Jensen Huang

The Reluctant Monopolist

Born: 17 February 1963, Tainan, Taiwan

For most of Nvidia’s history, Jensen Huang was building the wrong thing for the right reasons. He had co-founded the company in 1993 to build graphics processing units — chips designed to render three-dimensional images for video games, a market that the established semiconductor companies considered insufficiently serious for their attention. He spent the 1990s competing with ATI and 3dfx and S3 Graphics on benchmarks that were measured in frames per second on Quake. He spent the 2000s watching his chips become standard equipment in gaming PCs while the serious enterprise computing market continued to run on CPUs from Intel and AMD.

And then, around 2012, the researchers who were training deep neural networks discovered something that would change the trajectory of Huang’s company and, ultimately, of the entire technology industry: the same mathematical structure that made GPUs good at rendering graphics — thousands of small processing cores performing simple operations in parallel — made them exceptionally good at training neural networks. The matrix multiplications at the heart of backpropagation were, it turned out, exactly the kind of computation GPUs were designed to perform, and they performed it orders of magnitude faster than CPUs.

Huang did not cause this discovery, but he had, without knowing it, spent twenty years preparing for it. The GPU architecture he had championed, the CUDA programming model his engineers had developed to make GPUs programmable for general-purpose computation, the manufacturing relationships with TSMC that gave Nvidia access to the best available fabrication processes — all of these were in place before any AI researcher noticed they were useful. When they did notice, Nvidia was the only company positioned to supply what they needed.

Taiwan, Kentucky, and Oregon

Huang was born in Tainan, Taiwan, in 1963. His family emigrated when he was nine, and his trajectory through the American educational and corporate system was not smooth. He and his brother were sent, at ages nine and ten, to live with an uncle in the United States while their parents remained in Taiwan — an arrangement that placed them, briefly but memorably, in a school that served incarcerated youth in Kentucky. The experience gave Huang early exposure to the resourcefulness that comes with having few resources, and to the resilience that comes from being far from home.

The family eventually reunited in Oregon, where Huang completed secondary school before attending Oregon State University for his undergraduate degree in electrical engineering. He then worked at AMD as a microprocessor designer before completing a master’s degree at Stanford. At AMD he encountered the culture of semiconductor engineering at close quarters — the extraordinary precision required, the extraordinarily long timescales of chip design, the relationship between architectural decisions made years in advance and products that would compete in markets that did not yet exist.

In 1993, at thirty, he co-founded Nvidia with Chris Malachowsky and Curtis Priem. The founding insight was simple: the market for three-dimensional graphics was about to become very large, established chip companies were not pursuing it seriously, and a focused startup with the right architecture could establish a dominant position before the incumbents noticed. The insight was correct. The journey from correct insight to dominant company took fifteen years, required multiple near-death experiences, and produced a corporate culture that Huang has described as running on fear — the fear of the next competitor, the next architectural shift, the next thing that would make everything you had built irrelevant.

The GPU Architecture and CUDA

Nvidia’s first major product, the RIVA 128, was released in 1997 and was fast enough to establish the company as a serious competitor in the consumer graphics market. The GeForce 256, released in 1999, was the first chip that Nvidia described as a GPU, and it introduced hardware transform and lighting — moving geometric calculations from the CPU to the graphics chip — that became standard for the following decade.

The architectural philosophy behind these chips, which Huang drove with unusual consistency over a long period, was the inverse of the CPU philosophy. CPUs were designed to execute a small number of complex operations very quickly, with elaborate caching and branch prediction hardware to minimise latency. GPUs were designed to execute a very large number of simple operations simultaneously, trading latency for throughput. The tradeoff made sense for graphics, where the relevant computation was the same operation applied to millions of pixels, but it was not obvious, in the early 2000s, that it would be useful for anything else.

CUDA, released in 2006, was Huang’s most consequential strategic decision. It was a programming model and a set of development tools that made Nvidia GPUs programmable for general-purpose computation — not just graphics but scientific simulation, financial modelling, and anything else that could be expressed as highly parallel numerical computation. The decision to develop CUDA was controversial internally; it diverted engineering resources from the core gaming business without any obvious immediate commercial payoff. Huang pushed it through on the basis of a conviction that programmable GPUs would eventually be useful for things he could not yet specify.

