OpenAI said GPT-6 Astra was its first model trained using 100,000 GPUs.

A Telegram market roundup attributed to OpenAI’s president the statement that GPT-6 Astra was the first model trained on 100,000 graphics processing units, or GPUs. The same report said the model had crossed an “application threshold,” without supplying a technical definition or benchmark for that phrase.

NVIDIA chief executive Jensen Huang separately said GPT-6 Astra was trained using approximately 100,000 or more NVIDIA Grace Blackwell NVLink72 systems. His statement described the hardware platform but did not provide training duration, energy consumption, cluster location or total cost.

The source material does not clarify whether “100,000 GPUs” refers to the maximum concurrently active GPU count, the installed cluster size or the aggregate number used across the training run. It also does not provide model parameters, training-token count or measured utilization.

Huang stated that the progression from ChatGPT to o1 and then Astra occurred over four years. His additional statement that artificial general intelligence had arrived was his characterization, not a performance result established by metrics in the supplied sources.