NVIDIA has made a significant breakthrough in AI performance, particularly with Mixture of Experts (MoE) models, achieving a 10x increase in performance using its GB200 NVL72 AI cluster. This advancement comes from leveraging co-design performance scaling laws. The GB200 configuration, featuring a 72-chip setup and 30TB of fast shared memory, allows for expert parallelism, meaning only a portion of the model's parameters is activated based on the specific query, reducing the computational load. The performance leap was demonstrated with the Kimi K2 Thinking model, which has 32 billion activated parameters per forward pass. NVIDIA’s new architecture not only handles token batches more efficiently but also enhances communication between GPUs, addressing a major bottleneck in scaling MoE models. The use of the NVIDIA Dynamo framework further optimizes task management by distributing workloads across GPUs, allowing for better resource utilization. NVIDIA's progress is timely, as many frontier AI models increasingly rely on powerful AI servers. MoE models are becoming more popular due to their computational efficiency, making NVIDIA’s advancements particularly relevant in today’s rapidly evolving AI landscape. The firm's ability to enhance these models positions it well to capitalize on the growing demand for scalable AI solutions.