NVIDIA CEO Says GPT-6 Astra Marks the Arrival of AGI; Gary Marcus Points to Lack of Definition and Evidence

On September 6, 2026, NVIDIA CEO Jensen Huang posted on X that OpenAI's latest model, GPT-6 Astra—trained with more than 100,000 Grace Blackwell NVLink72 systems—marks the arrival of AGI. He also revealed that an additional 400,000 GPUs are gradually coming online. AI researcher Gary Marcus wrote on Substack the same day, arguing that the announcement provided neither a definition of AGI nor quantitative evidence, and that by agidefinition.AI's ten criteria, Astra satisfies only about two.

Factual Account

Huang's initial post on September 6 mentioned 300,000 systems; he later deleted it and revised the figure to more than 100,000 without explanation. OpenAI President Brockman promptly echoed the claim on the same platform, but did not directly confirm that AGI had arrived. Gary Marcus cited agidefinition.AI's standards in his rebuttal and noted that Huang had made a similar claim about another model within the previous six months. Multiple media outlets covered the episode, highlighting the tension between commercial interests and independent evaluation.

Mechanism Analysis

Huang's claim is directly tied to NVIDIA's hardware sales. NVIDIA has shifted from a design with eight chips per board to integrated NVLink72 systems combining 72 chips—a transition requiring billions of dollars in investment that can be recouped only through large-scale training. GPT-6 Astra is described as having been trained on more than 100,000 such systems. Gary Marcus's rebuttal focuses on the missing definition and insufficient evidence, arguing that capability claims need to correspond to quantifiable criteria rather than resting solely on declarations by a chipmaker.

Industry Impact

For the competitive landscape, this claim may accelerate follow-up moves by other chipmakers and AI labs, but it also raises questions about commercially motivated announcements. Across the supply chain, NVIDIA, as the GPU supplier, stands to benefit from growing demand, while developers need to assess whether the model's actual capabilities match the claims being made. Enterprise users face a selection risk: relying on a single vendor's statements could lead them to overlook discrepancies in third-party benchmarks.

For developers, independent evaluations should take priority over claims from a single source. When selecting models, enterprises should ask for quantitative results that meet multiple standards such as those from agidefinition.AI, and compare the scale of training hardware against actual task performance.

Comparisons and Precedents

Huang had already made a similar AGI claim about another model within the previous six months, and this repetition has drawn scrutiny over the consistency of the evidence. The reports also mention that a $100 billion partnership announced in 2025 was ultimately never signed, with OpenAI instead reducing its procurement of NVIDIA chips—a situation that contrasts with the current claim.

Strategic Assessment

Based on the available facts, the most likely development is that the market will continue to focus on the rollout progress of the 400,000 GPUs rather than on the AGI label itself. This assessment follows from the direct relationship between hardware delivery data and the claim.