Jensen Huang Explains Why the Token Factory Is AI’s Ultimate Form

On March 24, 2026, Nvidia founder Jensen Huang participated in a 150‑minute deep interview with Lex Fridman, discussing corporate strategy, AI compute architecture, agent technology, and the surrounding industry ecosystem.

Huang recapped Nvidia’s development path, noting that the company began with dedicated accelerator chips, then added programmable pixel shaders and FP32 compute units, and finally integrated CUDA across the entire GeForce GPU line—a series of strategic decisions that broadened Nvidia’s reach.

He introduced the Vera Rubin rack, a system designed specifically for agents. According to Huang, agents need tool‑calling capabilities, and their architectural requirements differ fundamentally from the previous generation of expert‑mixing models and inference‑focused systems.

Based on this evolution, Huang argued that the definition of a data center must be revised. Traditional data centers served information retrieval; generative AI shifts computation toward real‑time information generation, turning data centers from mere storage warehouses into “Token factories”.

In the token‑factory model, tokens become the core production element and value carrier. Huang predicts that as token production scales and standardizes, it will resemble the iPhone product line with tiered pricing and differentiated commercial value, eventually enabling Nvidia to reach a valuation of $10 trillion.

This shift also changes the compute proposition: the focus moves from “building stronger chips” to “building better systems” that can generate tokens at scale, directly linking compute power to industry value.

Finally, Huang outlined the “Agent Scaling Law”, which comprises four interdependent stages—Pre‑Training, Post‑Training, Test‑time, and Agent—that form a compute‑driven closed loop, and he highlighted a structural transition in compute demand toward supporting large‑scale agents.