NVIDIA Sets World Record for DeepSeek-V3 Pre-Training on GB300 NVL72

DeepSeek-V3 is a large language model of the Mixture of Experts (MoE) type with 671 billion parameters, designed for high efficiency in training and inference.

GB300 NVL72 is an NVIDIA infrastructure connecting 72 GPUs in a single node, optimized for training large MoE models.

It reduces training time and infrastructure costs, accelerating adoption of advanced AI models in government and private sectors across the region.

It represents extremely high compute efficiency, meaning faster training and better resource utilization, a key benchmark for enterprises planning AI investments.