Nvidia CEO Jensen Huang has made one of the strongest claims yet about the state of artificial intelligence: “AGI has arrived.”
His statement came after OpenAI launched GPT-6 Astra, its latest frontier model, on September 3. OpenAI says Astra has reached 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench.
But while those numbers show a major jump in capability, Huang’s statement remains a judgment, not a universally established declaration that artificial general intelligence has been achieved.
Why Astra Is Different
Astra is not being positioned simply as a better chatbot. OpenAI says it can use computers, browse the web, write and test software, analyse scientific data and handle multi-step professional tasks.
It can also create documents, spreadsheets and presentations while adapting when requirements change.
The computer-use numbers are particularly notable. On OSWorld 2.0, Astra scored 72.6%, compared with 65.7% for GPT-5.6 Sol. OpenAI says its latency simulations showed Astra completing the higher-performing tasks in about 47% less time per task.
That shift matters because the value of these models is increasingly being measured by what they can actually do, rather than simply how well they can answer a question.
Jensen Huang’s AGI Declaration
It was against this backdrop that Huang declared that AGI had arrived.
According to reporting on his statement, Huang said Astra had been trained using more than 100,000 Nvidia Grace Blackwell GPUs and that another 400,000 GPUs were coming online. The scale is significant because increasingly capable AI models require not just better algorithms but enormous amounts of computing capacity.
Still, the AGI label needs some caution. There is no single industry-wide test that determines when a model officially becomes artificial general intelligence. OpenAI itself describes Astra in terms of specific capabilities and benchmark performance rather than announcing that AGI has been formally achieved.
So Huang’s comment should be read as his assessment of how far AI systems have progressed, not as an independently certified scientific milestone.
Cybersecurity Raises Stakes
One of Astra’s more consequential achievements is outside the usual chatbot benchmarks.
OpenAI says Astra is its first model to reach the Critical level of cybersecurity capability under its Preparedness Framework. With the right tools and access, the company says Astra can identify previously unknown security flaws and develop new ways to exploit them across well-protected systems without someone directing every step.
That capability is also why OpenAI introduced stronger safeguards around the model. The company says Astra has additional monitoring and protections against harmful cyber actions, while its system card also flags challenges around monitoring the model’s reasoning behaviour under adversarial conditions.
For businesses, this makes the Astra rollout about more than productivity. The same capabilities that could help security teams can also create new risks if powerful models are misused.
AI Infrastructure Race Intensifies
The infrastructure story is already moving beyond announcements.
On September 8, Reuters reported that Nvidia-backed Firmus had signed a multi-year agreement with OpenAI to supply computing capacity from two data centres in Malaysia. Firmus said the deal would take its contracted capacity across customers above 900 megawatts.
That is a useful indication of where the AI race is heading. Training and running frontier models increasingly depends on securing data-centre capacity, processors and power, not just developing the model itself.
For Nvidia, this creates a direct link between advances in AI capability and demand for its infrastructure.
Rollout Was Not Smooth
The Astra launch itself had a rocky start.
After OpenAI introduced the model on September 3, some paying users were unable to access it as expected. Sam Altman apologised, calling the rollout “messy”. By September 4, he said Astra was available to Pro, Enterprise and Business Premium users in Work/Codex and through the API, with rollout to Plus and other users continuing.
That detail is worth noting because the real test for Astra will not be the phrase “AGI has arrived”. It will be whether the model can reliably perform useful work at scale, while companies can manage the cost, security and operational risks that come with giving AI systems more autonomy.
For now, the evidence points to a meaningful capability jump. Whether that jump deserves the AGI label is still a separate question.
