Jensen Huang and the contradictions of the AI age

Editorial illustration of Jensen Huang beside GPU hardware and server infrastructure, in mint, lilac and petroleum-blue tones

By Fabrizio Rinaldi. News context checked on 12 November 2026.

For years, Jensen Huang has been the personification of the AI boom: the engineer in the leather jacket who can explain a GPU to a room of developers and sell a national infrastructure project to a government. In November, a very different document has put him at the centre of public attention.

On 8 November, Nvidia published his message to employees acknowledging infidelity and apologising to his family and colleagues. “I was unfaithful to my wife,” he wrote. The admission followed reporting about a relationship with Alice Huang, a Chinese businesswoman associated with the Singapore-based cloud company Megaspeed International. Reuters could not independently verify the allegation that she was a Chinese intelligence asset, which China has denied. Huang’s own message did not identify his partner. An admission of infidelity is not an admission of espionage. [1] [2]

The corporate consequences are nevertheless serious. Nvidia’s board announced an independent investigation, while stating on 8 November that it had no reason to believe company secrets or national security had been compromised. On 10 November, its independent members appointed outside counsel. By 12 November, Reuters was reporting deliberations over Huang’s future, an announced removal. These are distinct stages of an unfolding case, not a completed verdict. [3] [4] [5]

Public reproduction of Jensen Huang’s 8 November 2026 message to Nvidia employees, acknowledging infidelity and supporting an independent board investigation
Huang’s employee message of 8 November 2026. Public document reproduction published by the New York Times; the text is also available in Nvidia’s newsroom. Read the full-size document. [20]

I am interested in the breach of trust because leadership includes accountability, especially when business relationships intersect with sensitive technology. I am much less interested in turning his family’s suffering into entertainment. The more useful task is to understand the person behind that influence, the company he helped build, and the assumptions inside his vision of our technological future. Neither the scandal nor the leather jacket is an adequate explanation of Jensen Huang.

From Taiwan to the American semiconductor industry

Born in Taiwan in 1963, Huang spent part of his childhood in Thailand before being sent to the United States with his older brother, ahead of their parents. He attended a boarding school in Kentucky under difficult circumstances and later settled with his family in Oregon. It was an immigrant childhood involving separation, adjustment and work, rather than the frictionless origin story that tends to emerge once a founder becomes successful. [6] [7]

He earned an electrical-engineering degree at Oregon State University in 1984 and a master’s at Stanford in 1992 while working full time. Before Nvidia, he designed microprocessors at AMD and worked at LSI Logic. The familiar Denny’s connection has two chapters: he worked there as a dishwasher and waiter, and years later held startup-planning meetings there with Chris Malachowsky and Curtis Priem. [7] [8] [10]

Those experiences help make sense of his practical vocabulary. He returns to constraints, architecture and the work required to make an idea function. Yet I would resist the temptation to explain everything about an adult’s decisions through childhood hardship. A biography can illuminate a way of thinking without becoming an excuse for conduct or a psychological diagnosis.

Remy from Ratatouille holds a piece of cheese and a strawberry together against a warm orange background.
Ratatouille: curiosity rarely starts with a business case. GIF source.

Nvidia was built for graphics before it became central to AI

Huang, Malachowsky and Priem founded Nvidia in April 1993. Their opportunity was the emerging market for three-dimensional graphics, particularly games, rather than a plan to dominate generative AI thirty years later. Huang’s strategic contribution matters, but it sits alongside the technical work of his cofounders and generations of engineers. [9]

The early company also made mistakes. Its NV1 chip used a graphics approach that diverged from the triangular polygons underpinning Microsoft’s Direct3D. Commercial uptake was disappointing, and the company had to change direction. In his Stanford account, Huang describes a reset that involved learning from graphics textbooks rather than defending the original approach simply because the team had invested in it. The lesson is less flattering, and more useful, than the idea of a genius whose first instincts were always correct. [10] [11]

Nvidia’s 1999 GeForce 256 was marketed as the first GPU, a graphics processing unit. The important advance was integrating substantial graphics work that would otherwise burden the central processor. Graphics required many operations to be carried out in parallel, and that capability would eventually have uses well beyond rendering a game. [6] [9]

CUDA, unveiled in 2006, helped turn that possibility into a usable computing platform. It gave developers a way to program Nvidia GPUs for general-purpose workloads, including scientific computation. Hardware alone was not sufficient: researchers needed tools that let them use its parallel processing without treating every problem as a graphics problem. [6] [9]

