numbers and the chip architectures are real people making decisions under uncertainty. There is the young Jensen Huang, nine years old, cleaning toilets in a Kentucky reform school, learning that the world is not fair but that you can survive it. There is the thirty-year-old Huang, sitting in a Denny's booth, drawing chip architectures on napkins, telling his co-founders that the future belongs to 3D graphics. There is the Huang of 1996, laying off half his workforce, knowing that the next chip has to work or the company dies. There is the Huang of 2006, investing in CUDA when nobody understood why. There is the Huang of 2016, personally carrying a supercomputer into OpenAI's office, betting that artificial intelligence would become the most important technology in human history. And there is the Huang of 2026, standing on stage at GTC in his leather jacket, sixty-three years old, presiding over the most valuable company on Earth, working seven days a week, driven by the same anxiety that has fueled him since the beginning.

In an era when many technology CEOs are interchangeable, when corporate leadership feels like a revolving door of MBAs and professional managers, Jensen Huang stands out as something different. He is a founder who has led his company for more than thirty years. He is an engineer who still understands the details of chip design. He is a strategist who sees further ahead than most. And he is a human being who has never forgotten what it feels like to be close to failure. "I'm afraid of going out of business every single day," he told Fortune in late 2025. From a man sitting atop a four trillion dollar empire, that statement tells you everything you need to know.

The story of Nvidia is not over. It may, in fact, be just beginning. The AI revolution that the company helped create is still in its early stages. Large language models are getting larger and smarter. Autonomous vehicles are getting closer to reality. Humanoid robots are moving from science fiction to early prototypes. Drug discovery is being accelerated by AI simulations. Climate models are being run on GPU clusters. Every one of these applications requires more compute, more GPUs, more of the chips that Nvidia designs and TSMC manufactures. The total addressable market is not just big. It is, if you believe Jensen Huang, effectively limitless.

But markets change. Technologies shift. Competitors adapt. Governments intervene. History is littered with companies that once seemed invincible and then were not. IBM dominated computing for decades and then missed the PC revolution. Intel dominated processors for decades and then missed the mobile revolution. Nokia dominated phones and then missed the smartphone revolution. Will Nvidia be different? Will the company that went from a booth at Denny's to the most valuable enterprise on Earth continue to grow, or will some unforeseen shift, some new architecture, some new competitor, some new technology, eventually unseat it?

Jensen Huang, characteristically, is not worried about what might go wrong. He is worried about what might not go right. He is worried about not moving fast enough, not investing boldly enough, not seeing the next zero billion dollar market before someone else does. He is worried about complacency, about the creeping institutional lethargy that afflicts every successful organization. And so he keeps the thirty-day-from-bankruptcy mantra alive, keeps pushing his sixty direct reports to move faster, keeps standing on stage in his leather jacket, telling the world about the future he sees.

At that Denny's on Berryessa Road in East San Jose, there is now a plaque commemorating the spot where Nvidia was founded. The restaurant still serves the same bottomless coffee. The vinyl booths are still there. But the company that was conceived in those booths has grown into something that the three founders could never have imagined. From forty thousand dollars in startup capital to four trillion dollars in market value. From three engineers drawing chip diagrams on napkins to forty-four thousand employees across the globe. From the NV1, a chip that nearly bankrupted the company, to the Blackwell and Rubin architectures that power the artificial intelligence revolution.

And through it all, one man has been at the helm. Jensen Huang, the dishwasher from Denny's, the table tennis champion from Oregon, the chip designer from AMD, the CEO who told Stanford students that he wished them pain and suffering. He built Nvidia by seeing the future before anyone else, by investing in zero dollar markets, by surviving when survival seemed impossible, and by never, ever believing that the company was safe.

Because it is not. It is thirty days from bankruptcy. It always has been. And that, perhaps, is the secret to everything.

Let me tell you a bit more about what makes the Nvidia ecosystem so powerful, because the hardware story, as dramatic as it is, only tells half the tale. The real source of Nvidia's dominance, the thing that keeps customers locked in and competitors locked out, is software. And not just any software. An entire galaxy of tools, libraries, frameworks, and platforms that make Nvidia's hardware not just fast but easy to use.

CUDA, the Compute Unified Device Architecture, is the foundation. When Ian Buck and John Nickolls created it in 2006, they made a decision that would prove to be one of the most consequential in computing history. They decided that CUDA would use standard C++, the language that most computer scientists and engineers already knew. This meant that a researcher who wanted to accelerate their code on a GPU did not have to learn a new language. They just had to learn a few CUDA-specific extensions and libraries. The barrier to entry was low, and the performance improvement was enormous. A neural network training job that might take a week on a CPU could finish in hours on a GPU. A fluid dynamics simulation that would take a month could be done in a day.

