OpenAI’s forthcoming GPT-5 large-language model was trained on NVIDIA’s current-generation H100 and H200 graphics processors and will be deployed on the new GB200 NVL72 artificial-intelligence server, according to information shared by NVIDIA’s data-center division. The GB200 NVL72 clusters 72 Blackwell-architecture GPUs and 36 Grace server CPUs in a single rack, interconnected by NVIDIA’s high-speed NVLink and NVLink-Switch fabrics. The configuration is designed to accelerate trillion-parameter models, offering what NVIDIA says is a significant jump in throughput and energy efficiency over earlier Hopper-based systems. Amazon Web Services plans to make similar multi-GPU configurations available through its SageMaker HyperPod service, giving customers access to the same infrastructure used for GPT-5. NVIDIA says its CUDA software stack has now been downloaded more than 450 million times, underscoring the company’s dominance in the market for AI computing silicon.