One other day, one other new AI massive language mannequin that’s supposedly higher than all earlier ones. Once I started penning this story, Elon Musk’s xAI had simply launched Grok 3, which the corporate says performs higher than its rivals towards a variety of benchmarks. As I used to be revising the article, Anthropic launched Claude 3.7 Sonnet, which it says outperforms Grok 3. And by the point you learn this, who is aware of? Perhaps a completely new LLM could have appeared. In January, in any case, the AI world was briefly rocked by the discharge of a low-cost, high-performance LLM from China known as DeepSeek-R1. A month later, folks had been already questioning when DeepSeek-R2 would come out.
The competitors amongst LLMs could also be laborious to maintain monitor of, however for Nvidia, the corporate that designs the pc chips—or graphics-processing items (GPUs)—that many of those massive language fashions have been educated on, it’s additionally enormously profitable. Nvidia, which, as of this writing, is the third-most-valuable firm on this planet (after Apple and Microsoft), was began three many years in the past by engineers who needed to make graphics playing cards for avid gamers. The way it advanced into the corporate that’s offering virtually all of the picks and shovels for the AI gold rush is the story on the core of Stephen Witt’s The Considering Machine. Framed as a biography of Jensen Huang, the one CEO Nvidia has ever had, the e-book can be one thing extra attention-grabbing and revealing: a window onto the mental, cultural, and financial ecosystem that has led to the emergence of superpowerful AI.
That ecosystem’s heart, after all, is Silicon Valley, the place Huang has spent most of his grownup life. He was born in Taiwan, the son of a chemical engineer and a trainer. The household moved to Thailand when he was 5, and some years later, his dad and mom despatched him and his older brother to the US to flee political unrest. Ultimately, his dad and mom relocated to the U.S. as properly, and Huang grew up within the suburbs of Portland, Oregon. Within the early Eighties, after majoring in electrical engineering at Oregon State (which on the time didn’t provide a computer-science main), he acquired a job at Superior Micro Gadgets. The corporate—then the poor cousin of the chip big Intel—was headquartered in Sunnyvale, California, close to US 101, the freeway that runs from San Jose to Stanford. Since then, Huang’s profession has unfolded inside a five-mile radius of that workplace.
Huang quickly left AMD for a agency known as LSI Logic Company, which constructed software-design instruments for chip architects, after which left LSI in 1993 to start out Nvidia with the chip designers Curtis Priem and Chris Malachowsky: He was proper on the right track “to run one thing by the age of thirty,” as he’d informed them he aimed to do. The corporate was coming into a crowded market for growing graphics playing cards, the pc {hardware} that’s used to render photographs and movies. Nvidia didn’t have an actual marketing strategy, however Huang’s boss at LSI really useful him to Sequoia Capital. One of many Valley’s most necessary venture-capital companies, Sequoia helped the corporate get off the bottom.
The graphics-card enterprise was constructed on a perpetual improve cycle that compelled builders right into a unending recreation of efficiency enchancment: An organization was solely pretty much as good as its final card. At numerous factors in these early years, Nvidia was one misstep away from chapter, and its unofficial motto grew to become “Our firm is thirty days from going out of enterprise.”
One will get the impression that Huang preferred it that method. He says his coronary heart fee goes down below strain, and to name him a relentless employee is to understate issues. “I ought to be sure that I’m sufficiently exhausted from working that nobody can preserve me up at evening,” he as soon as mentioned. His studying food plan options enterprise books (which he devours). He has no apparent politics (or at the very least by no means discusses them). He’s not a gaudy philanthropist. Although dedicated to his household, he’s additionally sincere: “Lori,” he says of his spouse, “did ninety p.c of the parenting” of their two kids. For the previous 30 years, his life has clearly revolved round Nvidia.
Huang’s reluctance to speak about himself makes him a difficult topic for Witt to deliver to life. However Nvidia’s workers, who virtually all discuss with Huang by his first identify, are effusive. They “worship him—I consider they’d comply with him out of the window of a skyscraper if he noticed a market alternative there,” Witt writes. He later provides that they see Huang “not simply as a pacesetter however as a prophet. Jensen was a prophet who made predictions about issues. After which these issues got here true.” He has a ferocious mood—referred to within the firm as “the Wrath of Huang”—and is infamous for publicly reprimanding, at size, staff who’ve made errors or didn’t ship. However he not often fires folks and, actually, conjures up intense devotion. One in all his key subordinates says, “I’ve been afraid of Jensen generally. However I additionally know that he loves me.”
