On September 6, 2026, Nvidia CEO Jensen Huang posted on X (formerly Twitter) the statement 'AGI has arrived' and congratulated OpenAI on the release of the GPT-6 Astra model, which was trained on more than 100,000 Nvidia Grace Blackwell accelerators at the Stargate site in Texas. OpenAI itself did not use the term AGI in its official announcement: the model, released on September 3, 2026, is currently available in a limited mode, and the milestone was declared by the hardware vendor on which it was trained. Nvidia also announced the connection of another 400,000 accelerators, and Huang's post gathered 2.6 million views in the first hours.

image

What happened

The starting point was the release of GPT-6 Astra: OpenAI released the model on September 3, 2026, in limited access and describes it as its most intelligent and balanced model with best-in-class results in computer use, software engineering, cybersecurity, science, and professional tasks. Three days later, on September 6, 2026, Nvidia CEO Jensen Huang posted on X (formerly Twitter) the phrase 'GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years' and the separate conclusion 'AGI has arrived'. In the same post, he clarified that 400,000 Nvidia accelerators 'will come online next', meaning that the computing power on which Astra was trained will be followed by a new infrastructure expansion. Business Insider recorded the industry's reaction in a separate article about the Nvidia CEO declaring the arrival of AGI and congratulating OpenAI.

Context

The infrastructure base of the statement is verifiable: Grace Blackwell is Nvidia's server platform combining Blackwell GPUs and Grace CPUs, and NVLink72 is a high-speed communication fabric that turns 72 accelerators in one rack into a single computing machine; more than 100,000 such GPUs are deployed at the Stargate site in Texas, where Astra was trained. The four-year path from ChatGPT (November 2022) through o1 to Astra, which Huang cited as evidence, is a trajectory of increasing model autonomy that has been discussed in the industry for years. At the same time, the term AGI still does not have a generally accepted operational definition, and OpenAI avoided it in its official announcement, describing the model through specific capabilities and access conditions. Huang's post should be read with a correction for the author's position: Nvidia is the largest GPU supplier, directly earning from the scaling of training, and its announced expansion of computing power is synchronized with the release of a model trained on its equipment.

Why this matters for the industry

For the industry, this is primarily a signal of a new computing economy: the frontier layer, where models are trained on 100K+ GPUs, is consolidating around a few players, and Nvidia publicly confirms this scale, promising the next expansion. The term AGI is becoming established in the industry's marketing language: its official use by the Nvidia CEO raises the bar of expectations for competitors (Anthropic, Google) and within OpenAI itself, increasing pressure to release comparable agentic systems. The 'AGI has arrived' narrative, recorded in a viral post, is highly likely to raise corporate AI budgets and investor interest in agentic systems, temporarily reducing the cost of market entry for startups with agentic products. The window of opportunity for startups is shifting from training foundational models to the application and agent layer: computer use, multi-step software engineering, and professional tasks. For product teams, Astra is a preview of the roadmap, not a launch: the meaning of the signal is what capabilities will soon appear in accessible models, and a reason to adjust plan timelines in advance.

Why this matters for users

Directly for readers, almost nothing changes today: Astra is released in limited access, and OpenAI has not published pricing, latency, or API details, so there is no reason to restructure current workflows. The practical benchmark is to monitor OpenAI's official announcement for expanded access, the appearance of an API, and pricing. Two signals will help assess the actual level of the model: independent verification of the claimed results (98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3) and the actual connection of the new Nvidia-promised computing power, on which the scale of the next models depends. If access is expanded, the first noticeable products based on such models will be agents that close end-to-end development tasks, and application automation without APIs through computer use. If independent checks confirm the claimed level, this episode will go down in history as a factual milestone; if not, 'AGI' will be firmly established as a marketing word.

What is still unknown / limitations

The statement 'AGI has arrived' is methodologically an assessment by an interested party — the CEO of a company that sold accelerators for training Astra — and not a technical verdict. The claimed benchmark results for Astra were obtained in limited access and have not yet been independently reproduced: without a description of the methodology, this is data, but not yet proof. The original version of Huang's post mentioned 300,000 GPUs, then the scale was corrected to approximately 100,000 — a sign that one should rely on the corrected version and OpenAI's announcement. The timing of expanded access, the conditions for the appearance of a public API and pricing, and the details of the restrictions on the model's cybersecurity capabilities that accompanied the release remain unknown.

Sources

Author

Look at AI, editorial team