NVIDIA CEO Jensen Huang said GPT-6 Astra was trained on more than 100,000 NVIDIA Grace Blackwell GPUs, adding that another 400,000 GPUs are “coming online next” as OpenAI continues to expand the compute infrastructure behind its frontier models.
Huang made the comments in a post on X on Monday, saying: “GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72.”
Huang also said, “From ChatGPT to o1 to Astra in 4 years. AGI has arrived,” congratulating OpenAI.
The AGI claim is Huang’s assessment, and OpenAI has described Astra as its most capable and aligned model, and president Greg Brockman also said this marks a ‘welcome to the AGI era’.
It can operate software, browse the web, handle spreadsheets and documents, write and debug software, conduct research and work on scientific and mathematical problems.
OpenAI said Astra scored 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench.
OpenAI conducted the training run at its Stargate site in Texas. The vice president of research Aidan Clark told Fortune that Astra was “the first time we’ve pretrained on more than 100,000 GPUs at our Stargate site in Texas.”
NVIDIA Grace Blackwell NVLink72 combines 36 Grace CPUs with 72 Blackwell GPUs in a liquid-cooled rack, with all 72 GPUs connected through a single fifth-generation NVLink domain.
The system provides 13.4TB of HBM3E GPU memory and 130TB/s of NVLink bandwidth.
That means the 100,000-plus GPUs referenced by Huang would represent roughly 1,389 GB200 NVL72 racks if expressed purely as 72-GPU rack equivalents. NVIDIA has not disclosed the actual deployment configuration.
The next generation is Vera Rubin NVL72, which combines 72 Rubin GPUs with 36 Vera CPUs and NVIDIA’s sixth-generation NVLink. NVIDIA says Vera Rubin is already in full production and is being positioned for large-scale agentic AI and reasoning workloads.
The platform is designed to reduce the GPU requirement for training while substantially lowering inference costs compared with Blackwell.
Your reaction