The Dishwasher

Jensen Huang got his first job at Denny’s when he was fifteen. He washed dishes, bused tables, worked the floor. He was a Taiwanese immigrant kid in Portland who had been sent to the United States at age nine without speaking English, enrolled at a boarding school in rural Kentucky where he was bullied regularly, and had learned — before he learned calculus or chip design or how to manage sixty direct reports — that the distance between the bottom and the top was not as far as people imagined. It was just work.

Twenty-two years after washing dishes at Denny’s, Huang walked into a different Denny’s — the one on Berryessa Road in East San Jose — and sat down in a booth with Chris Malachowsky and Curtis Priem. The coffee was cheap. The diner was quieter than any of their homes. They had $600 between them. Over Grand Slams and refills, they sketched out a plan for a company that would make chips for 3D graphics on personal computers.

That was April 5, 1993. The company they founded that morning is now worth $4.4 trillion — more than the GDP of Japan, more than every company on Earth except, intermittently, Apple. Jensen Huang is still the CEO. He has never left. He has never been fired, never been pushed out by a board, never pivoted to a different venture. For thirty-three years, he has run the same company. In 1996, competing against 3dfx in the graphics card market, Nvidia’s first product — the NV1 — failed so badly that the company laid off more than half its employees and came within weeks of running out of cash. Huang convinced investors to give him one more shot. The next chip, the RIVA 128, sold a million units in four months and saved the company. Then came the dot-com crash, which halved the stock. Then a decade of being dismissed as “just a gaming company.” Then the crypto mining boom of 2017-2018, which sent GPU prices and revenue soaring — followed by the crypto bust, which cratered demand and taught Huang a lesson about concentration risk that he has never publicly acknowledged but has clearly internalized. Then the AI revolution, which turned his chips into the most sought-after commodity on the planet.

In its most recent fiscal year, Nvidia reported $215.9 billion in revenue — up 65% year-over-year. Data center revenue alone was $197.3 billion. In the fourth quarter, ending January 2026, the company generated $68.1 billion in revenue, with data center accounting for $62.3 billion of it. The quarterly revenue of Nvidia’s data center business is now larger than the annual revenue of every semiconductor company in the world except Nvidia itself.

The stock hit $5 trillion in market capitalization in October 2025, making Nvidia the first company in history to cross that threshold. It pulled back to $4.4 trillion by early 2026, but the correction barely registered against the scale of the ascent: from $360 billion in January 2023, before the AI boom, to $4.4 trillion three years later. A twelve-fold increase in thirty-six months.

Huang, at sixty-three, is worth over $130 billion. He is the wealthiest person in the semiconductor industry by a factor of ten. He still wears the leather jacket.

This is the story of how he got here — and whether the position he has built is as unassailable as the market believes.

The CUDA Moat

The conventional explanation for Nvidia’s dominance is hardware. Nvidia makes the best AI chips. Everyone buys Nvidia chips. Therefore Nvidia dominates.

The real explanation is software.

In 2006 — seventeen years before ChatGPT, seven years before deep learning went mainstream, at a time when Nvidia was primarily a gaming GPU company — Huang made a decision that would prove to be the most consequential strategic bet in the history of the semiconductor industry. He launched CUDA.

CUDA — Compute Unified Device Architecture — was a software platform that allowed developers to use Nvidia GPUs for general-purpose computing, not just graphics rendering. The idea was simple in concept and radical in implication: GPUs, which were designed to perform thousands of parallel calculations simultaneously for rendering pixels on a screen, could also perform thousands of parallel calculations simultaneously for other purposes. Scientific simulations. Financial modeling. And, eventually, training neural networks.

The bet was expensive and unpopular. Nvidia spent years investing in CUDA with minimal commercial return. Analysts questioned why a graphics card company was building a software ecosystem. The gaming business was profitable and growing. CUDA was a cost center with uncertain payoff.

