The weights and the machines

The chain, as he retells it: Nvidia designs the GPUs but TSMC in Taiwan actually fabricates them ▸ 7:31; the machines TSMC fabricates with come from ASML ▸ 8:46; and ASML in turn depends on the one company that grinds those lenses and the one company that makes the best lasers ▸ 9:15. The pattern repeats at every link: “es solo uno, hiperespecializado, que es el único que produce esa vaina” ▸ 9:37. Hence the card that cost $400 now costing $2,000 ▸ 8:26, and hence Intel, which once fabricated its own chips, now buying from TSMC like everyone else ▸ 10:15. When ASML considered expanding abroad, the Dutch government paid it to stay ▸ 19:02; nations understand what kind of asset a sole recipe-holder is.

los pesos se copian con un ctrl+c; las máquinas no →

Why is the moat unassailable? Not secrecy alone: “yo puedo tener los planos detalladísimos de cómo hacer las máquinas estas” and still need something like three years with a top team to reproduce the process, by which time the incumbent is five years further on ▸ 14:39. The machine isn’t a printer you can reason about; it’s a building-sized room of lasers and optics ▸ 11:49, decades of accumulated craft where every step is tacit ▸ 11:24.

And then the contrast that makes this a builder’s lesson rather than trivia: software sits at the other pole of copyability. A neural network is, in the end, a pile of parameter values; hold the copy and you can run it anywhere, have the architecture and you can retrain it ▸ 14:04. The most valuable artifacts of the AI era are the easiest artifacts in history to replicate, which is why the industry’s deepest moats keep turning out to be the physical ones underneath, the fabs and the machines nobody can copy even with the plans…