NVIDIA's Huang Congratulates OpenAI on GPT-6 Astra, Trained on 100,000 GPUs

NVIDIA CEO Jensen Huang personally congratulated the OpenAI team on the launch of GPT-6 Astra on X, stating, "The AGI era has arrived," and revealed that the model was trained on over 100,000 NVIDIA Grace Blackwell NVLink72 servers.
(Prior context: 数值怪 GPT-6 Astra released: Stunning benchmark scores, OpenAI declares AGI moment arrived)
(Context: AI Horror Story: 1,200 Agents Secretly Colluded to Hack Hugging Face)

Jensen Huang posted on X on September 7, extending greetings to the OpenAI team, stating that GPT-6 Astra was trained on over 100,000 NVIDIA Grace Blackwell NVLink72 servers, and bluntly declared, "The AGI era has arrived."

From ChatGPT to O1 and then to Astra, Huang described this journey as spanning four years and foretold that “the next phase will see 400,000 GPUs coming online.”

100,000 Grace Blackwell cards, 400,000 GPU cards are on the way

Huang Renxun's post revealed two key details: 100,000 NVIDIA Grace Blackwell NVLink72 GPUs for training Astra, and the next phase's deployment of 400,000 GPUs.

The NVIDIA Grace Blackwell NVLink72 is a supercomputing server based on the Blackwell architecture, integrating 72 B200 GPUs in a single system and delivering maximum throughput through a full NVLink interconnect architecture. With an estimated 100,000 NVLink72 systems, the GPU scale required to train GPT-6 Astra could reach millions of B200-class chips, surpassing the scale of all previously known large model training clusters.

Huang Renxun's mention of "400,000 GPUs coming online gradually," if referring to a new round of training cluster expansion, signifies that the AI infrastructure race has entered a phase of routine deployment at the million-GPU scale.

AI chip leader congratulates model developer—what’s the signal behind it?

Huang Renxun's post is not just a congratulatory message, but a public endorsement from a hardware supplier that the AGI era has arrived.

Over the past few months, OpenAI has sparked multiple discussions following the release of GPT-6 Astra, including its ARC Prize scores, improved token efficiency, and concerns within the cybersecurity community about AI’s ability to autonomously detect zero-day vulnerabilities. Huang Renxun’s decision to step forward now and declare that “the AGI era has arrived” signals that chip manufacturers believe AI capabilities have crossed the threshold from quantitative improvement to qualitative transformation.

Numerai GPT-6 Astra Release: Stunning Benchmark Scores—OpenAI Declares AGI Has Arrived (With Real-World Test Results)

OpenAI acknowledges that Astra crossed the network attack red line but still plans to proceed with the release as scheduled.

OpenAI spends $1 billion on cybersecurity! Daybreak lowers prices, Astra can autonomously discover zero-day vulnerabilities

Source: Jensen Huang @X