The payoff came from an unexpected direction. Researchers at universities, working on machine learning, began experimenting with CUDA in the late 2000s and found that training neural networks on GPUs was dramatically faster than training on CPUs. By 2010, CUDA-accelerated deep learning was standard in academic ML research. By 2012, when AlexNet won ImageNet, the GPU was established as the essential piece of hardware for AI research. By 2015, the major cloud providers were building data centres full of Nvidia GPUs to serve AI workloads.

The AI Accelerator

Nvidia’s transition from a gaming chip company to the essential hardware infrastructure of the AI industry was not a strategic pivot so much as a recognition, gradually and then very quickly, of what the company had already built. The GPU architecture that made Nvidia chips good at gaming also made them good at training neural networks. The CUDA ecosystem that researchers had built on top of that architecture was not easily replicated by competitors. The manufacturing relationships with TSMC that Huang had cultivated for years gave Nvidia access to the most advanced fabrication processes.

The company introduced dedicated AI hardware with the Tesla series — GPU products sold not for gaming but for data centre computation. The V100, released in 2017, was the chip that trained GPT and many of the other large models of the era. The A100, released in 2020, was the chip that trained GPT-3. The H100, released in 2022, became so important for large language model training that its scarcity became a constraint on the entire AI industry and a subject of US government export control policy.

The export controls, introduced by the Biden administration in 2022 and extended subsequently, reflected a recognition that Nvidia’s chips were not merely commercial products but strategic assets. Restricting the export of H100s and their successors to China was an attempt to limit China’s ability to train frontier AI models, and it reflected a geopolitical assessment of AI chips as infrastructure comparable to semiconductors for weapons systems. Huang navigated this environment with visible discomfort — China was a major market for Nvidia, and the export controls required the development of alternative, less capable products for Chinese customers — but with the pragmatism that had characterised his entire career.

By 2024, Nvidia’s market capitalisation had briefly exceeded three trillion dollars, making it the most valuable company in the world. Huang, who owned approximately three percent of the company, became one of the wealthiest people on the planet. The valuation reflected a market assessment that AI would require essentially unlimited amounts of the computation that Nvidia chips provided, and that Nvidia’s architectural lead and software ecosystem moat would sustain its dominance for at least the medium term.

The Culture of Intensity

Huang has been unusually candid about the culture he has built at Nvidia, and about its costs. He has described the company as existing in a state of existential urgency — always one missed product cycle away from irrelevance, always aware that the architectural shift that would make GPUs obsolete might be developed by a competitor or a startup or a research lab. He has been explicit that this urgency is, in part, manufactured — that it serves a motivational function, that it keeps the organisation from becoming complacent in the way that market dominance can encourage.

He is famous for his leather jacket, which he wears at every public appearance, and for the tattoos on his arms, which document milestones in Nvidia’s history. He gives the impression of a man who decided, at some point, to be memorable, and who has been consistently memorable ever since. His keynotes at the GPU Technology Conference, which have grown into multi-hour events attended by tens of thousands, are theatrical in a way that few semiconductor company events are. He uses them to announce new chips, new software frameworks, new partnerships, and new visions of the AI future, all delivered in a style that combines genuine technical depth with showmanship.

The combination has been effective. Nvidia under Huang is a company that is simultaneously a hardware manufacturer, a software platform, a research organisation, and a cultural force in the AI industry. No single decision he made was as important as the series of decisions — the GPU architecture, CUDA, the data centre pivot, the manufacturing relationships, the software ecosystem investments — that compounded over thirty years into an infrastructure monopoly that the AI industry could not function without.

Key Works & Further Reading

Primary sources:

CUDA programming documentation (2006–present). The technical foundation of Nvidia’s AI dominance.
Nvidia GTC keynote addresses (2010–present). The best record of Huang’s evolving vision for the relationship between GPU computing and AI.
“NVIDIA H100 Tensor Core GPU Architecture” — Nvidia white paper (2022). The technical specification of the chip that became the infrastructure of the AI era.

Recommended reading:

Chip War — Chris Miller (2022). The most comprehensive account of the semiconductor industry and geopolitics that forms the context for Nvidia’s rise.
The Innovator’s Dilemma — Clayton Christensen (1997). The framework for understanding why incumbents missed the GPU opportunity that Nvidia exploited.
The Art of the Long View — Peter Schwartz (1991). Scenario planning methodology; illuminating for understanding the kind of long-horizon strategic thinking Huang has practised.
Turing’s Cathedral — George Dyson (2012). The history of the first programmable computers; essential for understanding the architectural lineage that leads to the GPU.