In a Silicon Valley culture known for “grindset” founders, Jensen Huang still manages to stand out. The Nvidia chief executive told Stripe CEO Patrick Collison earlier this year that he is either working, or thinking about work, every waking moment—and that he works seven days a week.
“If you want to build something great, it’s not easy. You have to suffer, you have to struggle, you have to endeavor,” Huang said. “And there are no such things that are great, that are easy to do.”
Well, no one doubts Huang has built something great. Under his leadership, Nvidia has positioned itself at the heart of the artificial intelligence boom. Its graphics processing units (GPUs), specialized for training and running the most powerful AI models, dominate that market, accounting for the overwhelming majority of GPUs sold into data centers in 2023. Nvidia’s share price has increased more than sevenfold since OpenAI’s ChatGPT debuted in November 2022, and the company is now among the most highly valued in the world, with a market capitalization of $3.4 trillion.
Demand for Nvidia’s most advanced GPU systems routinely outstrips supply—the entire 2025 production of its most advanced Blackwell chip is, according to a report from Morgan Stanley, already sold out. Elon Musk and Oracle founder Larry Ellison took Huang out for dinner at Nobu in Palo Alto to personally lobby him for larger allocations of his GPU production. Such hunger helps explain why Nvidia’s revenues for the current fiscal year—2025—are estimated to be $125 billion, more than double last year’s figure, which itself was more than double 2023’s tally. And its operating profit margin is north of 60%.
It’s not just Fortune 500 CEOs who are eager to meet with Huang. The White House has sought his views on AI, and he’s consulted with world leaders including Indian Prime Minister Narendra Modi and the UAE’s Sheikh Mohammed bin Zayed. The U.S. sees Nvidia’s leading edge in GPUs for AI as a key national security asset, and the Biden administration has restricted the sale of its more advanced chips to China—a move that might have been more damaging to Nvidia’s prospects if it hadn’t been seeing such explosive demand everywhere else.
Yet it is far from certain that Nvidia will be able to hold on its market position as it faces new threats, not just from its old competitor AMD, but a host of well-funded new startups eager to grab a slice of the AI computing market, as well as from the internal AI chip efforts of the large cloud computing companies that are also its best customers. Despite its success, Huang himself remains acutely aware that Nvidia’s leadership position could prove fleeting—which may explain his relentless work ethic. “I do everything I can not to go out of business,” he told a magazine reporter last year. “I do everything I can not to fail.”
From Denny’s to dominance
It was a long and uncertain path that brought Huang to such heights. Born in Taiwan, he came to the U.S. as a child and went on to earn degrees in electrical engineering from Oregon State University and Stanford. He worked on software and chip design for LSI Logic and AMD before leaving to cofound Nvidia in 1993.
At the time, the Santa Clara, Calif.–based startup was one of dozens springing up to build specialized graphics cards—they weren’t yet called GPUs—to enable computers to run video games faster. Nvidia was also among a new generation of “fabless” semiconductor companies—it designed the computer chips it sold, but it contracted out their manufacturing to foundries owned by others. Over the next three decades, Nvidia and its rival AMD emerged to dominate that market.
In the mid-2000s, artificial-intelligence researchers realized that GPUs could help them train and run large artificial neural networks—a kind of AI loosely based on how the human brain works—much more efficiently than conventional chips. Training large neural networks requires a chip to perform many of the same kinds of calculations millions or billions of times. Standard computer chips, called central processing units, or CPUs, can only perform one calculation at a time. GPUs, on the other hand, can perform many similar calculations in parallel, vastly accelerating the time it takes to run AI models. Huang presciently recognized the importance of this market and began promoting Nvidia’s chips specifically to AI researchers and engineers.
The key to Nvidia’s success, however, has been more than just designing ever faster and more powerful GPUs. The company has long taken a “full stack” approach: It designs not just chips, but also the software to run them and the cabling to connect them. In 2007, it introduced CUDA (Compute Unified Device Architecture), an open-source software programming language that helped coders run AI applications on GPUs. And it invested heavily in promoting CUDA and training engineers to use it. Today there are an estimated 5 million CUDA developers around the world. Their familiarity with CUDA has been a powerful factor in preventing rival AI-chip companies, which have mostly underinvested in creating similar software and developer communities, from challenging Nvidia’s dominance.
In 2019, Nvidia bought Israeli networking and switching company Mellanox for $7 billion. The deal gave Nvidia the technology to help its customers build giant clusters of tens or hundreds of thousands of GPUs, optimized for training the largest AI models. And Nvidia has continued to move up the stack, too—building its own AI models and tools, in an effort to encourage businesses to use generative AI. In 2023, Nvidia announced that it would, for the first time, begin offering its own AI cloud computing services directly to corporate customers, in a move that puts it in direct competition with the giant cloud “hyperscalers,” such as Microsoft, Google, and Amazon’s AWS, that are among its best customers.
“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.”
Jensen Huang, CEO and founder, Nvidia
Huang has fashioned himself as a rock-star founder-CEO, complete with a signature uniform of black leather jacket, black T-shirt, and black jeans. But unlike many of his tech peers, he comes off as self-deprecating and folksy in interviews. He jokes about cleaning toilets as a teenage busboy at Denny’s, and about his habit of getting up at 5 a.m. but then reading in bed until 6 a.m. because he feels guilty waking up his dogs too early. He admitted recently on a podcast with Rene Haas, the CEO of chip design company Arm and a former Nvidia employee, that he didn’t have any particular secret to hiring good people. “We’re not always successful, look how you turned out,” Huang ribbed Haas. “It’s always a shot in the dark.”
