With LLMs Exceeding Trillions of Parameters, Are Network Protocols an Overlooked Bottleneck?

In principle, the answer to this question has long been clear: Yes, and it's accelerating. The 'Special Action for Enhancing IPv6 Capabilities of Large AI Models,' launched by the Cyberspace Administration of China (CAC) jointly with local CACs in four regions and five leading large model companies, is essentially clearing the 'last mile' network layer obstacles for the scaled deployment of generative AI. My conclusion is: Full IPv6 support isn't optional; it's a necessary infrastructure upgrade for moving large model applications from the lab to industrialization. Below, I'll argue this from three angles: how network protocols impact distributed training, the inherent flaws of IPv4, and how IPv6 optimizes AI workflows.