Co-design the Whole System

On the Lex Fridman Podcast (#494, March 2026) Huang explained why Nvidia stopped being a chip company in practice. Once a model is too big for one GPU, speeding up a single part stops helping. He pointed to Amdahl’s law: “the amount of speed up you have for something depends on how much of the total workload it is.” The answer is to optimize everything at once: “across the entire stack of software from architectures to chips, to systems.”

The mechanism. He said the Rubin design “was completely focused on doing one thing, processing the LLM.” On suppliers, he said they delivered “Because I told ‘em what I needed. They understood what I need.” In other words, Nvidia specifies the whole system, and the supply chain builds to it.

The commitment. Nvidia’s FY2026 10-K reports outstanding inventory purchase and long-term supply and capacity obligations of $95.2B.

Doesn’t transfer without leverage over suppliers. “I told them what I needed” works when you are the customer every supplier wants. A small firm gets the standard catalogue. And the $95B of supply commitments that make co-design possible are also a large bet that demand holds.