Huang persisted. Over fifteen years, CUDA evolved from a programming toolkit into a comprehensive ecosystem — optimized libraries for deep learning (cuDNN, cuBLAS, TensorRT), tight integrations with PyTorch and TensorFlow, compilers, debuggers, profilers, and thousands of SDKs for specific workloads. When universities began teaching deep learning, they taught it on CUDA. When researchers published papers, they benchmarked on Nvidia GPUs. When companies hired ML engineers, they hired people who knew CUDA.

The cycle fed itself. More CUDA developers meant more GPU sales. More GPU sales funded more CUDA investment. More investment attracted more developers. By 2023, when the AI boom arrived, PyTorch and TensorFlow were optimized for CUDA first, every major ML framework assumed Nvidia hardware, and the global supply of AI engineers was trained on Nvidia’s stack. Switching to an alternative meant rewriting optimized kernels, retraining teams, revalidating performance pipelines, and accepting months of dead time. An engineering team that tried to migrate from CUDA to AMD’s ROCm in 2024 described the experience as “remodeling a house while living in it during an earthquake.”

Nvidia holds between 80% and 95% of the AI GPU market, depending on how you measure it. That number is not primarily a reflection of chip quality. It is a reflection of an ecosystem that took seventeen years to build and that no competitor has been able to replicate in three.

The Blackwell Machine

CUDA kept customers locked in. Blackwell kept them desperate.

Nvidia’s Blackwell architecture entered full volume production in early 2026 with a problem that most companies would envy: it could not be manufactured fast enough. The B200 GPU and the liquid-cooled GB200 NVL72 rack system — a configuration of 72 interconnected GPUs that functioned as a single computational unit — were designed specifically for the workloads that defined frontier AI: training trillion-parameter models and serving inference at scale.

The demand was, in Huang’s word, “insane.” Morgan Stanley reported in November 2024 that the entire 2025 production of Blackwell silicon was already sold out. The backlog reached 3.6 million units from cloud providers alone. Major hyperscalers were deploying roughly 1,000 NVL72 racks per week — 72,000 Blackwell GPUs entering data centers every seven days.

The revenue trajectory tracked the demand. Nvidia shipped approximately 2.52 million GB200 units, with data center revenue climbing from $39.1 billion in Q1 FY2026 to $62.3 billion in Q4 — a 59% increase within a single fiscal year. For the full fiscal year ending January 2026, data center revenue was $197.3 billion, up from $115.2 billion the prior year.

Huang told investors and analysts that the company had “visibility into $500 billion in cumulative Blackwell and Rubin revenue through the end of 2026.” The statement was carefully worded — “visibility” is not a guarantee — but the message was clear: the pipeline of committed orders extended far beyond what the company could produce.

The production ramp itself was a logistics achievement of unusual complexity. The GB200 NVL72 racks required coordination across dozens of suppliers — TSMC for the silicon, SK Hynix and Micron for HBM3e memory, a network of cooling system manufacturers for the liquid cooling infrastructure that these power-hungry systems demanded. Supply chain disruptions at any point could delay shipments and ripple through the AI industry, as models that needed to be trained on Blackwell hardware sat idle waiting for chips.

There were disruptions. Nvidia’s suppliers sent mixed signals about GB200 delivery timelines throughout 2025. Some systems shipped on schedule; others were delayed by weeks or months. The delays were manageable in aggregate but frustrating for individual customers who had organized entire engineering roadmaps around expected delivery dates.

The frustration pointed to a deeper asymmetry. Nvidia’s customers — the hyperscalers spending $700 billion on AI infrastructure in 2026 — were dependent on a single supplier for their most critical hardware. Google had TPUs. Amazon had Trainium. But for the workloads that mattered most — training frontier models at the bleeding edge of capability — Nvidia’s GPUs remained the default choice, and the default was not easily changed.

Sixty Direct Reports and No One-on-Ones

Nvidia’s organizational structure is unusual in ways that reflect Huang’s personality and convictions.

Huang has sixty direct reports. He does not conduct one-on-one meetings. When he has something to say — a criticism, a direction change, a strategic decision — he says it in group settings. He believes that information should flow freely, that hierarchies create silos, and that transparency eliminates the politics that slow large organizations down.

