Jensen Huang, CEO of Nvidia, wrote four words in a Sunday post on X: “AGI is here.” He credited OpenAI’s new GPT-6 Astra model and immediately added what the model had been trained on. When the head of the company that supplies chips for training virtually all major models announces that machines have achieved general intelligence, it is worth reading the rest of the post as well.
AGI stands for artificial general intelligence. It refers to a system capable of performing a broad range of intellectual work, rather than just one specific task. OpenAI itself defines it as highly autonomous systems that outperform humans at most economically valuable work. However, there is no precise definition of everything a model must accomplish that all researchers would agree on.
Huang presented three things in succession. He described the journey from ChatGPT through the o1 model to Astra as taking four years and pointed out that Astra was trained on more than one hundred thousand Nvidia Grace Blackwell NVLink72 systems. He congratulated the OpenAI team. And finally, he added that another four hundred thousand GPUs—the graphics processing units on which models are trained and run—are about to come online. In a single post, Huang linked his declaration about general intelligence to the announcement that OpenAI’s computing capacity would quadruple. Four hundred thousand processors require a facility, electricity, and money. Opposition to the construction of additional data centers is growing in the United States and elsewhere around the world.
GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years.
— Jensen Huang (@JensenHuang) September 6, 2026
AGI has arrived. Congratulations @OpenAI team.
400K GPUs coming online next.
OpenAI unveiled Astra last Thursday and described it as the most intelligent and best-tuned model in the world, capable of handling even the most demanding professional work. A user can give it a goal and let it complete a series of tasks with minimal intervention. It can search the internet, operate a computer, work with software, write code, and conduct research, and it makes better decisions in situations where it has incomplete instructions.
OpenAI published figures from its own OSWorld 2.0 test, which specifically measures computer use. Astra succeeded in 72.6 percent of the tasks and completed one in approximately forty minutes, while the previous GPT-5.6 Sol model needed around seventy-five minutes for the same task. This represents a 47 percent reduction in time per task.
Access to the model remains limited for now. You have to request it from OpenAI and provide the information required to obtain access. It is also available to subscribers of the ChatGPT Plus, Pro, Business, and Enterprise plans, through OpenAI’s API, and through AWS. It is being rolled out to subscribers gradually, so few people have had firsthand experience with it so far.
People at OpenAI view the term AGI differently
OpenAI’s leadership speaks about AGI more cautiously than its chip supplier. Greg Brockman, president of OpenAI, told reporters on Thursday: “Welcome to the era of AGI.” He added that, in his view, people will one day look back and think that AGI emerged around this time, and perhaps with this very model. Personally, he believes they are already there. However, the company does not explicitly describe Astra as AGI anywhere, and Brockman called it the first step toward it.
Sam Altman described AGI as a term that is not very useful, adding that it is more of a “meaningless marketing term.” On the Sources podcast, he said that it is, at best, a very poorly defined expression. Statements from leading figures in the AI world diverge: Huang says AGI has arrived, Brockman speaks of its beginning, and Altman calls it marketing.
Criticism focused on chips
Huang himself recently said something different. Before Sunday’s post, he said that AGI does not matter and that the term is essentially empty. According to experts, judging whether a system has crossed the threshold of general intelligence is similar to deciding whether a particular painting is a masterpiece.
Companies and their executives have a strong reason to claim that AGI has arrived, even if the same claim is then made again for the next model. Hector does not rule out the possibility that GPT-7 will one day also be described as the debut of AGI. According to him, we will not know when it actually happened until we can look back and identify the point at which models developed the ability to improve themselves.
Who pays whom for what
Demand for the chips that power models such as Astra is reflected in Nvidia’s results. In August, the company reported quarterly revenue of $96.2 billion, more than double the amount from the same period the previous year. Its data center division contributed $89 billion to that result. OpenAI, Meta, Anthropic, and Google all rely on its products.
The two companies’ partnership has already seen one unsuccessful attempt. Last year, Nvidia and OpenAI signed a letter of intent under which Nvidia was to help build at least ten gigawatts of data center capacity for OpenAI and invest up to one hundred billion dollars in the project. At the beginning of 2026, however, Huang explained that this amount had never been firmly committed, and the plan stalled.
A new version of the partnership arrived this February, when OpenAI announced a thirty-billion-dollar investment from Nvidia as part of a larger funding round. It also secured three gigawatts of capacity dedicated to running models and two gigawatts for training on Vera Rubin systems. These are intended to follow the Hopper and Blackwell architectures already in use at Microsoft, OCI, and CoreWeave. In its March funding announcement, OpenAI stated that Nvidia forms the foundation of its infrastructure and that its training fleet and most of its inference layer continue to use Nvidia graphics processing units.
Sources: finance.yahoo.com, businessinsider.com and techradar.com
