.svg){kind=spacer}
Details
- Leading AI models such as Kimi K2 Thinking and DeepSeek-R1 are achieving remarkable performance improvements using NVIDIA’s new mixture-of-experts (MoE) system.
- The GB200 NVL72 integrates 72 Blackwell GPUs with 36 Grace CPUs in a rack-scale system, addressing memory bandwidth and latency obstacles that have challenged AI model scaling.
- More than 60% of open-source models released in 2025 are anticipated to adopt MoE architecture, reflecting its rapid surge in popularity.
- The system unlocks dynamic computational routing, supporting advanced multimodal and specialized AI workloads.
- This MoE-driven leap in capability has fueled a 70-fold increase in AI model intelligence since early 2023.
Impact
NVIDIA's GB200 NVL72 ushers in a new era for AI infrastructure, dramatically slashing compute demands while supercharging the intelligence of modern models. The shift to MoE as the prevailing architecture signals a foundational change in deploying advanced AI across a wide range of industries, setting a new benchmark for efficiency and scalability.
