Nvidia’s Huang says AGI has arrived. OpenAI has not said it

Nvidia chief executive Jensen Huang said artificial general intelligence has arrived, pointing to OpenAI’s latest model¹. He made the comment on Sunday, days after OpenAI released Astra, its newest and most powerful model⁵, and used it to congratulate the company^3,4.

Huang said Astra was trained on over 100,000 Nvidia Grace Blackwell NVLink72 units². He declared AGI achieved for many practical tasks, dismissed traditional benchmarks, and set out plans for 400,000 AI agents alongside 40,000 human employees^9,10.

What OpenAI itself said

OpenAI shipped GPT-6 Astra to paid subscribers on 3 September, and president Greg Brockman told reporters the AGI bar had been cleared¹³. Astra was released last Thursday⁴.

The distinction is worth holding. The AGI claim has come from Nvidia’s chief executive and from OpenAI’s president in a launch briefing, rather than from any external assessment of the model^1,12.

Brockman called the model a generational leap and declared welcome to the AGI era during a media call¹¹. He used the launch briefing to announce the arrival of AGI, with computer use as the headline feature¹². Even so, the company as an institution has not made the claim that Huang has⁷.

The pushback, and why it matters

The conflict of interest is being stated openly by the outlets covering it. Huang says AGI has arrived, and he sells the machines that built it^7,8. That framing is not a smear, it is the disclosure an investor needs when a supplier certifies the arrival of the demand for its own product.

The technical objection is separate. Gary Marcus wrote that Huang had declared the race to AGI over, and that declaring victory without a definition simply muddies the waters⁶. There is no agreed benchmark against which the claim can be checked, which is precisely what Huang dismissed¹⁰.

The scale numbers attached to the claim are worth reading as a capital-spending statement rather than a scientific one. A model trained on over 100,000 of Nvidia’s own units², and a plan for 400,000 agents working alongside 40,000 people¹⁰, describe the compute Nvidia expects to sell as much as they describe a capability.

Policy is moving in the opposite direction at the same time. In the same week that Astra shipped, Congress moved to ban AGI¹³. For anyone holding the AI complex, the two developments define the risk: a demand narrative certified by the hardware vendor¹, and a legislature beginning to write rules against the thing being announced¹³.