The House That AI Built

Yesterday, Semafor reported a governance crisis at chip manufacturer Nvidia. Its chief executive, Jensen Huang, has admitted to having had an affair with Alice Huang, a Megaspeed executive suspected of spying for China. Nvidia’s board opened an independent investigation.¹ President Trump confirmed the FBI is also looking into the matter.² Bloomberg put Huang’s return since Nvidia’s 1999 IPO at around 700,000 per cent, the kind of number that turns »irreplaceable« from praise into diagnosis.³

»Nvidia supplies an estimated 90 per cent of high-end processors powering advanced AI systems. It’s the central node in a vast network of companies, data centres, and research programmes, resilient to ordinary disruption but exposed to the failure of its hub«

In his trademark black leather jacket, Huang has cultivated a rock-star persona. The world’s most important AI hardware company has built much of itself around him. He has no designated successor or chief operating officer, and around 60 direct reports. Nvidia has also operated without a chief compliance officer.⁴

Redundancy sounds redundant. The word suggests surplus, something left over and ripe for removal. In Britain, it means losing a job deemed unnecessary. But in complex systems, redundancy can make the difference between a system that fails and one that keeps functioning. The question is how resilience is achieved.

Biologists Gerald Edelman and Joseph Gally distinguish between redundancy and degeneracy. The former means identical elements capable of performing the same function; the latter, structurally different elements capable of doing so. Degeneracy can be more powerful: different components can compensate for one another without sharing the same weaknesses.⁵

But more components don’t automatically mean greater resilience. In Normal Accidents, sociologist Charles Perrow argued that backup systems in complex technologies can create new interactions and unexpected failure modes. More components might mean more ways to fail. The aim is not to duplicate every role, but to create independent ways of achieving essential outcomes without making the system harder to understand.⁶

Nvidia illustrates the distinction. Rather than appointing a dedicated compliance officer to oversee its legal and regulatory obligations, it assigned the role to a committee of three senior executives: its CFO, head of operations, and head of HR. That looks like redundancy: several people sharing a responsibility. But it offers little degeneracy. None has an independent compliance function; all remain occupied with their primary roles. Three names sit on the committee, but there is no alternative structure to take over if it fails.

In 2000, physicists Réka Albert, Hawoong Jeong, and Albert-László Barabási showed how scale-free networks withstand random node failures but remain vulnerable to the targeted removal of highly connected hubs. Nvidia supplies an estimated 90 per cent of high-end processors powering advanced AI systems. It’s the central node in a vast network of companies, data centres, and research programmes, resilient to ordinary disruption but exposed to the failure of its hub.⁷

In 2005, control theorist John C. Doyle described the Internet as »robust yet fragile«.⁸ Systems become good at handling expected failures, but vulnerable to those outside their design. Nvidia has spent years managing technical and commercial complexity. But its founder’s personal conduct becoming a board crisis, national-security question, and test of market confidence introduces another kind of complexity. It’s closer, as one Nvidia-watching academic put it, to »a day of reckoning« than to an orderly transition. Huang has led Nvidia since co-founding it in 1993. After 33 years, the company is so closely identified with him that his indispensability is itself a risk.

Consider Mixtral, an AI language model built from eight specialised neural networks, known as experts. When processing text, it breaks it into small units called tokens and sends each to just two experts, rather than activating all eight for every calculation.⁹ This distributes computational work across a system instead of concentrating it in one network. Nvidia makes the processors that run these distributed calculations. Yet the company has concentrated so much authority, knowledge, and responsibility in Jensen Huang that it has no obvious way to function without him. Its own single point of failure wears a black leather jacket.

References