Nvidia CEO Jensen Huang congratulated OpenAI on social media for launching GPT-6 Astra, adding a verdict of his own: AGI has arrived. He wrote that the model was "trained on more than 100,000 NVIDIA Grace Blackwell NVLink72" systems, and added that the path from ChatGPT to o1 to Astra took just four years, with the next batch of 400,000 GPUs coming online soon.
OpenAI offered two more measured statements. Aidan Clark, the company's VP of research training, said this was the first time OpenAI had pretrained a model on more than 100,000 GPUs at its Stargate site in Texas. President Greg Brockman was more cautious, saying it is "not unreasonable" to say the industry has entered the AGI era.
Astra's scorecard went public the same day: 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and a perfect score on ExploitBench. OpenAI also said the model pushed the known upper bound on prime gaps to 186. API pricing is set at $10 per million input tokens and $50 per million output tokens, with caching billed separately.
The 100,000 Figure: a Unit That Was Never Settled
Huang's line about "100,000-plus" has been read two ways: some outlets describe 100,000 systems, others 100,000 GPUs. NVLink72 refers to a full rack that links 72 GPUs together, so the two readings differ by a factor of more than 70. Aidan Clark's own phrasing — "pretraining on more than 100,000 GPUs" — points toward a GPU count rather than a rack count. Neither Nvidia nor OpenAI has clarified the unit, and no public material discloses the rack scale at the Texas site.
Also unexplained is which "next 400,000 GPUs" Huang meant — what hardware, where it will be deployed, or whether it is meant for training or inference. The line has been widely repeated since, but it remains a single social media remark with no matching deployment announcement.
Four Years, and Who Gets to Define AGI
Huang's timeline runs from ChatGPT to o1 to Astra, strung together by a steady scale-up in training size rather than any public benchmark that was cleared along the way. His case rests on compute and scores; Nvidia has not said what standard it used to conclude AGI has arrived.
OpenAI's own definition, still posted on its site, dates to its 2018 charter: highly autonomous systems that outperform humans at most economically valuable work. That wording has never come with a measurement method — how many jobs count as "most," or what baseline "outperform humans" is measured against, remain undefined. Benchmark scores can only speak to individual tasks, leaving a gap to that definition that nobody has closed.
Sam Altman has taken the opposite position. According to public reports, he considers AGI a vague, largely irrelevant marketing term. Two executives of the model's own maker, plus the head of its largest compute supplier, have landed on three different readings of the same launch.
For anyone trying to verify this from outside, only the price and the scores are checkable. At $10/$50 per million tokens, Astra costs more than the previous flagship, so bills for long-context work will climb; FrontierMath- and ExploitBench-style problems, meanwhile, remain some distance from everyday engineering work. Whether AGI has arrived has no citable determination right now — only three executives' differing choice of words.
Sources: Nvidia executive statements on social media, CocoLoop, OpenAI's model launch page, multiple tech media reports; GPU figures, benchmark scores and API pricing cross-checked individually.