NVIDIA (as reported by Reuters on December 3, relevant for December 4, 2025) demonstrated a significant performance leap in its new server solutions. Using 72-chip configurations with high-speed interconnects achieved a 10x speedup in serving Mixture-of-Experts (MOE) models. This improvement is critical for efficiently scaling and reducing the inference cost of modern massive LLMs, such as those from Chinas Moonshot AI and other developers.
NVIDIA: New Servers Boost MOE Model Inference by 10x