The decisive AI example arrived in 2012. Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton demonstrated a major advance in image classification with the neural network now known as AlexNet. Their paper describes training on two Nvidia GTX 580 GPUs. The breakthrough belonged to the researchers, while Nvidia supplied hardware and a programmable platform that helped make the experiment practical. [12]

That division of credit matters. Nvidia did not invent AI; it helped make increasingly ambitious AI computation feasible. Its subsequent importance grew through continued development of accelerators, software and the surrounding ecosystem. Models are what most users see, but the ability to train and run them depends on infrastructure, and Nvidia built a business around that requirement. [6] [9]

Training means using examples to adjust a model’s parameters; running the trained model to generate a result is inference. AlexNet illustrates the first task, while a system answering a request or classifying a new image performs the second. Both require computation. The research breakthrough was therefore also a demonstration of what becomes possible when algorithms have access to suitable hardware and software, not just an improvement in the appearance of computer graphics. [12]

For me, the interesting achievement is the connection between a commercially successful graphics business and a longer investment in computing that could serve unfamiliar problems. It required patience with a platform before every future application could be named. Once researchers demonstrated what that platform could do, Huang committed to the opportunity. That is substantial strategic judgement without requiring us to believe that he predicted ChatGPT in a Denny’s booth.

Wreck-It Ralph appears inside a pixel-art arcade game screen beneath the words I’M GONNA WRECK IT!
Wreck-It Ralph brings us back to the gaming world where Nvidia began. GIF source.

His philosophy combines ambition with an expectation of difficulty

At Stanford’s 2024 SIEPR summit, Huang argued that character and resilience are built through setbacks. His remark about wishing students “ample doses of pain and suffering” attracted understandable attention, alongside his claim that low expectations were an advantage. The point was not to set small goals, but to avoid expecting an easy journey or treating disappointment as a reason to stop. [13]

I understand the appeal of that attitude. A long project rarely survives if every obstacle is interpreted as evidence that the idea was worthless. But suffering does not become beneficial merely because a successful person survived it. There is a difference between learning from difficulty and creating avoidable difficulty for other people. A founder’s willingness to endure pressure is not a blank cheque to demand the same sacrifices from everyone around him.

His management philosophy is similarly worth examining beyond its headline. At Stanford, he described reducing organisational layers that distort information and rejecting the idea that possession of information should be the source of managerial power. Problems should be visible enough for people to reason through them together. This is an argument for fewer silos, not for a company without coordination or responsibility. [10]

It also contains a tension. A flatter structure can circulate knowledge more effectively while concentrating a large number of relationships around the chief executive. Removing an intermediate layer does not automatically distribute final authority. The current crisis makes that distinction unusually tangible: an organisation can be collaborative in its engineering and still depend heavily on one person’s judgement.

Andrew Neiman sits at a drum kit in a rehearsal room, striking the drums with focused intensity in Whiplash.
Whiplash makes the useful point that demanding practice is not the same as mistreatment. GIF source.

Sovereign AI needs a hardware conversation

Huang’s argument for sovereign AI starts from a legitimate concern. Countries should be able to produce intelligence from their own languages, history and institutional knowledge, rather than only consume systems designed elsewhere. At the World Governments Summit in 2024, he urged governments to build AI infrastructure and models reflecting their own data and culture. His 2026 Berlin keynote extended that message to Europe’s industrial expertise. [14] [19]

I find that more convincing than the assumption that one globally dominant model will adequately represent every society. Local capability can support languages, public services and specialised industries that would otherwise be peripheral to somebody else’s product roadmap. But “sovereign” is doing considerable work in this pitch.

How sovereign is a country’s AI if it controls the data and model but cannot replace the foreign supplier of the chips that make them usable?

That is not an argument that national sovereignty requires manufacturing every component domestically. It is an argument that data sovereignty, operational control and technological autonomy are different dimensions. A locally operated data centre can still depend on foreign-designed accelerators, software interfaces, replacement equipment and permission to acquire its next generation of computing capacity.

The hardware dimension includes the ability to obtain chips, maintain systems, move workloads to alternative architectures and understand which restrictions could interrupt access. Beyond that sit chip design and manufacturing capabilities, which cannot be created by relabelling an imported server as sovereign infrastructure. Controlling where data is stored addresses an important problem, but it does not answer all those others.

