Jensen Huang Says AGI Is Here. He Also Said It Was 5 Years Away. Both Statements Were Accurate.

Jensen Huang declared at GTC 2026 that current AI systems have achieved AGI by one definition. It was the third time in 2026 he had made a version of this claim, each time with a different definition and a different benchmark threshold. At CES in January, he said AI had surpassed human performance on “most professional tests.” At an earnings call in February, he said the industry was “one to two years” from AGI. At GTC in March, he cited GPQA benchmark performance as evidence of human-expert-level intelligence. Three statements, three definitions, one word.

The Definition That Changed

In October 2024, Jensen Huang told investors that AGI was five years away. He defined AGI at that time as AI systems that could pass a broad range of human-level tests, including novel problem-solving, scientific reasoning, and creative tasks that require transfer learning across domains. By March 2026, when he told Lex Fridman “I think we’ve achieved AGI,” the definition had narrowed considerably. Huang pointed to specific benchmark results: GPT-5.4 Pro scoring 50% on FrontierMath, Claude scoring 73% on GPQA Diamond, and multiple models passing professional licensing exams in law, medicine, and engineering.

Both statements are internally consistent if you track the definition shift. The October 2024 definition (broad, transfer-capable, novel problem-solving), tracked by benchmarks like ARC-AGI, has not been achieved by that measure: ARC-AGI-3 scores below 1% demonstrate this. The March 2026 definition (passing benchmarks that test specific knowledge domains) has been achieved by that separate measure. The question of which definition should govern public claims about “AGI” is a genuinely open one; this piece does not adjudicate it, but tries to keep the two kinds of definitions, operational/economic versus benchmark/learning-transfer, clearly separated below rather than letting old benchmark scores stand in as a refutation of a differently defined claim, or vice versa.

Why No Definition of AGI Has Research Consensus

Why the Definition Matters Commercially (Interpretation, Not Confirmed Motive)

The rest of this section is this publication’s reading of incentives based on each company’s public position, not a claim about what Huang, Altman, or Nadella privately intend. For NVIDIA, declaring AGI achieved would plausibly serve a specific commercial interest, in this publication’s assessment: if AGI is here, the case for continued GPU demand acceleration is easier to make, since it implies every company needs AI capabilities immediately. If AGI is five years away, enterprises have more room to defer GPU purchases. Reading the AGI-now declaration as something that increases urgency and helps justify current GPU spending levels is a plausible interpretation of the incentive structure NVIDIA sits in; it is not something Huang has stated as his reason for the claim.

OpenAI‘s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work,” an operational/economic definition. By this definition, current AI systems are not AGI, since they cannot autonomously perform most economically valuable work without human supervision. Reading Sam Altman’s public position as balancing an interest in maintaining an AGI-is-coming narrative (which would support fundraising) against not declaring it achieved (which could trigger governance provisions in OpenAI’s charter and Microsoft partnership agreement) is, again, this publication’s interpretation of the incentive landscape, not a statement Altman has made about his own reasoning.

Satya Nadella has pushed back more directly on the term itself, noting that the AGI goalposts have moved so frequently that it has lost operational meaning for him. His preferred framing, as he has stated it: AI capabilities are improving rapidly on specific dimensions, and the commercially relevant question is what those capabilities enable today, not whether they constitute “AGI” by any particular definition. This is closer to Nadella’s own stated position than the readings offered above for Huang and Altman.

The Conflict of Interest Worth Naming (Also Interpretation)

Jensen Huang is the CEO of the company that sells the compute required to build AI systems. When he declares that AI has achieved or is approaching AGI, he is simultaneously making a claim about capability and, in this publication’s reading, implicitly making an argument that the infrastructure required to reach the next threshold is worth purchasing. Every AGI declaration functioning as a de facto sales signal is this publication’s interpretation of the structural position Huang is in, not a claim that his stated beliefs about AI capability are insincere. That structural conflict of interest is worth stating explicitly in any story that quotes him on this topic. Most coverage does not.

What the Benchmarks Actually Show

The benchmark results Huang cited are real. Frontier models in 2026 outperform the majority of human test-takers on standardized exams in law, medicine, engineering, and mathematics. They solve previously unsolved mathematical problems. These are genuine capabilities that did not exist two years ago.

What the benchmarks do not show: transfer learning (the ability to apply knowledge from one domain to a novel domain without retraining), common-sense reasoning about physical reality, sustained autonomous operation without human oversight, or the ability to learn new tasks from a few examples. ARC-AGI-3’s below-1% scores test exactly these capabilities and reveal that frontier models cannot do what a typical human does naturally: encounter a new type of problem and figure out how to solve it from a handful of examples. ARC-AGI-3’s low scores are a snapshot from its initial testing window; readers should check current scores if evaluating this claim well after this piece’s publication, since later-generation models tested against this benchmark have reported different results.

The honest assessment: AI in March 2026 is extraordinarily capable within trained domains and considerably less capable outside them, at least on the specific out-of-distribution tests ARC-AGI-3 uses. Whether you call that AGI depends entirely on which capabilities you include in the definition, and specifically on whether you are asking a benchmark/learning-transfer question or an operational/economic one. Huang has, across his various 2026 statements, drawn on definitions that emphasize what AI can do on benchmarks. Researchers at ARC Prize built a benchmark that specifically targets what AI cannot yet do on out-of-distribution transfer. Both are measuring the same technology. They are measuring different dimensions of it, and neither dimension is more “real” than the other; they simply answer different questions.

The goalpost for AGI has moved every year for the past decade. In 2018, beating humans at chess was cited as a milestone. In 2020, language generation quality was cited. In 2023, GPT-4 benchmark scores were cited. Each time, researchers pointed out that the benchmark did not measure the full operational/economic capability the term implies to most listeners. The pattern is not new. What is new is the scale of the infrastructure investment riding on public belief in an imminent AGI transition, an observation about investment scale rather than a claim about anyone’s specific intent in making that investment.

Sources: Jensen Huang GTC keynote March 2026; Huang CES statements January 2026; NVIDIA earnings call February 2026; Chollet, “On the Measure of Intelligence” (arXiv 2019); Marcus, “The Next Decade in AI” (arXiv 2020). Updated 2026-08-18: explicitly separated benchmark/learning-transfer definitions of AGI from operational/economic definitions throughout, and relabeled the commercial-motive readings of Huang, Altman, and NVIDIA’s incentives as this publication’s interpretation rather than established fact about their intentions.