Huang’s humility and down-to-earth persona have made him an effective salesperson for Nvidia’s GPUs, and allowed him to build critical partnerships with top executives at companies such as OpenAI and Microsoft, as well as networking equipment makers like Broadcom.
It has also helped him maintain an unconventional management culture—particularly for a company that employs more than 30,000 people. Huang has 60 direct reports and is known, as Haas put it delicately, for “reaching down into different layers of the organization” (or, to put it less delicately, micromanaging). This flat structure can make Nvidia a tough place to work, but Huang sees it as critical to ensuring the organization is strategically aligned and nimble enough to stay at the cutting edge of rapidly evolving chip development and AI progress.
Huang says he is allergic to hierarchy and corporate silos. He doesn’t believe in one-on-one meetings. Instead, he prefers mass gatherings of his leadership team: He says all Nvidia execs should be able to learn from the feedback he provides to any one of them, and they should all benefit from watching him together as he puzzles through a problem.
Looming challenges to Nvidia’s top-dog status
For all his sometimes folksy charm, lately Huang has begun sounding increasingly prophetic and utopian. In his public comments, he has posited that the world is witnessing a new industrial revolution in which “AI factories” transform data and electricity into “intelligence tokens”—and in which there’s a fundamental shift in computing, with GPUs gaining at the expense of CPUs.
To keep the competition at bay, Nvidia has upped the tempo at which it is rolling out new generations of top-of-the-line GPUs, going from releasing a new model every other year to an annual release schedule. It is also buying out capacity at TSMC’s foundries, which manufacture all of Nvidia’s chips, to try to prevent competitors from being able to use TSMC’s facilities to produce rival products. It launched a software tool called Nvidia Inference Microservices (NIMs) that makes it easier for developers to set up and run existing AI models on cloud-based Nvidia GPUs without having to know as much about CUDA.
Some investors believe Nvidia will live up to Huang’s vision of making the GPU the essential hardware unit of all computing—and justify its outsize market cap. Bank of America’s equity analysts recently put forth a bullish scenario based on the tens of billions of dollars that Big Tech companies from Microsoft to Meta to Apple have announced they will invest in computing infrastructure over the next several years. The analysts noted that these purchases could translate into significantly higher growth for Nvidia’s data-center networking solutions, and they pointed out that TSMC has seemingly overcome production issues that had limited initial shipments of the Blackwell chip. They put a price target of Nvidia’s stock of $190 per share, 30% above its current record high.
Others are less sanguine. Businesses have struggled to figure out how to derive value from generative AI. Indeed, technology analytics firm Gartner says AI is entering what it calls “the trough of disillusionment”—in which people realize a much-hyped technology cannot live up to inflated expectations and drastically pare back spending on it. And while George Brocklehurst, a Gartner research vice president, says he expects this downturn to be short-lived, ending in 2027, it wouldn’t bode well for Nvidia’s revenues or stock price in the interim. At the same time, Brocklehurst says he expects AMD to begin to eat into Nvidia’s market share for data center GPUs and the hyperscale cloud companies to continue to invest in their own alternative to Nvidia’s chips. AMD has forecast it will sell $5 billion worth of such GPUs this year. “That is a reasonable toe in the door,” he says, and indicates that for the right price and performance, developers are willing to move away from CUDA. He says the market is increasingly intolerant of Nvidia’s near-monopoly position and the control and pricing power that gives it. (Nvidia says that it controls far less of the market for AI chips than critics charge when one also looks at the competition from chips that the hyperscale cloud companies, such as Google and AWS, design and produce themselves for their own data centers.)
Nvidia has also optimized its GPUs for training the largest, most powerful AI models. But when it comes to running applications on already trained AI models—which is called inference—there are indications that a number of new kinds of chips, including those from upstart AI-chip companies such as startups Groq and Etched, as well as offerings from AMD, and possibly even Intel, can match or outperform Nvidia’s GPUs at a lower cost. In addition, smaller AI models—which could even be run on devices like laptops or smartphones equipped with AI accelerators built by companies such as Qualcomm—may come to be the primary engines for many AI use cases. Nvidia currently doesn’t have significant hardware offerings in that market.
Another challenge: China and geopolitics. The Biden administration slapped export controls on the most sophisticated of Nvidia’s chips, preventing their shipment to China. Nvidia has created a “de-featured” version of its powerful H100 GPU, called the H20, which falls just below the thresholds of these export restrictions, and has proved popular in China—so much so that Chinese AI companies have become adept at using H20s, as well as their own homegrown GPUs from companies such as Huawei, to train AI models that are, by many measures, just as capable as those trained on Nvidia’s more potent GPUs.
These techniques for ringing performance out of ostensibly less capable chips may ultimately help companies elsewhere avoid having to pay top dollar to use Nvidia’s highest-end products. Moreover, an incoming Trump administration is likely to further ratchet up restrictions on chip sales to China, potentially hurting Nvidia’s sales. (China currently accounts for about 12% of Nvidia revenue.) And, of course, there’s always a risk of China moving against Taiwan militarily, which would disrupt Nvidia’s supply chain, which is heavily dependent on TSMC’s semiconductor foundries on the island.
Right now, Huang is pursuing the only sensible strategy available given Nvidia’s position and inflated investor expectations, says Alvin Nguyen, a senior analyst at Forrester Research. “He’s circling the wagons and trying to create this fear of missing out (on AI) among potential customers,” he says. But the strategy may only work for a little while. Long-term, given all the factors arrayed against it, Nguyen says, “I’d be very surprised if they were able to keep their dominance.”
Of course, competitors have bet against Huang before and been proven wrong. But the chips on the other side of the table have never been piled quite this high.
This article appears in the December 2024/January 2025 issue of Fortune.
