What NVIDIA’s chief actually said
On September 6, Jensen Huang said GPT-6 Astra had been trained on more than 100,000 NVIDIA Grace Blackwell NVLink72 systems, congratulated the OpenAI team and wrote: “AGI has arrived.” He also said another 400,000 NVIDIA GPUs would subsequently be brought into operation. [1 · Jensen Huang]
Business Insider confirmed the post and noted that OpenAI president Greg Brockman also describes Astra as the beginning of the artificial general intelligence era. OpenAI itself, however, speaks of a major leap in the model’s capabilities without offering a universal, testable definition of AGI. [2 · Business Insider] [3 · OpenAI]
Why the statement cannot be treated as proof
Huang’s words are confirmed, but they do not establish the existence of artificial general intelligence. The industry has no single accepted procedure after which a model earns that status; even strong results in programming, cybersecurity and computer use measure individual abilities, rather than general intelligence. [2 · Business Insider] [3 · OpenAI] [4 · Gary Marcus]
Researcher Gary Marcus separately noted that the statement contains neither a definition of AGI nor a set of criteria for testing the conclusion. The accurate editorial formulation is therefore “Huang called the model AGI,” rather than “NVIDIA or OpenAI created AGI.” [1 · Jensen Huang] [4 · Gary Marcus]
Sources
- Jensen Huang — post about GPT-6 Astra — September 6, 2026
- Business Insider — Huang declares the arrival of AGI — September 7, 2026
- OpenAI — GPT-6 Astra — September 3, 2026
- Gary Marcus — criticism of the AGI claim — September 6, 2026
Expert commentary
The post also serves as a statement about infrastructure scale. Mentioning more than 100,000 systems used in training and another 400,000 accelerators in the next wave links growth in model capabilities to further demand for NVIDIA hardware. [1 · Jensen Huang] [2 · Business Insider]
The AGI claim should be tested through independent, reproducible evaluations, consistent performance beyond prepared tasks and the model’s ability to transfer knowledge between fields without specialized adaptation, rather than through corporate wording. [3 · OpenAI] [4 · Gary Marcus]