Jensen's Decade

The stock that would become the most consequential equity in the world closed at $112.27 on October 14, 2022. Adjusted for splits that had not yet happened, the share price had fallen about sixty percent from its November 2021 peak. Nvidia’s market capitalization that day was a little under three hundred billion dollars, which made it the eighth largest American technology company by valuation, smaller than Tesla, smaller than Berkshire Hathaway. Two miserable forces had converged on the company. Ethereum, which had absorbed an enormous quantity of GeForce inventory through 2021, had completed its move to proof-of-stake in September and pushed the world’s GPU mining rigs into liquidation overnight. The same week, the United States Bureau of Industry and Security had published the October 7 export controls that wiped out Nvidia’s plans for selling A100 and H100 GPUs into Chinese data centers. The cryptocurrency cycle had ended, and the largest single foreign market had been closed by federal regulation. Inside Santa Clara, the projection deck for fiscal 2023 had been redrawn three times in two months. The data-center business was still growing, but no one outside the company had yet decided that growth was enough to matter.

Six weeks later, on November 30, 2022, OpenAI released ChatGPT. The chatbot had been trained on a cluster of roughly ten thousand Nvidia A100 GPUs, leased from Microsoft Azure under the partnership the two companies had announced three years earlier. Within five days the service had a million users. Within two months it had a hundred million. Every cloud-provider executive in Seattle, Redmond, and Mountain View who had been dragging on capital expenditure now had a clear directive from a board: order more H100s, and order them now. By February 22, 2023, when Jensen Huang stepped through Nvidia’s fourth-quarter fiscal 2023 earnings call, the company’s data-center revenue had grown eleven percent year over year to $3.62 billion, a number that looked respectable in isolation but did not yet describe what was happening in the order book. Huang told the analysts on the line that the industry had reached an inflection point, that a trillion dollars of installed data-center infrastructure was about to be rebuilt around accelerated computing, and that demand for the company’s chips was outrunning supply. The transcript reads, in retrospect, like a man trying to be heard over the noise of an avalanche he could already see.

The fiscal year that followed was one of the strangest financial events in modern American corporate history. In May 2023, on the first quarter call, Nvidia guided to eleven billion dollars of revenue for the next quarter, more than fifty percent above what Wall Street had penciled in. The stock jumped twenty-four percent in a single session, adding more than two hundred billion dollars of market capitalization in an afternoon. Three months later the eleven billion came in at thirteen and a half. By the time the full fiscal 2024 closed in January 2024, total revenue was $60.9 billion against $27 billion the year before, and the data-center segment alone had grown from $15 billion to $47.5 billion, a 217 percent jump. By the end of calendar 2024, total fiscal-2025 revenue was $130.5 billion, with $115.2 billion of that coming from the data center. Three years earlier, the data-center business had not yet existed at this scale; now it was the largest enterprise-hardware franchise in the world by a margin no competitor was within a generation of closing.

The men in Hsinchu, Seoul, and Taipei could see the trajectory in their fab schedules and their HBM stack orders before the analysts in New York could see it on a chart. Morris Chang’s foundry was building every flagship Nvidia design on its custom 4N node, a variant of TSMC’s 5-nanometer process tuned specifically for the H100 die. SK Hynix in Icheon, which had pulled ahead at HBM3 in a way that Samsung had not anticipated, was selling Nvidia every working stack it could pull off the line. Foxconn’s server arm in Taiwan, along with Wistron, Quanta, and Inventec, was assembling the DGX systems and the integrated server racks that hyperscalers were buying by the thousand. By the end of 2023, Taiwanese vendors were shipping more than seventy percent of the world’s Nvidia GPU servers, and the cumulative effect on Taiwan’s export numbers was visible at the macroeconomic level. A line item that had not been a category five years earlier was now driving the trade balance of an island.

The company that benefited from this convergence was a fabless GPU designer that had spent two decades preparing, accidentally and then deliberately, for exactly this moment.

The accidental preparation had begun in November 2006, when Nvidia released the G80 architecture inside a graphics card called the GeForce 8800 GTX. The chip was a five-hundred-eighty-one-million-transistor monster designed to render Microsoft’s then-new DirectX 10 graphics API, but its underlying organization was a quiet rebellion against the standard GPU layout of the era. Where every previous graphics chip had separated vertex shaders from pixel shaders into two specialized pipelines, the G80 unified them into a single pool of one hundred twenty-eight identical floating-point cores that the hardware scheduler could dispatch to either workload. This was a choice driven by graphics rendering, and it solved a graphics problem. It also turned the GPU, accidentally, into a general-purpose parallel processor.