On top of CUDA, Nvidia built a tower of specialized libraries. cuDNN for deep learning operations. cuBLAS for linear algebra. TensorRT for inference optimization. NCCL for multi-GPU communication. RAPIDS for data science. Triton Inference Server for deploying AI models at scale. Each library was meticulously optimized for Nvidia's hardware, squeezing out every last drop of performance. And each library became another link in the chain that bound developers to Nvidia's ecosystem.

The result was a flywheel effect. More developers used CUDA, which meant more software was optimized for Nvidia GPUs, which made Nvidia GPUs more valuable, which attracted more customers, which justified more investment in CUDA, which attracted more developers. This flywheel has been spinning for nearly two decades, and it has built up so much momentum that it is nearly impossible to stop. When AMD releases a new GPU that matches Nvidia's raw performance in benchmarks, it does not matter if the software ecosystem is not there. When Intel launches a new data center accelerator, it does not matter if TensorFlow and PyTorch are not optimized for it. Performance on paper is irrelevant if the software does not exist to harness that performance in practice.

Jensen Huang has sometimes described Nvidia as a "full-stack" company, meaning it provides everything from the silicon to the system software to the application frameworks. This is unusual in the semiconductor industry, where most companies specialize in either hardware or software. But Huang understood early on that the line between hardware and software was blurring, and that the companies that controlled both layers would have the greatest power. Apple understood this with the iPhone. Nvidia understood it with the GPU.

The full-stack approach extends beyond CUDA. Nvidia's DGX systems are turnkey AI supercomputers, complete hardware and software solutions that customers can plug in and start using immediately. The DGX-1, which Jensen hand-delivered to OpenAI in 2016, was followed by the DGX-2, the DGX A100, the DGX H100, and the DGX B200. Each generation offered dramatically more performance, and each came with a complete software stack optimized for AI training and inference. For AI startups and research labs that did not have the engineering resources to build their own infrastructure from scratch, the DGX systems were a godsend. They traded capital expenditure for speed, buying a complete solution rather than trying to assemble one from parts.

Nvidia also recognized that the networking layer was as important as the compute layer. When you train a large language model on tens of thousands of GPUs, the speed at which those GPUs can communicate with each other becomes the binding constraint. If the networking is slow, the GPUs spend most of their time waiting for data rather than computing. The Mellanox acquisition in 2020 gave Nvidia ownership of InfiniBand, the fastest data center networking technology, and the company has invested heavily in making its networking products work seamlessly with its GPUs. The NVLink interconnect, which connects GPUs within a single server, and the NVSwitch, which connects GPUs across multiple servers, are Nvidia-proprietary technologies that offer significantly higher bandwidth and lower latency than anything a third party can provide. This means that a data center built entirely with Nvidia components, GPU, networking, and software, will perform better than one that mixes Nvidia GPUs with third-party networking. The incentive to go all-Nvidia is powerful, and it is by design.

This brings us to the concept of the "Nvidia tax" that some in the industry have begun to complain about. Because Nvidia controls so much of the AI infrastructure stack, it can charge premium prices. The H100 GPU, which probably costs Nvidia a few thousand dollars to manufacture, sells for twenty-five to forty thousand dollars. Gross margins in Nvidia's data center segment exceed seventy percent, meaning that for every dollar of data center revenue, roughly seventy cents is profit before operating expenses. These are extraordinary margins for a hardware company, more typical of a software company or a luxury brand. And they are a direct consequence of Nvidia's monopoly-like position. Customers pay the premium because the alternative, spending months porting their code to a different platform with uncertain performance, is even more expensive.

Some customers have pushed back. The hyperscale cloud providers, in particular, have accelerated their custom silicon efforts partly out of a desire to reduce their dependence on Nvidia and its pricing. Google's TPUs, for example, are designed to run Google's proprietary AI models at a lower cost per computation than Nvidia GPUs. Amazon's Trainium chips serve a similar purpose for AWS customers. But these custom chips are available only within their respective cloud platforms. They are not sold on the open market. They do not have the broad software ecosystem that CUDA provides. And they often require significant engineering effort to use. For the vast majority of AI developers, Nvidia remains the path of least resistance.

The financial implications of this market position are extraordinary. Nvidia's revenue per employee, a common measure of workforce productivity, is among the highest of any company in the world. With approximately forty-four thousand employees generating over two hundred billion dollars in annual revenue, each Nvidia employee generates roughly five million dollars in revenue per year. For comparison, Apple generates about two point four million per employee, and Google generates about one point eight million. Nvidia's employees are, by this measure, among the most productive workers in corporate history.

