GPT-6 Astra trained on ~100,000 Nvidia GPUs; Nvidia plans 400,000-GPU scale-up

GG Brief Editorial Desk · AI-assisted, source attributed

OpenAI’s GPT‑6 Astra was trained using roughly 100,000 Nvidia Grace Blackwell NVLink72 server racks. Company leaders describe this as the first training run that used more than 100,000 GPUs and say a substantial portion of the compute went to safety and alignment work. Access to GPT‑6 Astra has begun for a limited set of organizations, with wider subscriber access planned soon. Nvidia plans to scale hardware up to about 400,000 GPUs to support future needs.

• What happened: GPT‑6 Astra’s training used roughly 100,000 Nvidia Grace Blackwell NVLink72 server racks, and OpenAI has started rolling out access to a limited number of organizations.
• Why it matters: Training at this scale required unprecedented compute and included significant GPU use for safety and alignment, indicating large resource commitments behind the model.
• Context: Company leaders characterized this as the first training run exceeding 100,000 GPUs and noted that a sizable share of the compute supported safety and alignment efforts.
• What next: Nvidia and partners plan to bring about 400,000 GPUs online, and OpenAI intends to expand subscriber access to GPT‑6 Astra soon.