Nvidia’s role in Huang’s proposal is explicit. Its technology enables national and enterprise AI, while his July 2025 message also promoted an American technology stack as the foundation on which civil models should run. Both positions can coexist: a country can gain meaningful control over its applications while a supplier gains a durable place inside its infrastructure. [15] [19]

The difficult part is making that bargain visible. Governments should distinguish an acceptable dependency from one they cannot manage, and test portability rather than merely promise it. Cultural sovereignty also needs scrutiny: a model trained locally does not automatically represent everyone in a country, or guarantee citizens any say in how their information is used.

I would describe sovereign AI as a set of capabilities to build and defend, not a label conferred by the location of a rack. The same systems-thinking approach that helps organisations understand dependencies is useful here: control over one part of a system should not be mistaken for control over the whole.

In The Wizard of Oz, an older man stands behind a green curtain beside a glowing machine while a woman watches; the GIF overlays the line Pay no attention to that man behind the curtain.
The Wizard of Oz makes hidden control visible, a neat fit for asking who really governs the hardware layer. GIF source.

China is a market, an ecosystem and a security question

Huang’s position on China makes more sense when understood through Nvidia’s platform strategy. In his May 2025 interview with Ben Thompson, he argued that excluding American technology would encourage alternative ecosystems to grow. He recognised the technical strength of Huawei and DeepSeek, rather than assuming that restricting access to chips would remove Chinese engineering ability. [16]

The commercial logic is straightforward. Hardware installations attract developers; developers build applications and tools; those applications make the platform more valuable to subsequent customers. For Nvidia, access to China concerns the future reach of that ecosystem as well as immediate sales. For the United States, Huang presents broad adoption of American technology as a source of influence. [15] [16]

It is a serious argument, not a neutral one. The person making it runs the company that benefits from broader access. Conversely, its commercial usefulness does not prove the argument is wrong. Policymakers have to weigh the possibility of strengthening rival platforms against the risks of supplying capabilities to prohibited users or applications. Those risks cannot be settled by a confident sales pitch, and national-security restrictions should not be treated as proof that an entire population is an enemy.

The rules also require more precision than “China is banned.” A US rule effective from January 2026 allows qualifying H200-class exports to China to undergo conditional, case-by-case licence review. That is not unrestricted access. Reuters reported in November that Chinese approval for large-volume imports remained a separate obstacle. Export permission, import permission and actual delivery are not interchangeable. [17] [18]

The allegations surrounding Huang do not establish that his policy position was secretly dictated by another government. That would require evidence. They do make independent scrutiny of sensitive relationships more important. A technology strategy should be evaluated on its merits, while possible conflicts and misconduct are investigated through a process capable of challenging the person who proposed it.

Captain Jean-Luc Picard faces another crew member in the foreground and gives a measured nod.
Star Trek: Picard leaves room for dialogue, even when interests differ. GIF source.

Agents turn a model into something that can act

Huang’s emphasis on agents concerns the transition from generating an answer to performing work. An agent combines a model with tools, knowledge and an orchestration system that lets it pursue tasks. Access to company files, applications or external services gives that system practical consequences beyond the quality of its prose. [19]

This is also a large commercial change. In his 2025 Stratechery interview, Huang described AI as partly labour, arguing that agents and robots could extend technology spending beyond conventional IT budgets into the cost of operations. Nvidia would then supply computation not only for software people use, but for software carrying out some of their work. [16]

His safety position deserves to be taken seriously. The Berlin keynote calls for sandboxing, monitoring and narrowly assigned privileges, and presents Nvidia’s OpenShell as a runtime designed to control an agent’s access to the systems around it. Giving an agent the minimum permissions needed for a task is much more sensible than granting general access and hoping it behaves. [19]

Still, a secure runtime is not the whole of responsible deployment. An agent can remain within its technical permissions and make a decision that should never have been delegated. Access control can prevent a system from opening an unauthorised file; it does not determine whether the information in an authorised file is complete, whether a proposed action is fair, or whether a person affected by that action can challenge it.

For a business, the practical work includes deciding which actions require human approval, how errors are detected, and who remains accountable. The ability to stop an agent matters, but so does recognising when it should not be given a decision in the first place. Huang’s engineering focus offers useful mechanisms. It should inform governance rather than substitute for it.

