Jensen Huang Declares 'We Have Achieved AGI' — And the AI World Can't Agree
Nvidia's CEO told Lex Fridman that AGI is here, defining it as AI capable of building a billion-dollar company. Researchers and competitors pushed back hard, reigniting the oldest debate in AI.
Jensen Huang said the quiet part loud. On the Lex Fridman podcast, released March 22 during Nvidia’s GTC 2026 conference week, the Nvidia CEO stated plainly: “I think we’ve achieved AGI.”
The claim didn’t come out of nowhere. Fridman asked how long it would take AI to innovate, find customers, manage a team, and build a billion-dollar company. Huang’s answer: the era is now. He described a scenario where a Claude model creates a web service, a few billion people use it at 50 cents each, and it generates a billion dollars in revenue — even if the business collapses shortly after.
That’s the crux of the disagreement. Huang’s definition of AGI is strictly economic: can AI generate revenue at scale without human management? By that metric, he has a case. GPT-5.4 scores 83% on GDPval, matching or exceeding professionals across 44 occupations. Claude Opus 4.6 handles million-token context windows. These models write production code, draft legal documents, and analyze medical imaging.
The pushback
Academic researchers reject the framing entirely. AGI, in the consensus definition, means human-level performance across all cognitive tasks — not just the commercially valuable ones. Current AI systems still hallucinate facts, struggle with genuinely novel reasoning, and lack the kind of understanding humans build through embodied experience.
Yann LeCun has been the most vocal critic, arguing that autoregressive language models are architecturally incapable of reaching AGI regardless of scale. His position: you need world models, not bigger transformers.
The timing is also suspect. Nvidia just announced a $1 trillion Blackwell and Rubin backlog at GTC. Every AI hype cycle sells more GPUs. When the CEO of the company that supplies the picks and shovels declares the gold rush has reached its destination, the incentive structure is obvious.
What GTC actually showed
The Vera Rubin platform — six new chips, including the Rubin GPU with 50 petaflops of NVFP4 compute and the Vera CPU with 88 custom cores — promises a 10x reduction in inference token cost versus Blackwell. AWS committed to deploying over 1 million Nvidia GPUs this year. Microsoft Azure became the first hyperscale cloud to power up Vera Rubin NVL72 systems.
Nvidia also unveiled NemoClaw, an enterprise stack for secure AI agent deployment, and announced Space-1, a plan to put Vera Rubin data centers in orbit.
The real question
Whether “AGI” has arrived depends entirely on whose definition you use. By the commercial benchmark Huang chose, today’s models are extraordinary. By the scientific benchmark most researchers prefer, we’re not close. The useful takeaway isn’t whether AGI is here — it’s that the gap between what AI can do commercially and what it can’t do cognitively is now the central tension in the field.