The deliberate preparation came alongside it. A Stanford computer-graphics graduate student named Ian Buck had spent the early 2000s writing a programming language called Brook that let researchers describe parallel computations and run them on graphics cards. In 2004, Nvidia hired Buck and paired him with a senior architect, John Nickolls, to design a programming model that could exploit the unified shader hardware. The platform they produced, released as a public software development kit on February 15, 2007, was called Compute Unified Device Architecture, abbreviated CUDA. It was a parallel-computing toolchain, free, available to any developer who owned a GeForce card, with all the supporting compilers, libraries, and debuggers that Nvidia could afford to write.

CUDA was not a profitable product. It was, by every Wall Street test of the period, a tax that Nvidia chose to pay. Every consumer GeForce GPU shipped with the silicon area and the software infrastructure required to run general-purpose compute workloads, even though the gamers buying those cards had no use for general-purpose compute. Industry analysts at the time estimated that Nvidia spent something on the order of half a billion dollars a year, every year, on engineering effort that produced no consumer-graphics revenue. By the time the project had been running for a decade, the cumulative R&D outlay attributable to CUDA was in the low double-digit billions. The board kept asking when the bet was going to pay off. Huang kept telling them it would. He would later describe CUDA, with characteristic bluntness, as the moat the company spent fifteen years digging before anyone realized there was a war.

What CUDA produced over those fifteen years was not just a programming language. It was a vertical software stack of the kind that Microsoft had once built around Windows. There was cuBLAS for dense linear algebra, cuDNN for deep-neural-network primitives, cuFFT for spectral analysis, NCCL for multi-GPU collective communication, NVCC for compilation, Nsight for debugging and profiling. PyTorch, TensorFlow, JAX, and every other framework that mattered for machine learning compiled their tensor operations to CUDA kernels first and ported to other platforms second, if at all. By 2024, the active developer base writing CUDA code numbered in the millions. AMD’s ROCm, the most credible alternative, had a fraction of the install base, no parity at the framework level, and a reputation among practitioners for instability that cost it the benefit of the doubt every time. The hardware competition could in principle produce GPUs with comparable peak teraflops at lower prices. The software competition was unable to match a stack that had been built incrementally over almost two decades by a company that had decided very early that the hardest part of selling parallel compute was making it easy to use.

By the time the AI boom arrived, the moat was not a slogan; it was a measurable cost of switching, in engineering hours, that no plausible AMD or hyperscaler-ASIC team could absorb on a multiyear training run.

The man who had insisted on the bet wore, on the day Blackwell was unveiled, a black leather jacket from a designer he could not have afforded as a young engineer. Huang, by then sixty-one, was the only founder among the original 1993 Denny’s cofounders still running the company, and the cleaning-toilets story he told from the GTC stage every year was the only piece of his Tainan-to-Tacoma-to-Oregon biography he allowed himself to recycle in public. The cinematic effect was not accidental. It was useful. He had spent a decade refining a public posture that did not fit any of the available CEO templates, neither the visionary monk in the turtleneck nor the defensive midwestern manufacturer nor the libertarian provocateur. He was a Taiwanese-American immigrant who pulled a chip out of the pocket of a leather jacket, told an audience of fifteen thousand engineers that the universe required bigger GPUs, and made the joke land. The jacket, which his wife and daughter had effectively chosen for him and which he wore with such consistency that it had become a trademark, communicated a precise message: the man inside had founded a graphics-card company in 1993, and he had not stopped being that engineer thirty years later.

The Graphics Technology Conference, GTC, had begun in 2009 as a small developer event in San Jose and had grown, year by year, into something closer to a cultural function. By March 2024, when Huang took the stage at the SAP Center in San Jose to introduce the Blackwell architecture, more than twelve thousand people were physically in the building and tens of thousands more were watching the live stream. The arena, normally home to a National Hockey League franchise, had been retrofitted into a keynote theater. Huang walked out, said the line that became the keynote’s headline, and then physically extracted a Blackwell chip from his jacket pocket and held it next to a Hopper for size. The audience laughed. The audience also understood. Blackwell was a two-die package, with a total of two hundred and eight billion transistors fabricated on TSMC’s 4NP node, joined by a ten-terabyte-per-second NVLink chip-to-chip interconnect. The package presented as a single GPU to software, achieved roughly two and a half times the FP8 training throughput of an H100 per chip, and produced, when assembled into the GB200 superchip with a Grace ARM CPU and bound into a seventy-two-GPU NVL72 rack, the first commercially available exaflop-class AI system in a single mechanical enclosure. The rack weighed three thousand pounds, drew a hundred and twenty kilowatts under load, required direct liquid cooling because air cooling could not absorb the thermal output, and delivered more than one and a quarter exaflops of FP4 inference performance. Hyperscalers had ordered them by the thousand before the keynote ended.

