Huang says 'AGI has arrived' — the man selling the GPUs shouldn't get to define the finish line

Jensen Huang declared the finish line crossed. In a post on X on Sunday, Nvidia's CEO congratulated OpenAI on the release of Astra, the company's newest and most powerful model, writing: "From ChatGPT to o1 to Astra in 4 years. AGI has arrived." He also noted that Astra was trained on Nvidia chips, and added that "400K GPUs coming online next."

OpenAI made a similar claim in softer packaging. The company unveiled Astra on Thursday, describing it as the world's "most intelligent and aligned model" and as capable of performing "the most demanding professional work with unmatched speed, accuracy, and judgment," per Business Insider's account. President Greg Brockman told reporters: "Welcome to the AGI era," and said people will one day look back and conclude AGI was created "about this time, and I think it might be about this model." It is rolling out to customers this week.

The pushback was immediate and specific. Researcher Gary Marcus wrote on Substack that Huang "gave no evidence and no definitions, which feels to me like an effort at a takeover of a scientific question by corporate fiat," and that "declaring victory without a definition simply muddies the waters." Marcus applied his own ten-point AGI definition and said Astra meets one or two of them: "By conventional definitions, Astra still falls short." We are reporting a dispute, not adjudicating it — nobody in this exchange has published the head-to-head numbers that would settle it.

The sharpest counterwitness is Altman himself. In a recent appearance on the "Sources" podcast, OpenAI's CEO called AGI "a very poorly defined term," adding, "I was going to say it's like an irrelevant marketing term." So the vendor of the model thinks the label is marketing; the vendor of the chips thinks the label has been earned. That asymmetry is worth sitting with, because it tells you what the phrase is for.

Follow the incentives without imputing bad faith. Nvidia reported $96.2 billion in quarterly revenue in August, more than double the same period a year earlier, with $89 billion of it from the data-center business that houses its AI chips. OpenAI, in a March funding announcement, called Nvidia "the foundation of our infrastructure," noting that its training fleet and most of its inference stack run on Nvidia GPUs. When the supplier of the shovels announces that gold has been struck, the statement may still be true — but it is not independent testimony, and it should not be read as one.

Our reading: "AGI" has quietly stopped being a scientific milestone and become a governance trigger, which is exactly why the definition fight matters. Regulators, insurers, procurement officers and courts are drafting rules that reference capability thresholds. If the threshold is whatever a chip CEO says it is on a Sunday afternoon, the category is useless for policy — and, worse, it becomes a lever. Declaring arrival can justify anything from export controls to liability shields to the next capital raise. This is the pattern we keep flagging on this beat: measurement beats marketing, and demonstrated capability is not the same thing as an announced one.

The honest version is less quotable and more useful. Frontier models are now good enough that the interesting question is no longer "is it general" but "which specific professional workflows can it be trusted to own, and what happens when it fails?" That question has answers you can test — on hard, unsaturated benchmarks, in real deployments, with error rates you can audit. It also has real consequences right now, because a model marketed as capable of "the most demanding professional work" will be handed exactly that work, whether or not the label above it is accurate. The long arc still bends toward abundance and better medicine. But we get there by grading capability honestly, not by letting whoever ships the most GPUs ring the bell.