Nvidia's (NVDA) Jensen Huang says GPT-6 Astra was trained on 100,000-plus Grace Blackwell GPUs, with 400,000 more coming
Key points
- Huang: Astra trained on "~100K+ NVIDIA Grace Blackwell NVLink72"
- "400K GPUs coming online next," location not stated
- Nvidia's CEO said "AGI has arrived"; OpenAI's president hedged
- Ellison has said Abilene will hold more than 450,000 GB200 GPUs
Nvidia (NVDA) chief executive Jensen Huang said Sunday that OpenAI trained GPT-6 Astra on more than 100,000 of Nvidia's Grace Blackwell GPUs, with another 400,000 GPUs coming online next, without specifying a timetable or location. "GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72," he wrote on X on September 6. "From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations @OpenAI team. 400K GPUs coming online next."
OpenAI's launch announcement on September 3 said the model was trained on more than 100,000 GPUs at its Stargate site in Texas, but did not name the chip and did not mention a further 400,000. Huang was replying to Chase Lochmiller, chief executive of Crusoe, who had written earlier Sunday: “I guess this makes Abilene the birthplace of AGI!” Crusoe is building the Stargate campus in Abilene, Texas, which Oracle operates for OpenAI.
Advertisement
What hardware powered Astra’s training?
Grace Blackwell is Nvidia's pairing of its Grace CPU with Blackwell GPUs, and NVL72 is the rack it ships in. Nvidia's product page says the GB200 NVL72 "connects 36 Grace CPUs and 72 Blackwell GPUs in a rack-scale, liquid-cooled design" with "a 72-GPU NVIDIA NVLink domain that acts as a single, massive GPU."
Read alongside OpenAI's figure, Huang's "~100K+" refers to GPUs, not racks. At 72 GPUs per rack, 100,000 GPUs is about 1,400 NVL72 racks, by our arithmetic, though Huang did not give a rack count.
Where are the 400,000 GPUs?
Huang did not say. One possible location is Abilene itself. Oracle chairman Larry Ellison said in October 2025, at Oracle's AI World conference, that the Stargate campus there will eventually house more than 450,000 Nvidia GB200 GPUs, Data Center Dynamics reported, on a site with a stated target of 1.2 gigawatts. An earlier plan, from March 2025, called for 64,000 GB200s there by the end of 2026. Whether Huang's 400,000 refers to that campus, to the Stargate sites OpenAI has announced in other states, or to something else, he did not say, and a training cluster need not match a campus's installed capacity.
Advertisement
Who has said AGI has arrived?
Huang's "AGI has arrived" is more categorical than Brockman's launch-day remarks. President Greg Brockman said on launch day, "It's not unreasonable to feel that we are now in the AGI era," and, of the moment AGI is later judged to have been created, "I think it might be about this model," Gizmodo reported, citing Wired. Huang has been saying a version of this for a while. On September 2, in a fireside chat with Commerce Secretary Howard Lutnick at the G20 Innovation Ministerial in Chapel Hill, North Carolina, he said, "In the next couple of years, we are going to achieve essentially what people call AGI," and added, "I would argue that we're practically there today," according to a transcript of the session.
The outside evidence so far is narrower than the word. On Saturday, in two independent tests, Astra ranked ahead of Anthropic's Claude Fable 5.1 on Arena's web-coding leaderboard and Signal65's 280-job enterprise benchmark. Neither tested general intelligence, and both came with caveats their authors stated.
What it means for Nvidia (NVDA)
The post does not establish anything about new orders, delivery timing or revenue. The training hardware was deployed before Astra's release, but Huang's post does not identify when Nvidia recognized the associated revenue. The 400,000 come with no date, no named buyer and no site. Nvidia's next quarterly report is due in November; its August results put its three largest customers at 54 percent of revenue, a concentration we examined last month. What the post does establish is who supplied the training hardware, stated by the supplier, with one forward-looking line attached: "400K GPUs coming online next."
Advertisement
