Jensen Huang Reveals 100,000 Nvidia GPUs Trained OpenAI's GPT-6 Astra

OpenAI has begun rolling out access to its latest model, GPT-6 Astra, to a limited number of organizations, with wider subscriber access planned to follow soon. According to Nvidia CEO Jensen Huang via X, the model was trained using approximately 100,000 Nvidia Grace Blackwell NVLink72 server racks. Furthermore, plans are already in motion to bring 400,000 GPUs online in anticipation of future computational needs.

Greg Brockman, the co-founder and president of OpenAI, noted in an interview with Stratechery that this deployment marked the first training run conducted on more than 100,000 GPUs. Alongside the hardware milestone, Nvidia's CEO asserted that Artificial General Intelligence has arrived, a sentiment echoed in discussions regarding whether current or upcoming models cross the threshold for matching or surpassing human capabilities on complex cognitive tasks.

According to OpenAI, GPT-6 Astra represents a state-of-the-art advancement across multiple AI benchmarks, including FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0. Promotional materials display conversational spoken prompts capable of assisting with tasks such as creating presentations, drafting online marketplace listings, and generating 3D printer files. OpenAI describes Astra as its most aligned model, with Brockman emphasizing that a substantial portion of the system's compute power and security work has been dedicated to safety and alignment.

Access to GPT-6 Astra remains limited during the initial rollout phase. The debut arrives amid circulating industry rumors regarding an upcoming initial public offering for OpenAI, as well as following recent developments surrounding isolated testing environments.