“If I don’t like the way something is going, I just say it aloud in a group setting,” Huang has explained. “I don’t take people aside.” The approach is, by most accounts, effective and uncomfortable. Employees describe Huang as demanding, direct, and allergic to hedging. He asks employees across the company to email him weekly with the five most important things they’re working on. He reads them.

He works seven days a week. When Stripe CEO Patrick Collison asked him about work-life balance during a public interview, Huang paused, then said he is either working or thinking about work during every waking moment. The audience laughed. Huang did not. He was stating a fact. At Stanford, when a student asked what advice he would give to aspiring entrepreneurs, Huang said: “I wish upon you ample doses of pain and suffering.” The audience laughed again. Again, Huang was serious. “Greatness is not intelligence,” he explained. “Greatness comes from character. And character isn’t formed out of smart people. It’s formed out of people who suffered.”

The flat structure — sixty direct reports, no intermediate management layers — allows Huang to reach deep into the organization. He can engage with an engineer working on a specific CUDA library or a sales executive negotiating a hyperscaler contract without going through multiple levels of translation. The downside is that it makes Nvidia brittle in a specific way: the company is organized around the assumption that Jensen Huang is available, engaged, and correct. There is no obvious successor. There is no second layer of leadership that has been groomed to make strategic decisions independently. The company’s most important competitive advantage — the ability to make fast, coordinated decisions across hardware, software, and business strategy — is inseparable from the person who makes those decisions.

Huang is sixty-three. He shows no signs of slowing down. But every company analyst who models Nvidia’s future must confront a question that the company itself does not publicly address: what happens when Jensen Huang is no longer CEO? The answer, based on the current organizational structure, is that nobody knows. And the fact that nobody knows is itself a risk factor for a $4.4 trillion company.

The culture Huang has built extends beyond the management structure. Nvidia’s engineering culture is characterized by an intensity that borders on obsession — a willingness to take multi-year bets (CUDA), an expectation that every employee operates at the limits of their capability, and a tolerance for failure that is conditional on the failure being ambitious. “No task is beneath me,” Huang has said, invoking the Denny’s dishwashing years. “I used to clean toilets.” The message is dual: humility about one’s status and ruthlessness about one’s output.

The culture works. The results prove it. But Glassdoor reviews tell a consistent story: exhilarating highs, relentless pace, and a CEO who remembers washing dishes and expects everyone else to remember too. Nvidia has avoided the executive exodus that gutted OpenAI and the talent bleed that weakened DeepMind. How long that holds, as the AI talent market tightens and competitors offer equity packages calibrated to poach Nvidia engineers, is an open question.

The Threats

Nvidia’s dominance is real. It is also under pressure from more directions than at any point in the company’s history.

The most significant threat comes not from a competitor building a better chip, but from customers building their own.

Google’s TPU program is the most advanced example. Google has been designing custom AI accelerators since 2015 and is now on its seventh generation — Ironwood — which delivers 42.5 exaflops of peak compute in a 9,216-chip pod configuration. Google trains and serves Gemini on TPUs, not Nvidia GPUs. Anthropic signed what was described as the largest TPU deal in Google Cloud history in November 2025, committing to hundreds of thousands of Trillium chips in 2026 and scaling toward one million by 2027.

The economics are stark. Midjourney, which had been running inference workloads on Nvidia A100 and H100 clusters, moved the majority of its image generation fleet to Google TPU v6e pods in Q2 2025. Monthly inference spending dropped from approximately $2.1 million to under $700,000 while maintaining the same output volume. A 67% cost reduction for an equivalent workload.

Amazon’s Trainium chips, while less publicized, represent a similar dynamic. Amazon has invested billions in custom silicon for its AWS AI infrastructure and is offering Trainium-based instances at prices that significantly undercut Nvidia GPU instances for certain workload profiles.

Custom ASIC shipments from cloud providers are projected to grow 44.6% in 2026, compared to 16.1% growth for GPU shipments. The growth differential signals a structural shift: hyperscalers are not replacing Nvidia GPUs entirely, but they are diverting incremental demand toward their own silicon, particularly for inference workloads where the cost advantages of custom chips are most pronounced.

