Reading Nvidia: On strategy, defensibility, and capital in the AI era

A few weeks ago, Jensen Huang, founder and CEO of Nvidia, sat down with Dwarkesh Patel for a long-form interview that, depending on how you read it, was either a masterclass in technical strategy or a remarkably consistent piece of corporate communications. It is in fact probably both at once, and that ambivalence is what makes the conversation worth taking seriously. It also raised several itches as I listened to it, which is what eventually grew into this series on strategy.

Huang’s core claims can be summarised with a formula. Input: electrons. Output: tokens. In the middle: Nvidia. The line lands like a marketing slogan. Read carefully, it is also a thesis about where AI value will accumulate over the next decade, and a position paper on what kind of company Nvidia intends to be in that future. The interview is the document this series annotates, pushes back against, and uses to open onto questions that go further than any single company.

Seven entries, organised around three concerns. The first concern is strategic: what kind of moats actually hold in an environment where models become commodities, where talent circulates, and where every architectural choice is a public roadmap. The second is economic: what happens to the structure of value when the build-out forces the tech industry back into industrial capex on a scale most of its current operators have never experienced. The third is societal: what an economy where AI is very strong does to labour, education, and the question of what makes a human valuable in the first place.

As usual at Libido Sciendi, we read the present standing on the shoulders of giants. Helmer on process power, Polanyi on tacit knowledge, Saxenian on regional advantage, Pareto and Markowitz on multi-dimensional optimisation, Williamson on transaction costs, Axelrod on repeated games, Porter on trade-offs, Teece on dynamic capabilities, Bowker and Star on classification as politics. Each of these thinkers has spent a career on a question that current analyst conversations on Nvidia tend to read too narrowly. We bring them in not to produce a complete academic treatment, but because their concepts cut through the noise.

A short note on what the series is not. It is not investment advice. It is not a forecast on Nvidia’s stock price. It is not an attempt to settle the question of whether Huang is right or wrong about the long-term durability of his position. The empirical question of who wins the AI infrastructure layer will be settled by Vera Rubin, Feynman, and the next two product cycles, not by analysis. What this series tries to do is offer better tools for thinking through what we are watching while it unfolds.

The seven entries

1. Nvidia, or the Repricing of a Watt

Read Part 1 → — What Jensen Huang’s mental model of Nvidia tells us about where AI value actually accumulates. The thesis paper of the series. Why the bottleneck of AI is shifting from compute to energy, why tokens-per-watt is becoming the binding benchmark, and why most of the gain in tokens-per-watt comes from system design rather than from lithography. With the arithmetic of the Hopper-Blackwell transition and three strategic implications.

2. The Moat That Cannot Be Coded: Nvidia, Frontier Labs, and Defensibility in the 21st Century

Read Part 2 → — Why tacit knowledge wins over codified knowledge, and what that means for who captures AI value. CUDA as institutional knowledge. Why frontier labs are structurally constrained at the top of the stack and will not subsume vertical AI. Why the United States has a structural advantage rooted in talent mobility. And why the rare companies that maintain dynamic capabilities tend to share two structural traits: an anchor in physical objects, and a living founder transmission.

3. Pareto Frontiers and Lock-In: When Mathematical Sub-Optimisation Is the Right Strategy

Read Part 3 → — What the inference market and Nvidia’s no-surge-pricing tell us about strategic decisions in multi-dimensional spaces. Why the same handful of companies keep showing up as durable winners not because they are smarter at the optimisation, but because they have correctly specified the function being optimised. With illustrations from the segmentation of inference and the absence of a formal contract between Nvidia and TSMC over thirty years.

4. The Return of the Real World: When the Boundaries Between Operators and Investors Erode

How the AI build-out is fusing corporate strategy with infrastructure capital, and what that says about a new economic landscape. Why hyperscalers are behaving more like industrial conglomerates than like software companies. Why infrastructure funds are becoming structural actors in the tech stack. Why CEOs are revealing themselves as asset allocators in a way that was always true but is now visible. The most ambitious entry of the series, on the institutional transformation underway.

5. The Job and the Task: What Will Make Us Valuable in the Age of Strong AI

Why the question is no longer which skills to acquire, but how to build a job that is not reducible to a sum of tasks. Starting from Huang’s distinction between the radiologist’s job (patient care) and her task (reading a scan), the paper builds a conceptual framework for thinking about what kind of human work survives an economy where AI does most of the tasks. With detours through Polanyi, Bourdieu, and Schön, and a closing argument on what the post-task economy says about how we should think about competence, knowledge, money, time, and the value of a human life.

6. A Note on China: The Limits of Lithographic Containment

What the bottleneck inversion thesis says about US export controls, Chinese AI capacity, and the gap that policy assumes versus the gap that actually exists. With the three marginal arguments Huang makes against the Dario Amodei position, and an honest critique of the concern Huang refuses to engage.

7. Drawing the AI Stack: Huang, Social Capital, and the Politics of Maps

Why every taxonomy of AI is a strategic act, and what comparing two of them reveals about where value actually accumulates. The closing entry of the series, which steps back from Nvidia specifically to ask the meta-question that has run through all of the others. With a comparison of Huang’s five-layer cake and Social Capital’s Physical/Digital stack, and a connection to our broader thread on classification as power.

How to read the series

The seven entries are designed to stand alone. A reader can pick any one of them as a single piece without having read the others. But they are also designed to compose. The thesis paper (entry 1) sets up the analytical frame. Entries 2 and 3 examine the two foundations of Nvidia’s durability (institutional knowledge as moat, rational decision-making in multi-dimensional environments). Entries 4 and 5 widen the camera to the broader economic and societal transformations underway. Entries 6 and 7 close the series, the first by addressing the geopolitical question explicitly, the second by stepping back into a meta-reflection on how we read these markets at all.

Source

• Dwarkesh Patel, interview with Jensen Huang, Dwarkesh Podcast, 15 April 2026 — YouTube.