AFL：大规模架构人工智能：从训练集群到推理驱动的基础设施白皮书（英文版）

AI infrastructure is evolving beyond a single architectural model. What began as uniform GPU training clusters has expanded into a set of specialized systems aligned to different workload requirements. Training remains important, but inference now accounts for 50–70%1 of total AI compute demand, introducing new considerations for how data centers are designed and built. Rather than representing a single workload category, inference spans high-throughput serving, multistep reasoning, disaggreg