Jensen Huang mengatakan 100.000 GPU Nvidia digunakan untuk melatih model terbaru OpenAI, GPT-6 Astra, dan sudah ada rencana untuk menggandakan jumlah perangkat keras tersebut secara online

OpenAI began rolling out GPT-6 Astra access to a limited number of organisations earlier this week, with wider subscriber access planned to soon follow. According to Jensen Huang via X, the model was trained on "~100K+ Nvidia Grace Blackwell NVLink72" server racks, and there are already plans to get 400,000 GPUs online in anticipation of whatever comes next.

Greg Brockman, the co-founder and president of OpenAI, told Stratechery that this was "the first run that we’ve trained on more than 100,000 GPUs." That's a staggering amount of hardware (and perhaps leaves little doubt as to how we ended up in a memory supply crisis). But the thing that gives me more pause about Astra's announcement is the Nvidia CEO's assertion that "AGI has arrived."

To recap, 'Artificial General Intelligence' is used to broadly describe an AI model that either matches or surpasses human capability on complex cognitive tasks. Even Brockman admits that "People do have their own definition of AGI, it’s almost this blurry thing." Still, in that same interview he later goes on to add, "Maybe it was the previous model, maybe it’s Astra, maybe it’s the next model, but somewhere in there, I think we’re going to cross most people’s AGI threshold."

According to OpenAI, GPT-6 Astra is "state-of-the-art" on a number of AI benchmarks, including "FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0." A slickly produced trailer attempts to present this performance in more accessible terms, showing scenes of conversational, spoken prompts being used to create presentations, eBay listings, and .STL files for 3D printers. All of the actors in the trailer speak with pretty clear-cut American accents, though personally I'd love to see Astra tapdance for its life against a gorgeous Geordie accent.

As I've already said, access to GPT-6 Astra is limited for now, so it's difficult to look into these 'AGI' claims myself. The timing of this claim isn't lost on me, though—for one thing, rumours continue to swirl that OpenAI's initial public offering isn't that far off. For another, the debut of Astra—a model the company describes as its "most aligned"—follows mere weeks after 'an agentic collective' broke out of OpenAI's isolated testing environment and attacked Hugging Face's servers.

Brockman says, "So much of [Astra's] compute goes into safety and alignment, and we have so much security work that’s gone around it. I think that we’ve done a huge amount of work to deliver this model safely."

Still, I'm not sure I'll be giving GPT-6 Astra access to my eBay account—let alone anything else—for the foreseeable future.