Nvidia's Jensen Huang Declares AGI Has Arrived With GPT-6 Astra as OpenAI's Chief Scientist Urges Voluntary Slowdowns

Two claims about the same model emerged within a day of each other, and they point to opposite conclusions. Nvidia chief executive Jensen Huang posted on X that OpenAI's GPT-6 Astra was trained on roughly 100,000 or more Nvidia Grace Blackwell NVLink72 systems, that the path from ChatGPT to o1 to Astra took about four years, and that "AGI has arrived," adding that 400,000 more graphics processing units are coming online next. The Russian-language channel AI ML Big Data relayed the claim with the same compute figures. Reporting on the post notes that Huang first published a larger figure, 300,000 systems, then deleted it and reposted the lower number, and that Nvidia has not explained the change.

OpenAI has not called Astra artificial general intelligence (AGI). On the same day, OpenAI chief scientist Jakub Pachocki published "An Alien Mind", arguing that no laboratory has solved alignment and monitoring well enough to justify scaling at maximum speed. Pachocki separates goal alignment, meaning whether a system pursues the objective it was given, from value alignment, meaning whether it holds and generalizes high-level principles under conflicting objectives and adversarial conditions. He writes that he expects, and hopes, that voluntary slowdowns become common until shared safety thresholds exist, enforced by third-party auditors, government agencies, or international bodies.

The incentives on both sides are worth examining. Huang sells the accelerators that Astra was trained on, and a declaration that the AGI threshold has been crossed supports demand for the next 400,000 units. Pachocki runs research at the lab that built the model, and he is arguing for a coordinated slowdown that would apply to competitors as well as to OpenAI. Neither statement is independently verified. What is checkable is the hardware description: an NVLink72 rack couples 72 processors inside one enclosure so they function as a single compute block, and a training run on 100,000 such systems is an investment that only a handful of buyers can finance.