At the 2026 China Unicom Partners Conference held on September 10, ZTE Corporation's President Xu Ziyang delivered a keynote address, highlighting the unprecedented challenges AI brings to network infrastructure. He pointed out that history has recorded three major surges in data traffic. The first occurred in the 1990s, when demand for fixed-line telephony drove the use of Programmable Switching Controllers and TDM technology to manage the explosion in voice traffic. The second wave came in the early 2000s, as IP and optical technologies solved the surge in internet traffic. A third spike emerged during the 4G and 5G mobile era, where LTE technology addressed the massive growth in wireless data consumption. Xu posed a forward-looking question: with the current AI-driven traffic explosion, what network preparations are essential?

He identified three key differences in AI's network demands compared to previous technologies. First, the served audience has shifted. Previously, network traffic was primarily generated by humans. Today, however, nearly 60% of internet traffic is machine-oriented. Moreover, AI agents are evolving from virtual to physical applications, driving a thousand-fold increase in requirements for cloud-edge-device connectivity. Second, the billing metric must evolve. Traditionally, billing was based on bits, but now it must transition to Token-based pricing. Call volumes for Tokens have grown exponentially, increasing a thousand-fold within just one to two years.

Xu emphasized that Tokens from models with different capabilities carry distinct values. He used a literary analogy to illustrate the point: "Li Bai wrote 'Raising my head, I see the moon so bright; lowering my head, I think of home,' that is Li Bai's Token. Its value is clearly different from a Token generated by a mediocre poet." Consequently, when adopting Token-based billing, there will inevitably be an exchange rate gap between Tokens produced by high-value models and those from standard models. He suggested that future telecom operators will need to conduct deeper research into these Token exchange rate differentials when implementing billing systems.

The third distinction lies in evolving demand capabilities. AI applications require frequent context handing-off, which places immense pressure on the uplink and downlink traffic of 5G networks. Specifically, the increase in uplink traffic imposes novel requirements on existing infrastructure. Furthermore, AI's many edge-agent collaborations depend on deterministic networking, which in turn demands extremely low latency. These requirements, Xu concluded, underscore the urgent need for network upgrades to support the AI era effectively.