Everybody writing about AI wants to write about the models. Stephen Witt wrote about the chips, and the result is the more illuminating book, because the constraint that has shaped this entire industry is not algorithmic cleverness, it is who can get compute and at what price. Nvidia became the most valuable company on earth in June 2024, thirty one years after being founded in a Denny's, and the path from graphics cards for video games to the substrate of the AI economy is genuinely strange. Witt had unusual access, to Jensen Huang himself, to early investors, to employees, and he uses it for texture rather than hagiography, mostly.
Huang emerges as relentless and difficult in specific rather than generic ways, a manager with an unusually flat organisation and an appetite for detail that would be intolerable to work for and appears to have been decisive. The single most valuable thing in the book, and the thing most accounts of this period get wrong, is the emphasis on CUDA. Nvidia's dominance is routinely explained as having better chips, and better chips are necessary but nowhere near sufficient. CUDA is the software layer that made those chips programmable for general computation, released years before there was an obvious market for it, and it created a decade of accumulated tooling, libraries and trained engineers that competitors cannot replicate by matching a specification.
That is a story about a long bet on developer ecosystem paying off spectacularly, and Witt tells it properly. Anybody trying to understand why AMD's competitive hardware has not translated into competitive position should read that section. The writing is good, which should not be remarkable and is. Witt wrote How Music Got Free, and the same ear for the detail that carries a scene is present here.
Technical material is explained without condescension and without the false precision that afflicts journalists writing about semiconductors. It won the Financial Times and Schroders Business Book of the Year for 2025. The structural weakness is the one inherent to access journalism, and it should be stated even though the book is good. When your subject grants unprecedented access, the resulting account tends towards their framing of events, and the questions that would be uncomfortable tend not to be the ones that get several chapters.
Nvidia's pricing power, its allocation decisions during the shortage, its relationships with the customers most dependent on it, the antitrust questions circling all of that: these are present but light. A more adversarial book would have been less readable and more complete. Competitors are similarly thin. The case that Nvidia's position is more contingent than it looks, that inference workloads may not favour the same architecture as training, that the hyperscalers building their own silicon represent a genuine long-term threat, all of it gets less attention than it warrants.
The energy and supply chain consequences are barely touched, and those are becoming the story. And the ending is a problem no author could solve. The book stops at a moment that already reads as a midpoint rather than a conclusion, which is the occupational hazard of writing about a live subject. Four point two.
Original access used well, the clearest available explanation of why software ecosystems beat hardware specifications, and the right book to read first if you want to understand the physical and commercial substrate under everything else in this field. Marked down for the sympathy that access buys and for leaving the case against its subject largely unmade.