Token原生的Scale Across网络

摘要：传统广域网架构的设计初衷是承载普通流量，并未考虑人工智能（AI）工作负载的特殊性。基于第一性原理，提出一种新的Token原生Scale Across网络，支持各类AI负载在广域网中的传送。该网络采用模型与基础设施协同设计、每任务流量工程以及统一的互联网协议第6版（IPv6）寻址方案。相比传统网络，其容量可扩展能力提升100倍，收敛时间提升10倍，成本效率提升50%。

关键词：模型与基础设施协同设计; 每任务流量工程，统一的IPv6寻址方案

Abstract: Traditional wide-area network (WAN) architectures are originally designed to carry regular traffic, and the specificities of artificial intelligence (AI) workloads are not taken into account. Based on first principles, a new Token-native Scale Across network is proposed to support the transmission of various AI workloads over the WAN. The network is built upon model‑infrastructure co‑design, per‑task traffic engineering, and a unified Internet Protocol version 6 (IPv6) addressing scheme. Compared with traditional networks, its capacity scalability is improved by 100 times, convergence time is reduced by 10 times, and cost efficiency is increased by 50%.

Keywords: model-infrastructure co-design; Per-job traffic engineering; unified IPv6 addressing scheme