Nvidia's Huang declares "AGI has arrived" — note who is declaring it, and on what evidence

Nvidia's chief executive settled a debate by fiat on Sunday. "From ChatGPT to o1 to Astra in 4 years," Jensen Huang wrote on X, per Business Insider. "AGI has arrived. Congratulations @OpenAI team." He also noted that Astra, OpenAI's newest model, was trained on Nvidia's chips.

The model itself was unveiled the previous Thursday. OpenAI called Astra the world's "most intelligent and aligned model" and said it can perform "the most demanding professional work with unmatched speed, accuracy, and judgment," with a rollout to customers beginning this week. Company president Greg Brockman told reporters: "Welcome to the AGI era," adding that people will one day look back and place AGI's creation "about this time, and I think it might be about this model." OpenAI's own working definition of AGI is "highly autonomous systems that outperform humans at most economically valuable work."

The pushback was immediate and, importantly, methodological rather than tribal. Researcher Gary Marcus wrote 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"; by his own ten-point test, he says Astra meets one or two criteria. The ARC Prize researchers — whose benchmark OpenAI itself highlighted in the launch — called the results a major advance but explicitly declined to call it AGI, noting the tests run in tightly bounded environments that do not reflect the open-endedness of the real world. Altman, meanwhile, has called AGI "a very poorly defined term" and "an irrelevant marketing term." And this is not Huang's first declaration: in March he told Lex Fridman "I think we've achieved AGI" under a definition involving building a billion-dollar company. Elon Musk said in January that humanity had "entered the technological singularity"; Marc Andreessen said in May that the threshold had been crossed months earlier.

The structural detail matters more than any of the quotes. The declaration comes from the supplier. OpenAI called Nvidia "the foundation of our infrastructure" in a March funding announcement, and frontier labs across the board — OpenAI, Meta, Anthropic, Google — depend on Nvidia silicon. An AGI announcement from the chip vendor is simultaneously a demand forecast for chips. That is not evidence of bad faith; it is an incentive structure, and readers should price it in the way they price an analyst note written by someone holding the stock.

This fits a pattern we keep flagging: bounded benchmarks saturate, and once they do, the marketing outruns the measurement. It is the same failure mode we documented in cybersecurity evaluations, where models score near-perfect on saturated tests while still failing expert unguided tasks. ARC Prize is being unusually honest about exactly this limitation regarding its own benchmark, which is to its credit and should be the norm.

Our read: "AGI" has stopped doing useful work as a word, and the sooner it is retired the better the conversation gets. The questions that still discriminate are unglamorous and answerable: what can the system do unguided, in messy environments, over long horizons, at what error rate and what cost? If Astra genuinely handles demanding professional work, that will show up over the next few months in deployment data, in the tasks companies stop staffing, and in hard unsaturated evaluations — not in a congratulatory post. We remain long-term optimists precisely because this class of tool is the shortest path we know to curing diseases and generating real abundance. But the near-term risk is not a runaway superintelligence. It is a single undefined word being asked to do the work of proof, and dragging benchmarks, policy and capital allocation along behind it.