At the AI Infra Summit, NVIDIA highlighted new collaborations, including Amazon's Annapurna Labs working on NVHBM custom high-bandwidth memory technology and d-Matrix integrating NVLink Fusion with Raptor XPUs. Lambda achieved a 23% performance-per-watt improvement and a 24% cluster-wide token throughput increase with DSX MaxLPS. Pinterest is leveraging the Blackwell platform for conversational AI. NVIDIA emphasizes optimizing "tokens per megawatt" as the new critical metric for AI infrastructure efficiency.