Inference has overtaken training, and the flywheel runs on tokens

The idea

NVIDIA's fiscal 2026 letter sets out how its founder thinks about the economics of AI. Inference, running models to answer questions, has overtaken training as the dominant workload, and the letter describes a loop: capable agents generate more tokens, more tokens need more compute, more compute produces more capable models and more demand. The twenty-year-old software platform is presented as one of the deepest moats, partly because improvements reach the entire installed base and keep older products valuable. The letter cites visibility into more than a trillion dollars of cumulative revenue from two product generations. Read beside the Hayden letter, it is the supply side of the same argument about whether the spending earns its return.

A compute vendor's case is a flywheel from agents to tokens to compute; the buyer's case is the return on that compute, and both have to be read together.

How to use it

Pair the two letters: list NVIDIA's flywheel claims on one side and the reinvestment-return equation on the other, and ask which assumptions have to hold for both to be right. The gap is where a thesis lives.