Jensen Huang says 100,000 Nvidia GPUs were used to train OpenAI’s latest model, GPT-6 Astra, and there’s already plans to bring quadruple that amount of hardware online
OpenAI has begun rolling out access to its latest large language model, GPT-6 Astra, with limited organizational access already available and broader subscriber access expected soon. According to Nvidia CEO Jensen Huang, the model was trained using approximately 100,000 or more Nvidia Grace Blackwell NVLink72 server racks—a massive computational milestone representing the first time OpenAI has trained a model on such a scale. Greg Brockman, OpenAI’s co-founder and president, confirmed this was “the first run that we’ve trained on more than 100,000 GPUs.”
A significant portion of the computational resources was dedicated to safety and alignment work, reflecting OpenAI’s continued investment in responsible AI development. Looking forward, Huang revealed that plans are already underway to expand the GPU infrastructure to 400,000 units, signaling dramatically increasing computational demands for future model iterations. This four-fold expansion underscores the rapidly scaling infrastructure requirements for frontier AI model training and hints at even more ambitious projects in development.