Jensen Huang: GPT-6 Astra Trained on ~100K+ Nvidia Grace Blackwell NVLink72 (twitter.com)

Jensen Huang, CEO of NVIDIA, announced the launch of GPT-6 Astra, a groundbreaking AI model trained on over 100,000 NVIDIA Grace Blackwell NVLink72 GPUs. This model represents a significant leap in capabilities, demonstrating state-of-the-art performance on various benchmarks, including FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0. Huang highlighted the rapid evolution from ChatGPT to this latest iteration in just four years, suggesting that we are on the cusp of achieving Artificial General Intelligence (AGI). The implications of GPT-6 Astra extend beyond impressive benchmarks; it is poised to accelerate scientific discovery as evidenced by its leading performance on Terminal-Bench Science 0.1 and HealthBench Pro. Moreover, Huang proposed a novel approach to democratize access to AI infrastructure by suggesting the tokenization and fractional ownership of the 400,000 GPUs that are coming online. This could enable broader participation in AI development and facilitate funding for operational costs, fostering a more inclusive AI environment.