Reading Nvidia: On strategy, defensibility, and the return of industrial AI
Moats, systems, and the physical economy behind the AI build-out.
A few weeks ago, Jensen Huang, founder and CEO of Nvidia, sat down with Dwarkesh Patel for a long-form interview that was, depending on how you read it, either a masterclass in technical strategy or a remarkably consistent piece of corporate communications. It was probably both, which is why it was worth taking seriously.
Huang’s core claim could be reduced to a formula:
Input: electrons. Output: tokens. In the middle: Nvidia.
I originally planned seven essays around the interview. Three published pieces later, that structure felt increasingly artificial. The subjects kept widening, from chips to energy, then to institutional knowledge, supplier relationships, inference markets, capital allocation, geopolitics and labour. The later essays risked becoming separate investigations held together mostly by the fact that Huang had mentioned them.
Rereading the first three pieces suggested a simpler thread.
The first asks what happens when the useful unit of performance is no longer a chip in isolation, but a system made of accelerators, memory, networking, software, cooling and power. The second asks where defensibility lives when so much of the underlying technology can be copied, bought or open-sourced. The third asks how to make decisions once value depends on relationships and trade-offs across that wider system.
The same question appears in each case: what is the right unit of analysis?
For Nvidia, the answer has kept expanding. GPU performance matters, but the economics increasingly sit at rack, cluster and data-centre level. CUDA matters, but so do the people, supplier relationships and feedback loops that make the software useful across successive hardware generations. A transaction can be locally attractive and still destroy value if it weakens a relationship the system depends on.
The final essay follows that expansion into the physical economy. Nvidia is now involved not only in chips, racks and networking but in AI-factory design, long-dated supply commitments, land and power availability, and attempts to make AI compute financeable by institutional capital. The interesting question is no longer whether Nvidia is a chip company or a software company. It is how far a company can coordinate a system it does not own.
This leaves four essays, each widening the frame once.
The four entries
1. Nvidia, or the Repricing of a Watt
Read Part 1 → What Huang’s “electrons to tokens” formula says about the economics of AI infrastructure. The paper starts with tokens per watt and the Hopper-to-Blackwell transition, then asks how much of useful performance is determined above the transistor. The October update keeps the thesis but drops an over-confident attempt to decompose the efficiency gain between lithography and system design.
2. The Moat That Cannot Be Coded
Read Part 2 → CUDA as institutional memory, Polanyi’s tacit knowledge, and the parts of competitive advantage that are difficult to reproduce even when the underlying technical ideas are public. The updated version also takes model progress more seriously: tacit knowledge is not permanently outside AI’s reach, and frontier labs are not mechanically condemned to commoditisation. The stronger claim is about privileged learning loops, access and accumulated practice.
3. Pareto Frontiers and Lock-In
Read Part 3 → Why apparently sub-optimal decisions can be rational once the objective function includes time, counterparties and future cycles. Pareto, Markowitz, Williamson and repeated games provide the vocabulary; Nvidia’s pricing decisions and its relationship with TSMC provide the case study.
4. The Return of the Real World
Read Part 4 → The series closes where Nvidia itself has moved: from system design into the industrial and financial conditions required to deploy those systems. Power, land, construction schedules and capital are becoming part of the compute problem. Nvidia is trying to coordinate them without owning the whole stack. That may extend its moat. It also exposes the company to a much larger set of assumptions about demand, utilisation, financing and technological obsolescence.
How to read the series
The four essays can still be read independently. Together, they make one claim that was less clear when the series began:
In AI, an increasing share of strategic value sits in the coordination of complements rather than in any single component.
That is not unique to Nvidia. The same pattern appears higher in the stack, where the usefulness of a model depends on the harness around it: context, tools, proprietary data, permissions, feedback and workflow. Better components matter enormously. They do not remove the problem of putting the system together.
The interview remains the starting document, not an authority. Huang is describing the world from the position of a company that benefits when customers believe optimisation should happen at the system level. The point of the series is to take that claim seriously enough to test it, including where the evidence cuts against Nvidia.
Source
- Jensen Huang, interview with Dwarkesh Patel, Dwarkesh Podcast, 15 April 2026. YouTube.