The path from Hopper to Blackwell to Rubin was the visible armature of the company’s strategy, but the architecture every two years was no longer the differentiator. The DGX systems and NVL72 racks were system-level products. They depended on the high-bandwidth memory that South Korea was suddenly the only country in the world able to make in volume. They depended on the InfiniBand and Ethernet switching fabric that Nvidia had acquired in 2019 by writing a $6.9-billion check for an Israeli company called Mellanox, an acquisition that closed in April 2020 and that, with the benefit of hindsight, looked like one of the better-timed corporate purchases of the decade. Mellanox had been the company that owned InfiniBand, the ultra-low-latency interconnect technology that bound supercomputer clusters into a single coherent fabric, and the NVLink Switch System that knit seventy-two Blackwell GPUs into a single shared-memory domain at one hundred thirty terabytes per second of bisection bandwidth was a direct descendant of Mellanox technology. Without the 2020 acquisition, the GB200 NVL72 would not have existed. With it, Nvidia could sell not just the GPU but the network and the rack and the management software, and could capture, across the system, gross margins north of seventy percent on workloads where the customer’s alternative was to design its own custom silicon and its own networking from scratch.

This was the AI factory doctrine. Huang articulated it in a sequence of keynotes through 2024, beginning at the World Government Summit in Dubai in February and continuing through the GTC keynote in March and the Computex keynote in Taipei in early June. The argument was that the data center was no longer a place where general-purpose computers ran a heterogeneous mix of workloads. The data center was now a factory, and its product was tokens. Raw materials, electricity and data, came in at one end. Tokens of generated text, image, code, and embedding came out at the other. Every nation, Huang argued in Dubai, would need to build sovereign AI infrastructure, because the tokens encoded the country’s history and culture and language and could not be safely outsourced to a foreign provider. Every enterprise, Huang argued in San Jose, would need to think of itself as running an AI factory rather than buying software. The pitch was unsubtle, and it was unsubtle on purpose: every customer, public and private, large and small, in every country that was not directly under U.S. export control, ought to be ordering Nvidia systems by the rack. By mid-2024, Saudi Arabia, the United Arab Emirates, France, India, Japan, South Korea, and a handful of other countries had announced sovereign AI initiatives that all converged on the same shopping list.

The Computex keynote on June 2, 2024, at the National Taiwan University Sports Center in Taipei, was where the cultural dimension of Huang’s role became unmistakable. The audience that morning was made up of engineers, executives, and journalists from across the Taiwanese supply chain, the men and women who actually ran the fabs and the ODMs and the assembly lines that turned Nvidia’s designs into hardware. They had, for thirty years, been the second-tier press subjects of an American semiconductor industry whose marquee figures were Andy Grove and Steve Jobs and Elon Musk. Now, on a stage in their own city, the most important customer in the world was a man who had been born up the coast in Tainan. Local reporters had begun calling the phenomenon Jensanity, after the Linsanity that had attached itself a decade earlier to the basketball player Jeremy Lin. Crowds followed Huang through Taipei restaurants. Vendors at the night market asked him to autograph baseball caps. The Taiwan that had absorbed his elementary-school years and produced his foundry partner could now claim a cultural figure whose stock chart, for the moment, was the most important number in global capitalism.

The stock chart had crossed three trillion dollars on June 5, 2024, three days after the Computex keynote. By June 18, Nvidia’s market capitalization had reached $3.34 trillion, briefly making it the most valuable public company on earth, ahead of both Microsoft and Apple. Eleven days earlier, on June 7, the company had executed a ten-for-one forward stock split, which had nothing to do with valuation but everything to do with the optics of an equity whose pre-split price had crossed a thousand dollars and was making retail brokerage screens uncomfortable. The split-adjusted price at the moment Nvidia became the most valuable company in the world was about $135 a share, against the $108 split-adjusted low of October 2022. In twenty-one months, an investor who had bought at the trough had multiplied their position by roughly fourteen.