Huang’s best energy as a CEO has been his willingness to make huge, dangerous bets when alternatives current themselves. The primary of these got here when he modified the structure of Nvidia’s chips from serial processing to parallel processing. Witt calls this transfer “a radical gamble,” as a result of as much as that time, no firm had been capable of make promoting parallel-processing chips a viable enterprise.
Serial computing is the way in which your laptop’s central processing unit works: It executes one instruction at a time, very, very quick. Witt likens it to telling one supply van to drop off packages in sequence. In contrast, “Nvidia’s parallel GPU acts extra like a fleet of bikes spreading out throughout a metropolis,” with the drivers delivering every package deal at roughly the identical time. The coding required to make parallel processing work was far more advanced, however in the event you may do it, you had entry to huge quantities of computing energy.
Initially, all that energy was used primarily to make laptop video games look and carry out higher. However then Huang took one other huge danger, remaking Nvidia’s GPUs in order that they might additionally course of huge knowledge units, of the sort scientists would possibly use. As one Nvidia govt places it, “You’ve gotten a online game card on one aspect, however it has a change on it. So that you flick that change, and switch the cardboard over, and abruptly the cardboard turns into a supercomputer.”
The fascinating factor about this choice was that Huang didn’t know who would possibly wish to purchase a supercomputer within the guise of a graphics card, or what number of such folks had been on the market. He was simply betting that in the event you make highly effective instruments accessible to folks, they may discover a use for them, and at a scale to justify the billions in funding.
That use—and it was huge—turned out to be synthetic intelligence, specifically neural-network expertise. As Witt notes, simply as parallel processing was revolutionizing computing, an analogous revolution was occurring in AI analysis—although nobody at Nvidia was taking note of it. AI had gone by a collection of boom-and-bust cycles as researchers tried completely different strategies, all of which finally failed. A type of strategies was neural networks, which tried to imitate the human mind and permit the AI to evolve new guidelines of studying by itself. If you practice these networks on huge databases of photographs and textual content, they’ll, over time, establish patterns and change into smarter. Neural networks had lengthy been peripheral, partly as a result of they’re black bins (you’ll be able to’t clarify how the AI is studying, or why it’s doing what it’s doing), and partly as a result of the computing energy required to make a high-performance neural community function was out of attain.
Parallel-processing GPUs modified all that. Instantly, AI researchers, if they might write software program properly sufficient to get probably the most out of the chips, had entry to ample computing energy to permit neural networks to evolve at a unprecedented tempo. In 2009, Geoff Hinton, one of many godfathers of AI analysis, informed a convention of machine-learning specialists to go purchase Nvidia playing cards. And in 2012, one among Hinton’s college students, Alex Krizhevsky, strung collectively two Nvidia GPUs and constructed and educated SuperVision (which he later renamed AlexNet). It was an AI mannequin that would, for the primary time, establish photographs with startling accuracy, largely as a result of, in Witt’s phrases, “the GPU produced in half a minute what would have taken an Intel machine an hour and what would have taken biology 100 thousand years.”
Huang didn’t instantly acknowledge the significance of what had occurred. When he spoke at Nvidia’s annual GPU Expertise Convention in 2013, he by no means talked about neural networks, speaking as an alternative about climate modeling and laptop graphics. However a couple of months later, after an Nvidia researcher named Bryan Catanzaro made a direct pitch to him in regards to the significance of AI, Huang had what Witt calls a “Damascene epiphany”: He positioned one other huge guess, basically remodeling Nvidia from a graphics firm into an AI firm over the course of a weekend. This guess was much less dangerous than his earlier ones, as a result of regardless that Nvidia had rivals who additionally constructed GPUs, none of them had actually designed theirs for use as supercomputers. Nonetheless, going all in was prescient—developments comparable to massive language fashions had but to take off—and is what has turned Nvidia into an almost $3 trillion firm.
That weekend feels as if it had been the compressed end result of Nvidia’s story, which isn’t empirically true. The 12 years that adopted have been extremely eventful, and extremely worthwhile, as the corporate has saved enhancing its chips, servicing the insatiable urge for food for computing energy created by the emergence of LLMs, and keeping off rivals (lots of whom are Nvidia’s prospects, now constructing their very own chips). However the foundations for that pivot, and all that ensued, had been already in place when Huang determined to behave on his AI perception.