Now let us step back and consider the broader significance of what Nvidia has built, because the company's impact extends far beyond its own financial statements. Nvidia's GPUs are the physical substrate on which the AI revolution is being built. Every time you ask ChatGPT a question, every time you use Google's AI-powered search, every time an autonomous vehicle navigates a city street, every time a pharmaceutical company uses AI to design a new drug molecule, Nvidia hardware is almost certainly involved. The company has become a foundational layer of the global economy, as essential in its own way as the electrical grid or the internet.

This is not an exaggeration. When Jensen Huang says that the transition from CPU-centric computing to GPU-accelerated computing is a trillion dollar opportunity, he is pointing to a genuine paradigm shift. For fifty years, the computing industry was organized around the central processing unit, a chip designed to handle a wide variety of tasks sequentially. The CPU was a generalist, good at everything, great at nothing. The GPU, by contrast, is a specialist, designed to handle thousands of parallel calculations simultaneously. For workloads that can be parallelized, and AI workloads are among the most parallelizable in computing, GPUs are not just faster than CPUs. They are orders of magnitude faster. The transition from CPUs to GPUs for AI workloads is analogous to the transition from horse-drawn carriages to automobiles. It is not a marginal improvement. It is a categorical change.

And Nvidia is the company that enabled this change. Not single-handedly, of course. The researchers who developed deep learning algorithms deserve enormous credit. The companies that collected and labeled the training data deserve credit. The cloud providers who built the data centers deserve credit. But Nvidia provided the computational foundation without which none of it would have been possible. As Jensen Huang told an audience at GTC, "We didn't just ride the AI wave. We created the surfboard."

But every wave eventually crests. And the question that investors, analysts, competitors, and policymakers are all asking is: what happens next? The bears argue that Nvidia's current growth rate is unsustainable, that AI spending will eventually plateau, that customers will diversify away from Nvidia, and that the regulatory environment will become increasingly hostile. The bulls argue that AI is still in its infancy, that the total addressable market is far larger than current spending levels suggest, that Nvidia's software moat is impregnable, and that the company's execution machine, led by the most capable CEO in the technology industry, will continue to outperform.

The truth is probably somewhere in between. Nvidia's market share will likely decline from its current levels as competitors improve their products and customers diversify their supply chains. The regulatory environment will likely become more challenging, particularly in Europe and China. And there is always the risk of a genuine technological disruption, a new computing paradigm that makes GPUs obsolete, though no such paradigm is visible on the horizon as of 2026.

But even if Nvidia's market share declines from ninety percent to sixty percent, even if its revenue growth slows from triple digits to double digits, the company will remain one of the most important and valuable enterprises in the world. It has built something that is very difficult to replicate: a complete computing platform, from hardware to software to networking, optimized for the most important workload of the twenty-first century. It has a thirty-year head start in GPU design. It has a twenty-year head start in GPU software. It has a monopoly-like position in the fastest-growing segment of the technology industry. And it has Jensen Huang, a founder-CEO who has been at the helm since day one and shows no signs of stepping down.

When Nvidia was founded in 1993, the personal computer was still a novelty for many people. The internet was a curiosity known mainly to academics and government researchers. Smartphones did not exist. Social media did not exist. Cloud computing did not exist. Artificial intelligence was a field in its third decade of unfulfilled promises. Into this world walked three engineers from Silicon Valley who believed that graphics would change computing. They were right, but not in the way they expected. They thought they were building chips for video games. They ended up building the computational foundation for the age of artificial intelligence.

That is the story of Nvidia. It is a story about vision and timing, about betting on zero dollar markets and being patient enough to wait for them to materialize. It is a story about near-death experiences and last-chance chips, about laying off half your workforce and still believing you can win. It is a story about a CEO who works seven days a week, who wears the same leather jacket to every event, who got his company's logo tattooed on his arm, and who tells Stanford students that he wishes them pain and suffering. It is a story about CUDA and ChatGPT, about the RIVA 128 and the H100, about a booth at Denny's and a four trillion dollar empire.

And it is a story that is still being written. Every day, in data centers around the world, thousands of Nvidia GPUs are training the next generation of AI models, models that will be smarter, faster, and more capable than anything we have seen before. Every day, engineers at Nvidia are designing the next generation of chips, pushing the limits of what silicon can do. Every day, Jensen Huang is thinking about the next zero billion dollar market, the next big bet, the next opportunity that others cannot yet see.

Because that is what Nvidia does. It sees the future before anyone else. And then it builds the hardware to make that future real.

Thirty days from bankruptcy. Always. That is the Nvidia way.

...more

Share Corrp.com | Company Histories

Sign up to save your podcasts

Corrp.com | Company Histories

FAQs about Corrp.com | Company Histories:

How many episodes does Corrp.com | Company Histories have?

Corrp.com | Company Histories episodes:

FAQs about Corrp.com | Company Histories:

How many episodes does Corrp.com | Company Histories have?