AGI in a tweet. On September 7, 2026, Jensen Huang posted a sentence on X that was destined to go on repeat. AGI, artificial general intelligence, has arrived. The boss of Nvidia reacted to the launch of GPT-6 Astra, the new flagship model of OpenAI. Le Journal du Coin already knew it for a completely different reason, first classified by OpenAI at the highest level of cyber risk on its scale. This time, the question no longer concerns security but the very nature of the tool. Does Astra think, or complete sentences with extraordinary talent?
GPT-6 Astra blows up the scores, not the AGI debate
On X, Jensen Huang summed up the technical prowess in a handful of numbers. GPT-6 Astra was trained on more than 100,000 Nvidia Grace Blackwell NVLink72 processors, the most powerful generation of chips in the group. The boss of Nvidia recalls the pace: from ChatGPT to o1, then to Astra, in four years. Four hundred thousand additional GPUs are coming online immediately, enough to drive the next one even before its release.
On paper, the beast impresses. OpenAI says Astra is its most successful model ever deployed at scale. But measured performance and general intelligence do not necessarily overlap. A system can excel on a specific set of tests without reasoning like a human when faced with a new problem. This is precisely where the problem lies.
“GPT-6 Astra, trained on 100,000+ NVIDIA Grace Blackwell NVLink72 GPUs. From ChatGPT to o1, then to Astra in 4 years. Artificial general intelligence has arrived. Congratulations to the @OpenAI team Coming soon: 400,000 GPUs in service.”
What is General AI, AGI?
You have already experienced this mechanic without knowing it. Type the start of a sentence into ChatGPT, it completes with the word it thinks is most likely, over and over, until the entire sentence is complete. This is what most researchers still call generative AI: a model trained to predict, word by word, the most probable sequence of a text from a gigantic mass of data. Nothing in this mechanism implies an understanding of the world comparable to ours, nor a will of our own. In other words, Astra guesses better than its predecessors. He doesn't think any better though.
Artificial general intelligence, AGI, is therefore distinguished from generative AI. It would designate a system at least as competent as a human on almost any cognitive task. The problem is that no numerical threshold, no single test has reached consensus to say when this milestone has been crossed. This vagueness suits everyone. Huang can claim victory, his opponents can deny it, and no one is wrong on paper.
Moreover, OpenAI itself does not talk about AGI in its security file on Astra. The company classifies the model at the Critical level of its cyber risk grid, a level never reached before, capable according to it of unearthing previously unknown flaws and exploiting them on systems that are however well protected. A formidable, targeted, measurable capacity. Not proof of universal consciousness.
Why Nvidia has every interest in declaring victory over AGI
Jensen Huang is not a neutral commentator on the ChatGPT-o1-Astra martingale. Nvidia sells the chips that make it possible, and its turnover has doubled to $96.2 billion on the momentum driven by this race for ever more demanding models. The closer the AGI appears, the more reason the tech giants have to order bursts of GPUs by the hundreds of thousands.
The figures cited by Huang are not false. The 100,000 Blackwell chips exist, the next 400,000 too, and Astra displays scores that its predecessors did not come close to. But between noting measurable progress and decreeing the arrival of general intelligence, there is a whole reasoning that the boss of Nvidia carefully avoids. He sells graphics cards, not philosophical definitions. Astra has not changed the definition of AGI. He just changed who wields it.