Through all of this, the political environment around Nvidia was deteriorating in a specific and structural way. The October 2022 export controls had cut the A100 and H100 out of China; the A800 and H800 variants Nvidia engineered to slip beneath the bandwidth threshold sold for almost exactly twelve months before the October 2023 update closed the loophole; the H20 inference part that followed could legally enter China and was ordered by ByteDance and Alibaba in the hundreds of thousands, but the company’s China revenue, which had peaked in fiscal 2022 at roughly $7 billion, was a fraction of what underlying demand would have supported.

Huang’s response was to become, by necessity, one of the most active corporate diplomats in the world. He met American senators on Capitol Hill, the prime ministers of Vietnam and Japan, and the crown prince of Saudi Arabia. After Donald Trump won the 2024 election, Huang met repeatedly with the incoming and then sitting administration, lobbying to keep some level of Chinese sales open and arguing in public that excessively tight controls would simply hand the Chinese AI hardware market to Huawei. He attended a state dinner at Mar-a-Lago in early 2025 and a White House reception in July, on the day Nvidia became the first company to cross a four-trillion-dollar market capitalization. He met Xi Jinping’s ministers in Beijing on a series of trips through 2024 and 2025, arriving each time in the same uniform he wore at GTC. The geopolitical theater Andy Grove and Morris Chang had played in their own ways thirty and forty years earlier was now being run by a man whose product was a system rather than a chip, whose installed base in 2025 ran every leading frontier-AI lab in the world.

The Blackwell ramp was not entirely smooth. In the second half of 2024, Nvidia disclosed that the platform had encountered a yield problem rooted in the package design. The CoWoS-L advanced packaging that bound the two reticle-sized GPU dies, the HBM stacks, the silicon interposer, and the substrate had run into a coefficient-of-thermal-expansion mismatch that produced warping during assembly and intermittent system failures. Nvidia respun the chip’s top metal layers and a portion of the packaging masks at TSMC, accepted a roughly three-month delay on volume shipments, and absorbed the financial cost rather than pass the schedule slip onto customers. Huang took the public blame at the company’s earnings call in late August 2024, said the defect was Nvidia’s responsibility and not TSMC’s, and promised that the ramp would resume in the fourth quarter. It did. By February 2025, when the company reported its fourth-quarter fiscal-2025 numbers, Blackwell had contributed roughly eleven billion dollars of revenue in a single quarter, the fastest product ramp in the history of semiconductors. Three months earlier, at Computex 2024, Huang had already announced the successor architecture, named Rubin after the astronomer Vera Rubin, slated for 2026, with a Rubin Ultra in 2027 and a roadmap that committed Nvidia, for the first time in its history, to an annual cadence of new data-center platforms rather than the historical two-year cycle.

The annual cadence was a statement of intent. The only way to maintain the moat, in Huang’s view, was to keep moving it, on the bet that the AI build-out had at least another decade to run and the capital to fund it was effectively bottomless. Whether the assumption would hold was another question. The DeepSeek release in January 2025, which would briefly knock six hundred billion dollars off Nvidia’s market capitalization in a single trading session, was already foreshadowed by the quieter rise of model-efficiency research. AMD’s Instinct MI300X had begun shipping in December 2023 to Microsoft and Meta. Google, Amazon, Microsoft, and Meta were pouring money into custom silicon. Huawei’s Ascend 910C was reportedly being deployed inside Chinese hyperscaler clusters in volumes Beijing did not advertise. Each was a small bet against the proposition that Nvidia’s lead was permanent.

By the close of fiscal 2025 in late January 2025, none of those bets had yet materially altered the data. Nvidia represented roughly seven and a half percent of the S&P 500’s total market capitalization, the largest single-company weighting since the index began in 1957. The data-center business had grown 142 percent year over year. The company’s gross margin sat in the mid-seventies. The order book, by every disclosure the management was willing to make, was constrained on the supply side rather than the demand side, which is to say that the only thing limiting Nvidia’s revenue was Nvidia’s ability to convince TSMC to allocate more wafers and SK Hynix and Micron to deliver more HBM stacks. By any reasonable definition of the term, Nvidia had become the hinge on which the global economy of artificial intelligence turned. Every dollar of capital expenditure announced by Microsoft, Amazon, Google, and Meta passed through Nvidia’s order book on its way to the rest of the supply chain. Every sovereign AI initiative announced by Riyadh, Abu Dhabi, Tokyo, Paris, or Seoul was, in financial terms, a downstream subscription to Nvidia’s product roadmap. Every export-control debate in Washington was a debate about which countries got how many of which Nvidia products and on what schedule.