Goldman Sachs has privately estimated that Google TPUs could reach 35% market share in AI inference by Q4 2026. If that estimate proves accurate, Nvidia faces an uncomfortable choice: cut inference chip prices by 40-50% to remain competitive, or cede the inference market and focus on training, where its performance advantage remains widest.

AMD represents a more conventional competitive threat. AMD’s MI350 series, launched in 2025, was the fastest-ramping product in the company’s history. The MI450 “Helios” systems, scheduled for Q3 2026, promise rack-scale performance leadership for specific workload profiles. AMD’s chips are generally cheaper than Nvidia’s equivalents, and the company has made progress in building a software ecosystem — ROCm — that, while still far behind CUDA in maturity, is closing the gap.

Then there are the abstraction layers. OpenAI’s Triton compiler, MLIR (Multi-Level Intermediate Representation), and similar hardware-agnostic tools allow developers to write high-performance code that runs across different accelerator architectures without being locked into CUDA. These tools are still early, still less optimized than CUDA for most workloads, and still require significant engineering effort. But they represent a strategic threat to the lock-in mechanism that has been the foundation of Nvidia’s dominance for seventeen years.

Huang’s response to these threats has been characteristically aggressive. At CES 2026, he unveiled the Rubin platform — six new chips, including the Rubin GPU, Vera CPU, and a complete networking stack, designed to deliver a 10x reduction in inference token cost and a 4x reduction in the number of GPUs needed to train mixture-of-experts models compared to Blackwell. Rubin uses TSMC’s 3nm process with HBM4 memory and enters full production in the second half of 2026. Rubin Ultra follows in 2027, with a 576-GPU rack configuration delivering 15 exaflops of FP4 inference compute. Beyond Rubin, the roadmap includes an architecture named Feynman — after the theoretical physicist — with no public details and no expected timeline.

The product cadence itself is the strategy. By releasing a new architecture annually — Hopper in 2022, Blackwell in 2024, Rubin in 2026, Feynman presumably in 2028 — Nvidia forces customers to commit to a continuous upgrade cycle and makes it harder for competitors to catch up. The moment a rival chip approaches parity with Nvidia’s current generation, Nvidia releases the next generation. The treadmill never stops.

Huang addressed the competitive pressure directly at a February 2026 event, saying the $700 billion in hyperscaler AI capex was “just the start of something far bigger.” He also made a move that surprised some observers: Nvidia invested $5 billion in Intel in September 2025, at $23.28 per share, alongside a collaboration to co-develop custom x86 CPUs integrated with Nvidia GPUs via NVLink. The investment in a struggling competitor was read as both a strategic hedge and a signal that Nvidia saw value in controlling the CPU-GPU integration stack rather than ceding it to AMD or Arm.

In another unexpected move, Huang publicly said in March 2026 that Nvidia’s $30 billion investment in OpenAI “might be the last” — suggesting the company was reconsidering its role as both supplier and investor in the AI companies that were also its customers. The statement hinted at a strategic clarity that had always been part of Huang’s approach: make the picks and shovels, sell to everyone, and don’t get entangled in the fortunes of any single customer.

The $700 Billion Question

Nvidia’s financial position, at this moment, is nearly impregnable. Every frontier AI model is trained on Nvidia GPUs. The backlog exceeds supply. CUDA is the standard. The roadmap extends three generations out. At $68.1 billion in quarterly revenue, the company is generating cash faster than it can deploy it.

The vulnerability is not in the present. It is in the structure of the present.

Nvidia’s revenue is extraordinarily concentrated. Data center accounted for 91% of total revenue in Q4 FY2026. Within data center, a small number of hyperscalers — Microsoft, Google, Amazon, Meta, and a handful of sovereign AI programs — represent the majority of orders. If any of these customers significantly reduced purchases, shifted to custom silicon, or experienced a reduction in AI-related revenue that caused them to pull back on infrastructure spending, Nvidia’s revenue would decline sharply.

