NVIDIA China VP: A Single Agent Task Can Require 50-500 Model Calls

On October 22, 2026, Lai Junjie, vice president of engineering and solutions at NVIDIA China, discussed NVIDIA's “token economics” framework in a podcast. He said tokens are shifting from a technical metric to a business metric as AI moves from pilots to scaled deployment, and enterprises need to track token output, efficiency, and business value like a production system.

Lai explained that a token is the basic unit AI models use to process and generate information, and can be seen as the product unit of an AI factory. The cost per token is determined by a data center's total hourly operating cost divided by the number of tokens delivered in real time in that hour, with GPUs, CPUs, networks, storage, and power all affecting overall efficiency.

With agents entering enterprise environments, one agent task may involve 50 to 500 or more consecutive model calls, and multi-agent collaboration further amplifies token consumption, making demand harder to predict. Lai said token value should be assessed by its intelligence content and interactivity, and always in relation to a specific use case—in medical and financial scenarios, for example, accuracy and reliability are hard requirements.

NVIDIA frames token economics around four pillars: value, demand, supply, and monetization. The value pillar asks whether AI output solves a business problem; demand covers usage scale across users and scenarios; supply concerns improving efficiency and lowering cost through models, compute, and infrastructure; and monetization measures whether token usage turns into business returns.

Lai suggested enterprises shift their evaluation from input metrics such as computing investment to output metrics such as business results. The conversation was part of a podcast with Guangzhi Intelligent (光锥智能), an industry media outlet, and also covered token demand forecasting, infrastructure efficiency, and enterprise AI operations.