The data did not stop. Through the fiscal year that closed January 25, 2026, Nvidia’s revenue grew another 65 percent to $215.9 billion, with the data-center segment reaching $197.3 billion against $115.2 billion the prior year. The fourth fiscal quarter alone produced $68.1 billion of revenue, of which $62.3 billion came from the data center, the kind of numbers that most companies were happy to print on an annual basis. The networking business that the Mellanox acquisition had built out, branded as Spectrum-X and Quantum InfiniBand, was now an eleven-billion-dollar quarterly run rate by itself. The fiscal-2027 first-quarter guidance, issued in late February, called for $78 billion of revenue against a backdrop in which Nvidia explicitly assumed zero Data Center compute revenue from China. The Trump administration’s H20 reversal had produced a complicated middle ground in which Nvidia could in principle resume H20 sales to China in exchange for a fifteen-percent revenue share with the U.S. government, a structure the company described internally as a tax on uncertainty and that the China sales team had stopped factoring into base-case forecasts. Even with China off the books, the order trajectory implied a fiscal 2027 in the neighborhood of three hundred billion dollars of revenue. No company in the history of the semiconductor industry had ever printed those numbers. No company in the history of any industry had grown so much, so fast, with such concentration on a single product family.

The man at the center of it had been, in 1996, a cofounder of a company that had a single payroll cycle of cash on hand. He had walked away from a Sega contract that would have funded another quarter of operations because he believed the architecture was technically inferior to what the market would demand. He had sent a letter, in 1997, to a Taiwanese foundry founder he had never met, asking for capacity. Twenty-eight years later, that foundry’s most advanced node was almost entirely allocated to his designs, the chips fabricated on it were the binding constraint on the most ambitious technological build-out since the railroads, and the leather jacket he wore on stage was photographed more often than the chip he was holding. Nvidia had not become the hinge of the global economy by inventing a single decisive technology. It had become the hinge by stacking, over thirty years, a series of decisions that had each looked, at the time, like a survival move. The fabless gamble in 1993. The TSMC partnership in 1998. The CUDA bet in 2006. The Mellanox acquisition in 2019. The willingness to dig a software moat for fifteen years before it produced a dollar of differentiated profit. By 2025, every one of those decisions was producing compounding returns, and the cumulative effect was a company whose business represented something the semiconductor industry had not previously contained: a single firm, headquartered in California, fabricating in Taiwan, assembling in Taiwan, networking in Israel, memory-supplied by South Korea, that controlled the throttle on the global rate of artificial-intelligence progress.

The throttle would, in due course, be tested. The DeepSeek release was already two weeks away when the fiscal-2025 numbers landed. The Trump administration’s tariff and export-control posture was about to start moving in directions even Huang’s lobbying could not fully shape. The energy supply available to data centers in Texas, Virginia, and Arizona was beginning to constrain how fast new AI factories could actually come online. The hyperscalers’ custom-silicon programs were, slowly, starting to take measurable share. None of it had yet broken the basic shape of the market, but every line on Huang’s roadmap was now being read by competitors and adversaries who were no longer willing to grant the assumption that Nvidia’s lead was structurally permanent.

For now, in the early spring of 2025, the leather jacket was on a stage in San Jose again, the GB200 NVL72 was ramping in volume, the Rubin platform was a year out, and the man who had once washed dishes at a Denny’s around the corner was the chief executive officer of the most valuable company in the history of the public markets. He had a phrase he liked to use about the work, which the trade press had picked up and recycled until it had lost most of its sting. Our company, he had said in more than one interview, is always thirty days from going out of business. He had said it in 1996, when it was almost literally true, and he had kept saying it in 2024, when no responsible analyst believed it. The line was not, in either era, a piece of false modesty. It was a description of a personal operating posture that had survived the founding crisis, the dot-com crash, the 2008 financial crisis, the cryptocurrency busts of 2018 and 2022, the October 7 export controls, the Blackwell yield issue, and several smaller emergencies the public never heard about. Whether it would survive what came next was the question the rest of the decade would answer.