The $700 billion in combined hyperscaler capex for 2026 is predicated on the assumption that AI demand will continue to grow faster than infrastructure can be built. That assumption may prove correct. But it is an assumption, and it depends on AI applications generating sufficient economic value to justify the infrastructure investment. If enterprise AI adoption plateaus, if open-source models become good enough to reduce the need for frontier-scale training, or if inference costs drop faster than expected due to algorithmic improvements or custom silicon, the infrastructure buildout could decelerate. Nvidia would feel it first.

The geopolitical dimension adds complexity. Nvidia has become, without seeking the role, a tool of American foreign policy. The company’s chips are the primary target of U.S. export controls on advanced AI technology. Every restriction on what Nvidia can sell to China is, in effect, a restriction on Nvidia’s revenue — imposed by its own government. U.S. export controls have restricted Nvidia’s ability to sell its most advanced chips to Chinese customers. Nvidia has developed compliance-specific chip variants — the H20 and its successors — that operate within the regulatory limits, but these chips are less capable and less profitable than the unrestricted versions sold to Western hyperscalers. China represents a massive potential market that is, for now, largely inaccessible for Nvidia’s highest-margin products.

And then there is the succession question. Jensen Huang is Nvidia. Not in the metaphorical sense that Steve Jobs was Apple, but in the operational sense that Huang’s daily involvement in product decisions, architectural choices, and strategic direction is embedded in how the company functions. The sixty-direct-report structure works because Huang is there. The annual architecture cadence works because Huang drives it. The ability to simultaneously manage hardware, software, and business strategy works because one person holds all three in his head.

The company has not publicly addressed succession planning. There is no known heir apparent. The organizational structure has no natural second-in-command. For a company worth $4.4 trillion — more than the combined value of most national stock exchanges — the single-point-of-failure risk is unusual and, by any governance standard, concerning.

The Leather Jacket

In January 2026, Huang spoke at Davos. He told the audience that AI was “the largest infrastructure buildout in human history” and that every country should treat AI capacity as essential national infrastructure, like electricity or roads. The audience — heads of state, central bankers, CEOs — listened with the deference typically reserved for the people who control things that governments need and cannot build themselves.

That deference is the measure of Nvidia’s position. Jensen Huang is not a head of state. He does not command an army or control a currency. But he controls the supply of the commodity that every government and every corporation in the world has decided it needs: AI compute. He is the bottleneck.

The position is precarious in the way that all bottleneck positions are precarious. Everyone who depends on Nvidia is, simultaneously, motivated to find an alternative. Google is building TPUs. Amazon is building Trainium. AMD is shipping MI350. The abstraction layers are maturing. The custom ASIC market is growing at three times the rate of GPU shipments.

But the bottleneck holds. CUDA is seventeen years deep. Blackwell is sold out through mid-2026. Rubin is in production. The architecture cadence keeps the treadmill moving faster than anyone can catch it. And the man running the treadmill works seven days a week, reads emails from individual engineers, and has been doing this for thirty-three years without a break.

Whether anyone — Google, AMD, the open-source abstraction layers, the custom ASIC wave — can break the bottleneck within the next five years is an open question. The company that dominates the most consequential market in the global economy was founded over cheap coffee at a Denny’s by three friends with $600.

Huang still goes back to that Denny’s. In 2023, after Nvidia crossed $1 trillion in market cap, he returned to the booth where it started. Denny’s put a commemorative plaque on the table. Huang ordered the Super Slam — the same thing he always orders.

The dishwasher built a $4.4 trillion company. History says no position this dominant lasts forever — not Standard Oil, not IBM, not Intel. Jensen Huang’s thirty-three-year track record says something different. He has been written off before. He nearly went bankrupt before. He survived the dot-com crash, the crypto bust, and a decade of irrelevance before the world decided it needed what he had been building all along.

Don’t bet against the dishwasher. But don’t assume the treadmill runs forever, either.

Published March 6, 2026. This investigation covers Jensen Huang’s leadership and Nvidia’s strategic positioning through early 2026.

About the Author

Gene Dai is the co-founder of OpenJobs AI, focusing on AI-powered recruitment technology and the intersection of artificial intelligence with enterprise software.