These foundations, The Considering Machine makes clear, weren’t laid by Nvidia alone. Certainly, amongst Witt’s key contributions is to point out that Nvidia’s success can’t be understood aside from the tradition and economic system of Silicon Valley (and of tech extra usually). Take the straightforward truth of free labor markets. One catalyst of the Valley’s success, because the scholar AnnaLee Saxenian has famously argued, was a freewheeling, risk-taking tradition that inspired staff to depart corporations for rivals or to start out their very own companies. And that depended, partially, on the truth that noncompete clauses had been unenforceable in California. Nvidia’s historical past exemplifies this: not simply Huang’s mobility, however that of his early hires as properly. Later, one of many firm’s favourite ways was to poach its rivals’ greatest engineers and coders—dangerous type, maybe, however a very good enterprise tactic.
Nvidia additionally benefited from the analysis investments made by the federal government and universities. One of many essential breakthroughs in unlocking the ability of parallel computing, as an illustration, was an open-source programming language known as Brook, which a gamer and Stanford graduate pupil named Ian Buck developed with a gaggle of researchers in 2003, counting on a Protection Division grant. Alex Krizhevsky and his companion Ilya Sutskever (who later helped begin OpenAI) had been grad college students on the College of Toronto when Krizhevsky devised AlexNet. The competition during which the mannequin demonstrated its accuracy, the ImageNet problem, was designed by a Stanford laptop scientist named Fei-Fei Li. And as that lineup demonstrates (Krizhevsky and Sutskever had been born within the Soviet Union, Li in China), immigration has been central to the historical past of not simply Nvidia however AI usually.
Sensible financial options of the ecosystem mattered as properly. An important was the rise of unbiased chip foundries: factories that serve many various corporations and make chips on order. Nvidia’s partnership with Taiwan Semiconductor Manufacturing Firm, the best-known of those factories, allowed it to change into a dominant participant by specializing in designing and writing software program for its chips; Nvidia didn’t must spend money on precise manufacturing, which might have required prohibitive quantities of capital.
Lastly, Nvidia benefited from persistence, and its board’s willingness to place long-term considering forward of short-term income. Due to the gaming market, Nvidia was virtually all the time a worthwhile firm, however its inventory value dropped almost 90 p.c two completely different instances; it didn’t admire for a full 10 years after the dot-com bubble burst, whereas the corporate was spending billions turning its graphics playing cards into supercomputers. One acquainted indictment of American capitalism is that it’s too short-term-focused. Within the tech trade, at the very least, the trajectory of Nvidia (and plenty of different corporations) means that’s a bum rap.
To make sure, Huang himself was central to Nvidia’s success: He has run the corporate basically on his personal (as Witt places it, he has had “no right-hand man or girl, no majordomo, no second-in-command”), and he’s made the daring strikes. What’s extra, he appears to have completed so and not using a hint of doubt. A number of folks within the AI trade—together with the folks coaching LLMs—have raised considerations about AI’s risks, however Huang is just not one among them. For him, Witt writes, “AI is a pure drive for progress.” Huang doesn’t fret that it could eat all of our jobs, or change artists, or go rogue and determine to wipe out humanity.
In actual fact, when Witt, stricken with existential nervousness about how AI will change the world, asks Huang whether or not a few of these considerations could be price pondering, he’s subjected to one among his legendary tirades:
“Is it going to destroy jobs?” Huang requested, his voice crescendoing with anger. “Are calculators going to destroy math? That dialog is so previous, and I’m so, so uninterested in it,” he mentioned. “I don’t wish to discuss it anymore … We make the marginal price of issues zero, era after era after era, and this precise dialog occurs each single time!”
You could possibly write this off for example of Upton Sinclair’s adage “It’s tough to get a person to grasp one thing, when his wage relies upon upon his not understanding it!” However the truth that Huang talks about AI by way of its affect on “marginal prices” shouldn’t be lowered to mere opportunism: It matches proper in with the single-minded concentrate on efficiency that has pushed him from Nvidia’s starting. Witt at one level calls Huang a “visionary inventor.” The imaginative and prescient Huang has been in thrall to, although, appears to be much less about grand future targets, and extra about instruments—about making the quickest, strongest chips as effectively as attainable. “Existential danger” has no place in that imaginative and prescient. Huang’s unapologetic stance on AI is bracing in its method, particularly in distinction with the general public hand-wringing of many AI chieftains, fretting in regards to the risks of their LLMs whereas persevering with to develop them. However he’s in impact making the most important, riskiest guess ever—not only for Nvidia, however for all of us. Let’s hope he’s proper.
This text seems within the Could 2025 print version with the headline “The New King of Tech.”
If you purchase a e-book utilizing a hyperlink on this web page, we obtain a fee. Thanks for supporting The Atlantic.