Mickey Mouse in his red robe and blue wizard hat watches enchanted brooms carry two buckets across a cave-like room.
In Fantasia, delegation gets ahead of its permissions surprisingly quickly. GIF source.

Optimism about growth is not a guarantee about work

Huang’s broader outlook is that cheaper and more capable computing will expand ambition: organisations will find more things to do rather than simply purchase less. The history of Nvidia makes that belief understandable. A technology developed for games helped create demand in fields that its original market did not contain. [9] [16]

His employment optimism is not quite the claim that nobody loses a job. In the 2025 interview, he explicitly acknowledged that some jobs would be lost while predicting that many would be created. That remains a forecast, not a measured outcome of the next phase of AI. [16]

My concern is the distance between an expanding economy and the experience of an individual worker. A company can produce more while needing fewer people for a particular task. New roles can emerge in a different place, require different qualifications or arrive after someone has already lost their income. Construction work for a data centre does not automatically compensate a displaced office worker.

The social outcome depends on investment choices, training, bargaining power and how gains are distributed. Those are human and institutional decisions, not properties guaranteed by a GPU. An optimistic account of technological possibility becomes more useful when it includes the cost of the transition and a credible responsibility for the people carrying it.

Black-and-white scene of Charlie Chaplin partly hidden among large interlocking factory gears in Modern Times.
In Modern Times, the productivity machine speeds up while the worker still bears the strain. GIF source.

Understanding Huang without surrendering judgement

Huang deserves recognition for helping build a computing platform that changed what researchers and companies could attempt. His story shows technical adaptation, sustained investment and a capacity to connect engineering with emerging markets. It also shows the value of work done by people whose names attract much less attention than the chief executive’s.

His vision offers serious ideas: local AI capability, engagement with a technically capable China, controlled agents and an economy that uses computing to attempt more. Each deserves examination because each combines a real opportunity with unresolved dependencies and commercial interests.

The current crisis should test accountability, not become an excuse to erase Nvidia’s contribution or to presume guilt on allegations that remain unproven. Equally, technological achievement should not exempt a leader from scrutiny.

For me, the useful position is to take his engineering seriously while retaining our own judgement about the society built around it. That means being able to explain which dependencies a country can manage and which decisions a business has delegated, while giving the people asked to adapt a credible share in the benefits. Huang can help expand what computing makes possible; deciding what we entrust to it remains our responsibility.

A red pill and a blue pill float over a green glow while orange discs repeat the words THE CHOICE IS YOURS.
A Matrix-inspired reminder that listening does not mean surrendering the choice. GIF source.

References

  1. Nvidia: Jensen Huang’s message to employees, 8 November 2026.
  2. Reuters: Who is Alice Huang, 10 November 2026.
  3. Nvidia board statement announcing an independent investigation, 8 November 2026.
  4. Nvidia independent directors appoint outside counsel, 10 November 2026.
  5. Reuters: Board deliberations and crisis updates, 12 November 2026.
  6. Computer History Museum: Jensen Huang’s biography and contribution to computing.
  7. Oregon State University: Jen-Hsun Huang alumni biography.
  8. Stanford Technology Ventures Program: Jensen Huang biography.
  9. Nvidia corporate timeline: Founding, graphics, CUDA and AI.
  10. Stanford GSB: Jensen Huang on decisions, information and organisation, 25 April 2024.
  11. Jon Peddie: Nvidia’s quadratic processor, the NV1, 22 June 2021.
  12. Krizhevsky, Sutskever and Hinton: ImageNet classification with deep convolutional neural networks, 2012.
  13. Stanford SIEPR: Huang on AI, resilience and leadership, 7 March 2024.
  14. Nvidia: Huang’s sovereign AI remarks at the World Governments Summit, 12 February 2024.
  15. Nvidia: Huang’s Washington and China visits, 14 July 2025.
  16. Ben Thompson: Interview with Jensen Huang on chip controls, AI factories and enterprise pragmatism, 19 May 2025.
  17. US BIS: Conditional licence review for advanced computing commodities, 15 January 2026.
  18. Reuters: Chinese import approval and Nvidia chip deliveries, 5 November 2026.
  19. Nvidia: GTC Berlin keynote on sovereign AI and agent security, 21 October 2026.
  20. New York Times: Public reproduction of Huang’s employee message, 9 November 2026.

Discover more from INNOVATION WAVES

Subscribe now to keep reading and get access to the full archive.

Continue reading