NVIDIA CEO Jensen Huang mentioned the training infrastructure of OpenAI's latest model, "GPT-6 Astra," on his official X account. Huang revealed that the model was trained by clusters of over 100,000 "NVIDIA Grace Blackwell NVLink72" and congratulated the OpenAI team, stating that "AGI (Artificial General Intelligence) has arrived."
For the training of GPT-6 Astra, GB200 NVL72 units equivalent to a scale of 1,400 racks were used. This system densely integrates 36 Grace CPUs and 72 Blackwell GPUs, designed to make the entire rack function as a single logical accelerator.
Huang also touched upon the fact that the next "400,000 GPUs" are scheduled to go operational. With the securing of massive computing resources, the importance of physical infrastructure challenges in data centers, such as power supply and heat dissipation management, is increasing.
Source:
- AGI一歩手前のAstraは1,400本のラックスケールシステムGB200 NVL72でトレーニング（想像図あり）。AIデータセンター・インフラ案件として世界最高峰：経営者が読むNVIDIAのフィジカルAI / ADAS業界日報 by 今泉大輔 - オルタナティブ・ブログ (Google News: Data Center Power, 2026-09-09)